eurasip journal

Vehicular Ad Hoc Networks

Guest Editors: Hossein Pishro-Nik, Shahrokh Valaee, and Maziar Nekovee

EURASIP Journal on Advances in Signal Processing

EURASIP Journal on Advances in Signal Processing


Guest Editors: Hossein Pishro-Nik, Shahrokh Valaee,and Maziar Nekovee

Copyright © 2010 Hindawi Publishing Corporation. All rights reserved.

This is a special issue published in volume 2010 of “EURASIP Journal on Advances in Signal Processing.” All articles are open accessarticles distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction inany medium, provided the original work is properly cited.

Editor-in-ChiefPhillip Regalia, Institut National des Telecommunications, France

Associate Editors

Adel M. Alimi, TunisiaYasar Becerikli, TurkeyKostas Berberidis, GreeceA. Enis Cetin, TurkeyJonathon A. Chambers, UKMei-Juan Chen, TaiwanLiang-Gee Chen, TaiwanSatya Dharanipragada, USAKutluyil Dogancay, AustraliaFlorent Dupont, FranceFrank Ehlers, ItalySharon Gannot, IsraelSamanwoy Ghosh-Dastidar, USANorbert Goertz, AustriaM. Greco, ItalyIrene Y. H. Gu, SwedenFredrik Gustafsson, SwedenUlrich Heute, GermanySangjin Hong, USAShih-Syuan Huang, TaiwanJiri Jan, Czech RepublicMagnus Jansson, SwedenSudharman K. Jayaweera, Mexico

S. Jensen, DenmarkMark Kahrs, USAMoon Gi Kang, Republic of KoreaWalter Kellermann, GermanyLisimachos P. Kondi, GreeceAlex Chichung Kot, SingaporeC.-C. Jay Kuo, USAErcan E. Kuruoglu, ItalyTan Lee, ChinaGeert Leus, The NetherlandsT.-H. Li, USAHusheng Li, USAMark Liao, TaiwanShoji Makino, JapanStephen Marshall, UKC. Mecklenbrauker, AustriaGloria Menegaz, ItalyRicardo Merched, BrazilMarc Moonen, BelgiumChristophoros Nikou, GreeceSven Nordholm, AustraliaPatrick Oonincx, The NetherlandsDouglas O’Shaughnessy, Canada

Bjorn Ottersten, SwedenJacques Palicot, FranceAna Perez-Neira, SpainWilfried Philips, BelgiumAggelos Pikrakis, GreeceIoannis Psaromiligkos, CanadaAthanasios Rontogiannis, GreeceGregor Rozinaj, SlovakiaMarkus Rupp, AustriaWilliam Allan Sandham, UKBulent Sankur, TurkeyErchin Serpedin, USALing Shao, UKDirk Slock, FranceYap-Peng Tan, SingaporeJoao Manuel R. S. Tavares, PortugalGeorge S. Tombras, GreeceDimitrios Tzovaras, GreeceBernhard Wess, AustriaJar Ferr Yang, TaiwanAzzedine Zerguine, Saudi ArabiaAbdelhak M. Zoubir, Germany

Contents

Vehicular Ad Hoc Networks, Hossein Pishro-Nik, Shahrokh Valaee, and Maziar NekoveeVolume 2010, Article ID 864032, 1 page

Driver Drowsiness Warning System Using Visual Information for Both Diurnal and NocturnalIllumination Conditions, Marco Javier Flores, Jose Marıa Armingol, and Arturo de la EscaleraVolume 2010, Article ID 438205, 19 pages

Multiobjective Reinforcement Learning for Traffic Signal Control Using Vehicular Ad Hoc Network,Duan Houli, Li Zhiheng, and Zhang YiVolume 2010, Article ID 724035, 7 pages

Design and Experimental Evaluation of a Vehicular Network Based on NEMO and MANET,Manabu Tsukada, Jose Santa, Olivier Mehani, Yacine Khaled, and Thierry ErnstVolume 2010, Article ID 656407, 18 pages

Traffic Data Collection for Floating Car Data Enhancement in V2I Networks, D. F. Llorca, M. A. Sotelo,S. Sanchez, M. Ocana, J. M. Rodrıguez-Ascariz, and M. A. Garcıa-GarridoVolume 2010, Article ID 719294, 13 pages

Improvement of Adaptive Cruise Control Performance, Shigeharu Miyata, Takashi Nakagami,Sei Kobayashi, Tomoji Izumi, Hisayoshi Naito, Akira Yanou, Hitomi Nakamura, and Shin TakeharaVolume 2010, Article ID 295016, 8 pages

Reducing Congestion in Obstructed Highways with Traffic Data Dissemination Using Ad hoc VehicularNetworks, Thomas D. Hewer, Maziar Nekovee, and Peter V. CoveneyVolume 2010, Article ID 169503, 10 pages

Reliable Delay Constrained Multihop Broadcasting in VANETs, Martin Koubek, Susan Rea,and Dirk PeschVolume 2010, Article ID 753256, 13 pages

Traffic Flow Condition Classification for Short Sections Using Single Microwave Sensor,Muhammed G. Cinsdikici and Kemal MemisVolume 2010, Article ID 148303, 13 pages

Hindawi Publishing CorporationEURASIP Journal on Advances in Signal ProcessingVolume 2010, Article ID 864032, 1 pagedoi:10.1155/2010/864032

Editorial


Hossein Pishro-Nik,1 Shahrokh Valaee,2 and Maziar Nekovee3

1 University of Massachusetts Amherst, Amherst, MA 01003, USA2 University of Toronto, Toronto, ON, Canada M5S 1A13 University College London, London WC 1E 6BT, UK

Correspondence should be addressed to Hossein Pishro-Nik, [email protected]

Received 5 October 2010; Accepted 5 October 2010

Copyright © 2010 Hossein Pishro-Nik et al. This is an open access article distributed under the Creative Commons AttributionLicense, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properlycited.

With vehicular ad hoc networks gaining an ever-increasinginterest to serve a diverse variety of applications in today’sintelligent transportation systems, it was not at all surprisingfor the guest editorial team to receive a handful of sub-missions for this special issue addressing different aspectsand test-beds of vehicular networks. In sum, 8 papers wereaccepted to be published in the special issue. An interestingnote to make is that 5 of the accepted papers had an actualexperimental implementation carried out in the road andunder real-world conditions. This certainly helps to justifytheir application and usefulness for future deployment by theindustry and authorities.While all papers address enhancingthe safety and efficiency of driving, each of them addresses acertain aspect of this issue.

The paper by M. J. Flores et al., “Driver DrowsinessWarning System Using Visual Information for Both Diurnaland Nocturnal Illumination Conditions,” seeks to locate,track, and analyze both the drivers face and eyes to computea drowsiness index under varying light conditions (diurnaland nocturnal).

In their paper “Multiobjective Reinforcement Learning forTraffic Signal Control Using Vehicular Ad Hoc Network,” D.Houli et al. propose a new multiobjective control algorithmbased on reinforcement learning for urban traffic signalcontrol, named, multi-RL.

M. Tsukada et al. in “Design and Experimental Evaluationof a Vehicular Network Based on NEMO and MANET,”present a policy-based solution to distribute traffic amongmultiple paths to improve the overall performance of avehicular network.

The paper “Traffic Data Collection for Floating Car DataEnhancement in V2I Networks” by D. F. Llorca et al. presents

a complete vision-based vehicle detection system for floatingcar data (FCD) enhancement in the context of vehicular adhoc networks.

S. Miyata et al. in “Improvement of Adaptive CruiseControl Performance” propose a more accurate method fordetecting the preceding vehicle by radar while cornering.

The paper “Reducing Congestion in Obstructed Highwayswith Traffic Data Dissemination Using Ad hoc Vehicular Net-works” by T. D. Hewer et al. presents a message-dissemi-nation procedure that uses vehicular wireless protocolsto influence vehicular flow, reducing congestion in roadnetworks.

M. Koubek et al., in “Reliable Delay Constrained MultihopBroadcasting in VANETs,” focus on mechanisms that improvethe reliability of broadcasting protocols, where the emphasisis on satisfying the delay requirements for safety applications.

Finally, M. G. Cinsdikici and K. Memis in “TrafficFlow Condition Classification for Short Sections Using Sin-gle Microwave Sensor” seek to identify the current trafficcondition by examining the traffic measurement parametersand taking into account occupancy as another importantparameter of classification.

We hope this special issue can help the research commu-nity further its understanding of this emerging field.

Hossein Pishro-NikShahrokh ValaeeMaziar Nekovee

Hindawi Publishing CorporationEURASIP Journal on Advances in Signal ProcessingVolume 2010, Article ID 438205, 19 pagesdoi:10.1155/2010/438205

Research Article

Driver Drowsiness Warning System Using Visual Information forBoth Diurnal and Nocturnal Illumination Conditions

Marco Javier Flores, Jose Marıa Armingol, and Arturo de la Escalera

University Carlos III of Madrid, C/. Butarque 15, 28991 Leganes, Madrid, Spain

Correspondence should be addressed to Arturo de la Escalera, [email protected]

Received 23 November 2009; Accepted 21 June 2010

Academic Editor: Hossein Pishro-Nik

Copyright © 2010 Marco Javier Flores et al. This is an open access article distributed under the Creative Commons AttributionLicense, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properlycited.

Every year, traffic accidents due to human errors cause increasing amounts of deaths and injuries globally. To help reduce theamount of fatalities, in the paper presented here, a new module for Advanced Driver Assistance System (ADAS) which deals withautomatic driver drowsiness detection based on visual information and Artificial Intelligence is presented. The aim of this systemis to locate, track, and analyze both the drivers face and eyes to compute a drowsiness index, where this real-time system worksunder varying light conditions (diurnal and nocturnal driving). Examples of different images of drivers taken in a real vehicle areshown to validate the algorithms used.

1. Introduction

ADAS is part of the active safety systems that interact to alarger extent with drivers to help them avoid traffic accidents.The goal of such systems is to contribute to the reductionof traffic accidents by means of new technologies; that is,incorporating new systems for increasing vehicle security,and at the same time, decreasing danger situations that mayarise during driving, due to human errors. In this scenario,vehicle security research is focused on driver analysis. In thisparticular research, a more in-depth analysis of drowsinessand distraction is presented [1].

Drowsiness appears in situations of stress and fatigue inan unexpected and inopportune way and may be producedby sleep disorders, certain types of medications, and even,boredom, for example, driving for long periods of time. Thesleeping sensation reduces the level of vigilante producingdanger situations and increases the probability of an accidentoccurring.

It has been estimated that drowsiness causes between10% and 20% of traffic accidents, causing both fatalities dead[2] and injuries [3], whereas within the trucking industry57% of fatal truck accidents are caused by this problem[4, 5]. Fletcher et al. in [6] have stated that 30% of alltraffic accidents have been caused by drowsiness, and Brandt

et al. [1] have presented statistics showing that 20% ofall accidents are caused by fatigue and lack of attention.In the USA, drowsiness is responsible for 100000 trafficaccidents yearly producing costs of close to 12.000 milliondollars [7]. In Germany, one out of four traffic accidentsoriginate from drowsiness, while in England 20% of all trafficaccidents are produced by drowsiness [8], and in Australia1500 million dollars has been spent on fatalities resultingfrom this problem [9].

In this context, it is important to use new technologiesto design and build systems that are capable of monitoringdrivers and to measure their level of attention during thecomplete driving process. Fortunately, people in a state ofdrowsiness produce several visual cues that can be detectedin the human face, such as

(i) yawn frequency,

(ii) eye-blinking frequency,

(iii) eye-gaze movement,

(iv) head movement,

(v) facial expressions.

By taking advantage of these visual characteristics, com-puter vision is the most feasible and appropriate technology

2 EURASIP Journal on Advances in Signal Processing

available to deal with this problem. This paper presents thedrowsiness detection system of the IVVI (Intelligent Vehiclebased Visual on Information) vehicle [10]. The goal of thissystem is to automatically estimate the driver’s drowsinessand to prevent drivers falling asleep at the wheel.

This paper is laid out as follows. Section 2 presentsan extensive review on the state of the art consideringdifferent lighting conditions. A general framework of theproposed method is presented Section 3. There are twosystems, one for diurnal and another nocturnal driving. Bothhave a first step for face and eye detection, followed for asecond step for face and eye tracking. The output of bothsystems is a drowsiness index based on a support vectormachine. A deeper explanation of both systems is presentedin Sections 4 and 5 where the similarities and differences ofboth approaches are highlighted, and the results are shown.Finally, in Section 6, the conclusions are presented.

2. Related Work

To increase traffic safety and to reduce the number of trafficaccidents, numerous universities, research centers, automo-tive companies (Toyota, Daimler Chrysler, Mitsubishi, etc.),and governments (Europe Union, etc.) are contributing tothe development of ADAS for driver analysis [5], usingdifferent technologies. In this sense, the use of visualinformation to obtain the state of the driver drowsiness andto understand his/her behavior is an active research field.

This problem requires the recognition of human behav-ior when in a state of sleepiness by means of an eyeand facial (head) analysis. This is a difficult task, evenfor humans, because there are many factors involved, forinstance, changing illumination conditions and a variety ofpossible facial postures. Considering the illumination, thestate of the art has been divided in two parts; the firstprovides details on systems that work with natural daylightwhereas the second deals with systems which operate withthe help of illumination systems based on near infrared(NIR) illumination.

2.1. Systems for Daylight Illumination. To analyze driverdrowsiness several systems have been built in recent years.They usually require the problem to be simplified to workpartially or for specific environments; for example, D’Orazioet al. [11] have proposed an eye detection algorithm thatsearches for the eyes within the complete image and haveassumed that the iris is always darker than the sclera. Usingthe Hough transform for circles and geometrical constraintsthe eye candidates are located; next, they are passed to aneural network that classifies between eyes and noneyes. Thissystem is capable of classifying eyes as being open or closed.The main limitations to this algorithm are as follows. It isapplicable only when the eyes are visible in the image, andit is not robust for changes in illumination. Horng et al.[12] have presented a system that uses a skin color modelover an HSI space for face detection, edge information foreye localization, and dynamical template matching for eyetracking. By using color information from the eyeball, thestate of the eye is defined; thus the driver’s state can be

computed, that is, asleep or alert; if the eyes are closed forfive consecutive frames, the driver is assumed to be dozing.Brandt et al. [1] have shown a system that monitors driverfatigue and lack of attention. For this task, the Viola Jones(VJ) method has been used [13] to detect the driver’s face.By using the optical flow algorithm on eyes and the headthis system is able to compute the driver’s state. Tian andQin in [2] have built a system which verifies the state of thedriver’s eye. Their system uses Cb and Cr components ofthe YCbCr color space; with a vertical projection functionthis system locates the face region and with a horizontalprojection function it locates the eye region. Once theeyes are located the system computes the eye state using acomplexity function. Dong and Wu [3] have presented asystem for driver fatigue detection; this is based on a skincolor model on a bivariate Normal distribution and Cb andCr components of the YCbCr color space. After locating theeyes, it computes the fatigue index using the distance ofthe eyelid to classify whether the eyes are open or closed; ifthe eyes are closed for five consecutive frames, the driver isconsidering to be dozing, as in Horng’s work. Branzan et al.[14] also have presented a system for drowsiness monitoringusing template matching to analyze the state of the eye.

2.2. Systems Using Infrared Illumination. As a result of noc-turnal lighting conditions, Ji et al. in [4, 15] have presented adrowsiness detection system based on NIR illumination andstereo vision. This system locates the position of the eye usingimage differences based on the bright pupil effect. Later, thissystem computes the blind eyelid frequency and eye gaze tobuild two drowsiness indices: PERCLOS (percentage of eyeclosure over time) [7] and AECS (average eye closure speed).Bergasa et al. [5] have also developed a nonintrusive systemusing infrared light illumination this system computes thedriver’s vigilance level using a finite state automata (FSM)[16] with six different eye states that compute several indices,among them, PERCLOS; this system is also capable ofdetecting inattention considering a facial posture analysis.Other research work based on this type of illumination hasbeen presented by Grace [17], where the authors measureslow eyelid closure. Systems using NIR illumination workwell under stable lighting conditions [5, 18]; however, thesesystems present drawbacks for applications in real vehicles,where the light continually changes. In this scenario, if thespectral pupils disappear, then the eye detection processbecomes more complex.

3. System Design for Drowsiness Detection

This paper presents a system which detects driver drowsinesswhich works for both day and night time conditions andfollows the classification presented in the state of the art.

This composition has allowed two systems to beobtained, one for day and a second for night time conditions.The first works with natural daylight illumination and thesecond with artificial infrared illumination. It is interesting tonote that both systems operate using grayscale images takenwithin a real vehicle.

EURASIP Journal on Advances in Signal Processing 3

Initialize Captureimage

Facedetection

Success?No

Yes

Faceanalysis

Eyedetection

Distraction?Eyestate

Drowsiness?

Success?

Face and eyetracking

Activatealarm

Activatealarm

YesYes

YesNoNo

(a)

Initialize CaptureNIR image

Eyedetection

Success?No

Yes

Face poseestimation

Distraction?Eyestate

Drowsiness?

Faceand eyetracking

Activatealarm

Yes

NoYes

No

(b)

Figure 1: Algorithm scheme: (a) day system, (b) night system.

The general scheme of both systems is shown in Figure 1,where six modules are presented as follows:

(i) face detection;

(ii) eye detection;

(iii) face tracking;

(iv) eye tracking;

(v) drowsiness detection;

(vi) distraction detection.

Each one of these parts will be explained in the followingsections.

4. Day System Design

In this section, the daytime system based on the algorithmschematic shown in Figure 1(a) will be described, where thevisual information is acquired using a digital camera.

4.1. Face Detection. To locate the face, this system uses theVJ object detector which is a machine learning approach forvisual object detection. This makes use of three important

features to make an efficient object detector based on theintegral image, the AdaBoost technique and the cascadeclassifier [13]. Each one of these elements is important toefficiently process the images and in near real-time withcorrect detections as high as 90%. A further important aspectof this method is its robustness for changing light conditions.However, in spite of the above-mentioned features, itsprincipal disadvantage is that it cannot extrapolate and doesnot work appropriately when the face is not in front of thecamera axis. This particular case occurs when the drivermoves his/her head. This shortcoming will be analyzed lateron in this paper.

Continuing with the algorithm description, when thedriver’s face is detected, it is enclosed within a rectangleRI (region of interest) which is defined by its left-topcorner coordinates P0 = (x0, y0) and bottom-right cornercoordinates P1 = (x1, y1), as can be observed in Figures 2(a),2(b), and 2(c). The size of the rectangle has been determinedfrom experimental analysis developed on the face databasethat has been created for this task.

4.2. Eye Detection. Locating the position of the eye isa difficult task as different features define the same eye


(a) (b) (c)

Figure 2: Viola Jones method.

RIR RIL

(a)

RIR RIL

(b)

Figure 3: Eye rectangles RIR and RIL.

depending, for example, on the area of the image where itappears and on the color of the iris, but the main problemthat occurs when driving is the changes in the ambientlighting conditions.

Once the face has been located within the rectangle RIdescribed in the previous section, using the face anthropo-metric properties [19] which are derived from a face databaseanalysis, two rectangles containing the eyes are obtained.This system uses RIL for the left eye rectangle and RIR forthe right eye rectangle; this is shown in the following fourequations and also in Figure 3:

(u0L, v0L) =(x0 +

w

6, y0 +

h

4

),

(u1L, v1L) =(x0 +

w

2, y0 +

h

2

),

(u0R, v0R) =(x0 +

w

2, y0 +

h

4

),

(u1R, v1R) =(x1 − w

6, y1 − h

2

),

(1)

where w = x1 − x0 and h = y1 − y0.After the previous step; the exact position of each eye

is searched for by incorporating information from the grey-level pixels. The main idea here is to obtain a random samplefrom the pixels that belong to the eye area, and then, to adjusta parametric model. Figure 4 shows this procedure where arandom sample has been extracted in (a), and an ellipticalmodel has been adjusted in (b). In this case, the eye state isindependent, that is, it may be open or closed.

y

x

(a) (b)

Figure 4: (a) Random sample, (b) eye parametric model.

To extract the random sample, the following algorithmhas been proposed. Let I(x, y) ∈ [0, 255] be the pixel valueat the position(x, y), then do as follows.

(i) Generate image J using the following:

J(x, y

) = I(x, y

)−m

σ, (2)

where m and σ are the mean and the standard devi-ation, respectively. These parameters are computedover the previously located eye rectangles.

(ii) Generate image K using

K(x, y

) =⎧⎨⎩J(x, y

)− 256∗ δ1 if J(x, y

) ≥ 0,

256∗ δ2 + J(x, y

)if J

(x, y

)< 0,

(3)

where δ1 = max(0, ceil(J(x, y)/256) − 1), δ2 =max(1, ceil(|J(x, y)|/256)), and ceil(x) is the functionthat returns the smallest integer larger than x.


(a) (b)

(c) (d)

Figure 5: Eye location using RL and RR: (a) grayscale image, (b) binary image (B), (c) gradient image (G), and (d) logarithm image (L).

(a) (b)

(c) (d)

(e) (f)

Figure 6: Expectation maximization algorithm over the spatial distribution of the eye pixels: (a) eye image, (b) ellipse parameters: center,axes, and inclination angle. (c), (d), (e), and (f) show other examples of this procedure.

(iii) Obtain the binary image, B, from image K usingOstu’s method [20] which calculates an automaticthreshold (Figure 5(b)).

(iv) Compute the gradient image, G, using the Sobelhorizontal (Sx) and vertical (Sy) edge operatorfollowed by an image contrast enhancement [21](Figure 5(c)).

(v) Compute the logarithm image [22], L, where theobjective here is to enhance the iris pixels that formthe central part of the eye (Figure 5(d)).

With the pixels that have been extracted from images B,G, and L, it is possible to obtain the previously mentionedrandom sample. This sample presents an ellipse shapewhere an elliptical model has been adjusted over this usingthe expectation maximization algorithm (EM) [23]. Specialattention is paid to the center of the ellipse, because, itallows the exact position of the center of the eye center tobe obtained. The ellipse axes determine the width and heightof the eyes. The result is shown in Figure 6(b).

The main reason behind using pixel information froma random sample is due to the fact that head movements,illumination changes, and so forth, do not allow complete


(a) (b) (c) (d) (e)

Figure 7: Examples of a face database which contain faces with different orientations: (a) left profile, (b) front view, (c) right profile, (d)down profile, and (e) up profile.

(a) (b) (c)

Figure 8: Mask for face training and its result.

eye pixel information to be obtained, that is, only partialinformation of the eye in images B, G, and L is available,where the elliptical shape prevails. This random informationmakes it feasible to use an algorithm that computes theparameters of a function which approximate the eye ellipseshape. EM computes the mean, variance, and the correlationof the X and Y coordinates that belong to the eye. Theinitial parameters required to run the EM are obtained froma regression model adjusted using the least square method.The number of iterations of the EM algorithm is set to 10,and the sample size is taken to be at least 1/3 of the rectanglesarea RIR. These parameters will be used in the eye stateanalysis presented below.

4.3. Tracking. There are a number of reasons for using atracking module. The first is due to problems that wereencountered using the VJ during this research. Anotheris related with the necessity to track the face and eyescontinuously from frame to frame. A third reason is to reducethe search space thus satisfying the real-time conditionrequirement. The tracking process has been developed usingthe Condensation Algorithm (CA) in conjunction withNeural Networks (NNs) used for face tracking and withtemplate matching for eye tracking.

4.3.1. The Condensation Algorithm. This contribution imple-ments the CA that was proposed by Isard and Blake [24,

25] to track active contours using a stochastic approach.CA combines factored sampling (Monte-Carlo samplingmethod) with a dynamic model that is governed by the stateequation

Xt = f (Xt−1, ξt), (4)

where Xt is the state at instant t and f (·) is a nonlinearequation and depends on the previous state plus a whitenoise. The goal here is to estimate the state vector Xt with thehelp of system observation, which are the realization of thestochastic process Zt governed by the measurement equation

Zt = h(Xt,ηt

), (5)

where Zt is the measurement system at time t and h(·) is anonlinear equation that links the present state plus a whitenoise. The processes ξt and ηt are both white noise termsand are independent of each other. Also, these processesin general are non-Gaussian and multimodal. It must bepointed out that Xt is an unobservable underlying stochasticprocess.

4.3.2. Neural Networks. Neural networks are used in a widevariety of pattern recognition and classification problems.Figure 7 shows several face examples used to train thebackpropagation neural network used in this algorithm.

Before the neural network is trained, a two-part prepro-cessing step is necessary.


Viola-JonesNumber of detections: 8Frame Number: 1

t = 1

(a)


t = 10

(b)


t = 50

(c)


t = 89

(d)

Figure 9: The driver’s face has not been found by the Viola Jones method during several time instants.

020

4060

Time80

100 170190

VJ face positionTrue face position

210

x

230250

200195190185180175170165160155

y

Figure 10: Example where the VJ method does not find the driver’sface in a 100-frame sequence.

(i) Contrast modification using gamma correction givenby (6) with γ = 0.8 where this value has beendetermined experimentally [26]

J(x, y

) = I(x, y

)γ. (6)

(ii) Remove the contour points by means of a maskedAND operation shown in Figure 8(a).

Next, the characteristic vector which consists of the pixelgray-level values from the face image is extracted. The rate ofclassification following training is greater than 93%.

4.3.3. Face Tracking. Previously, it has been mentioned thatthe VJ method presents problems when detecting facesthat deviate from the nominal position and for differentorientations; thus, to correct this disadvantage a face trackingmethod has been developed. To demonstrate this shortcom-ing, Figure 9 shows several different time instants wherethe VJ method is not capable of finding the driver’s face.Figure 10 presents an extended example, where the trueposition and the VJ position are represented over a framesequence. The true position has been retrieved manually.

The main problem of the VJ method is that it is onlyable to locate the human face when it is positioned in frontof the camera. This drawback leads to an unreliable systemfor driver analysis throughout the driving process which ishighly dynamic, for example, when looking at the rearviewor wing mirrors. Much effort has gone into correcting thisproblem resulting in an efficient tracker which has beenimplemented using CA combined with a backpropagationneural network.

Through recursive probabilistic filtering of the incomingimage stream, the state vector

Xt =(xc, yc,uc, vc,w,h

)T ∈ R6 (7)

for a driver’s face is estimated for each time step t. It ischaracterized by its position, velocity, and size. Let (xc, yc)represent its center position, (uc, vc) its velocity in both x andy directions and (w,h) the size in pixels. The measurementvector is given by

Zt =(xc, yc,w,h

)T ∈ R4. (8)

The dynamics of the driver’s face has been modeled as asecond-order autoregressive process AR(2), according to

Xt = A2Xt−2 + A1Xt−1 + ξt , (9)

where A is the transition matrix proposed in [25], andξt represents the system perturbation at time t. The mostdifficult part of the CA is to evaluate the observation density

function. In this contribution to compute the weight π( j)t =

p(zt | xt = s( j)t ) for j = 1, . . . ,N , at time t, a neural network

value in the range of [0, 1] has been used; this provides anapproximation of the face and nonface in conjunction withthe distance and with respect to the face to track. This issimilar to the work performed by Satake and Shakunaga[27] who have used sparse template matching to compute

the weight π( j)t of the sample s

( j)t for j = 1, . . . ,N . In

this contribution, the neural network value is used as anapproximate value for the weights.

The density function of the initial state is p(x0) =N(z0,Σ0), where z0 is computed using the VJ method, and


(a) (b)

Figure 11: One time step of the Condensation algorithm (a) predicted region, (b) particles regions.

020

4060

Time80

100 170190

True positionEstimated position

210

x

230250

200195190185180175170165160155

y

Figure 12: Trajectory of the real and estimated face-center in a 100-frame sequence using the proposed tracker.

Table 1: Result of face tracking.

Driver Total frames Tracking failure Correct rate

D1 960 60 93.75%

D2 900 22 97.55%

D3 500 45 91.00%

D4 330 15 95.45%

D5 1400 50 96.42%

Σ0 is given in [4]. A particle representation at time t isshown in Figures 11(a) and 11(b) and Figure 12 depicts thetracking process in which the green circle is the true positionand the red cross characterizes a particle or hypothesis,whereas Figure 13 shows the probability over time. Thistracker is highly flexible as the neural network includesfaces and nonfaces for different head orientations undervarious illumination conditions. Table 1 presents furtherresults for several sequences of drivers faces. The sequencescome from the driver database, which was created to performthese experiments. The true face position has been retrievedmanually.

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 40 50 60 70 80 90 100

Time

Pro

babi

lity

Figure 13: Estimated value of the a posteriori density of the face-center in a 100-frame sequence using the proposed tracker; the faceis detected in the fourth frame.

4.3.4. Eye Tracking. For this task, the state of the eye ischaracterized by its position and velocity over the image. Let(x, y) represent the eye pixel position at time t and (u, v) beits velocity at time t in the x and y directions, respectively.The state vector at time t can, therefore, be represented by

Xt =(x, y,u, v

)T ∈ R4. (10)

The transition model is given by (11) which is a firstautoregressive model AR(1)

Xt = AXt−1 + ξt. (11)

The evaluation of the observation density function isdeveloped by a template matching strategy [27] that wastruncated to reduce false detections. CA is initialized whenthe eyes are detected with the method described in theprevious section plus a white noise, that is, similar to the facetracking case. Figure 14 depicts the eye trajectory trackingand Figure 15 shows the compute value of the a posterioridensity function of each eye, both on a sequence of 100images. Table 2 shows the eye tracking results that have


250300

350400

x 450500 0

2060

Time

80 100

True right eyeTrue left eye

Estimated right eyeEstimated left eye

200

195

190

185

180

175

170

y

Figure 14: Trajectory of the real and estimated eyes center in a 100-frame sequence.

Table 2: Result of eye tracking.

Driver Total frames Tracking failure Correct rate

D1 960 20 97.91%

D2 900 30 96.60%

D3 500 8 98.40%

D4 330 14 95.75%

D5 1400 90 93.57%

been obtained from several sequences of images during theexperiments.

4.4. Eye State Detection. To identify drowsiness using eyeanalysis it is necessary to know its state, that is, openor closed, through the time and to develop an analysisover time, that is, to measure the time that has passedfor each state. Classification of the open and closed stateis complex due to the changes in the shape of the eye,among other factors, the changing position and the facerotations and variations of twinkling and illumination. Allof these problems make it difficult to reliably analyze the eye.To overcome these shortcomings a supervised classificationmethod has been used, more specifically, a Support VectorMachine (SVM). Figure 16 presents the proposed scheme foreye state verification.

4.4.1. Support Vector Machine. SVM classification [28–30] isrooted in statistical learning theory and pattern classifiers; ituses a training set, S = {(xi, yi) : i = 1, . . . ,m}, where xi is thecharacteristic vector in Rn, yi ∈ {1, 2} represents the class, inthis case 1 for open eyes and 2 for closed eyes, and m is thenumber of elements of S. From a training set, a hyperplane isbuilt that permits classification between two different classesand minimizes the empirical risk function [30].

Mathematically, SVM consists in finding the best solu-tion to the following optimization problem:

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 40 50 60 70 80 90 100

Time

Pro

babi

lity

Right eyeLeft eye

Figure 15: Estimated value of the a posteriori density of the eyecenter in a 100-frame for right and left eyes; the eyes are detected inthe fourth frame.

SVM Open or closed?

Figure 16: SVM scheme for eye state verification.

minα

f (α) = 12αTQα− eTα

s.t. 0 ≤ αi ≤ C, i = 1, . . . ,m

yTα = 0,

(12)

where e is an m by the 1 vector, C is an upper bound, Q isan m by m matrix with Qij = yi y jK(xi, xj), and K(xi, xj)is the kernel function. By solving the above quadraticprogramming problem, the SVM tries to maximize themargin between the data points in the two classes and to min-imize the training errors simultaneously. Figure 17 depictsinput space mapping to a high-dimensional feature spacethrough a nonlinear transformation and its maximizationprocess.

4.4.2. Eye Characteristic Extraction Using a Gabor Filter.The Gabor filter was used by Daugman for image analysis,changing the orientation, and scale [18] where these are


Nonlineartransformation Hiperplane

Feature spaceInput space

Hiperplane

Class 1Class 2

Figure 17: SVM representation.

multiscale and multiorientation kernels. They can be definedby the complex function

g(x, y, θ,φ

)

= exp

(−x2 + y2

σ2

)exp

(i2πθ

(x cos

(φ)

+ y sin(φ)))

,

(13)

where θ and φ are the scale and orientation parameters andσ is the standard deviation of the Gaussian kernel whichdepends on the spatial frequency to be measured, that is,θ. The response of the Gabor filter to an image is obtainedfrom a 2D convolution operation. Letting I(x, y) representthe image and G(x, y, θ,φ) denote the response of a Gaborfilter with scale θ and orientation φ to an image at point (x, y)on the image plane by G(·) is obtained using

G(x, y, θ,φ

) =∫∫

I(p, q

)g(x − p, y − q, θ,φ

)dp dq. (14)

Various combinations of scales and orientations are morerobust for the classification between open and closed eyes.Three scales and four orientations have been used to generateFigure 18; these are {1, 2, 3} and {0,π/4,π/2, 3π/4} whichwere obtained experimentally over an image size 30 by 20.

Once the response of the Gabor filter is obtained, theeye characteristic vector is extracted using a subwindowprocedure described by Chen and Kubo [31] and denotedby d ∈ R360. This vector is computed using (15) for eachsubwindow of size 5 by 6. Figure 19 shows the subwindowdiagram

dθ,φi = 1

30

∑y=1:5

∑x=1:6

G(x, y, θ,φ

)i = 1, . . . , 20. (15)

To perform this analysis a training set has been builtwhich consists of both open and closed eyes. The imagescome from diverse sources, under different illuminationconditions and are from different races. A further importantaspect of this eye database is that it contains images ofdifferent eye colors, that is, blue, black, and green. Figure 20shows several examples of this database.

Previous to SVM training, it is crucial to preprocess eachimage where this procedure involves histogram equalization,

Table 3: Result of eye state analysis.

Driver Total frames Eyes open Eyes closed Correct rate

D1 960 690/700 258/260 98.90%

D2 900 520/560 339/340 96.27%

D3 500 388/400 99/100 98.00%

D4 330 150/170 152/160 91.61%

D5 1400 891/980 401/420 93.19%

filtering using a median filter, followed by the sharpen filter.The median filter is used to reduce image noise, and thesharpen filter enhances the borders.

The main objective of the SVM training is to obtain thebest parameters and the best kernel that minimizes (5). Afterseveral SVM training experiments, it was decided to use theRBF kernel, that is, K(xi, xj) is exp(−γ‖xi − xj‖2), C = 30,and γ = 0.0128, where these parameters achieve a hightraining classification rate of close to 93%.

Table 3 presents results using this method computed forseveral sequences of drivers. It demonstrates a high correctrate of classifications.

4.5. Drowsiness Index. The eye-blinking frequency is anindicator that allows the level of driver drowsiness (fatigue)to be measured. As in the works of Horng et al. [12]and Dong and Wu [3], if for five consecutive frames orduring 0.25 seconds the eye is identified as being closed thesystem issues an alarm cue, PERCLOS [7], which is alsoimplemented in this system.

Figure 21 presents an instantaneous result for this systemon a driver’s image whereas Figure 22 pictures the drowsinessevolution index graph for a driver’s drowsiness sequence.Additionally, other examples are show in Figure 25.

4.6. Distraction. It is estimated that 20% of traffic accidentsare caused by driver distraction [1]. To detect this char-acteristic the driver’s face should be studied because thepose of the face contains information about one’s attention,gaze, and level of fatigue [4]. To verify driver distraction thefollowing procedure has been implemented.


(a) (b)

Figure 18: Gabor filter for θ = {0, 1, 2} and φ = {0,π/4,π/2, 3π/4}.

Figure 19: Subwindow images from the Gabor filter.

Figure 20: Examples from the eye database for daylight illumina-tion.

4.6.1. Face Orientation. The driver’s face orientation isestimated using the eye position, with

θ = tan−1

(Δx

Δy

), (16)

Open

Alert

Figure 21: Day system instantaneous result.

0

20

40

60

80

100

Perc

enta

ge

0 200100 300 400 500 600 700 800 900

Time

(a)

0

20

40

60

80

100

Perc

enta

ge

0 200100 300 400 500 600 700 800 900

Time

(b)

Figure 22: Drowsiness index graph for a 900-frame sequence of adrowsy driver: (a) Perclos, (b) Horng-Dong and Wu index.


X

Y

θ

Figure 23: Face orientation.

−20−15−10−5

05

101520

An

gle

0 2010 30 40 50 60 70 80 90 100

Time

Figure 24: Head-orientation monitoring over time in a 100-framesequence.

where Δx = x2 − x1, Δy = y2 − y1, (x1, y1), and (x2, y2)correspond to the left and right eye positions, respectively.Equation (17) provides the classification limits. Figure 23depicts an example of the face orientation and in Figure 24 anextended example is shown for the driver’s face orientationusing information from the eyes for a sequence of images

Left if θ > 8◦,

Front if |θ| ≤ 8◦,

Right if θ < −8◦.

(17)

4.6.2. Head Tilt. The method described above presentsproblems when a monocular camera is used, and so, toovercome this drawback, this contribution has implementeda head-tilt based on neural networks. Keeping in mind thatthe driver face database is composed of face examples forfive different orientations, the face is passed to the neuralnetwork to determine its orientation, specifically for the upand down cases. If the system detects that the face positionis not looking straight on, an alarm cue is issued to alert thedriver of a danger situation.

5. Night System Design

In this part of the work, the night system will be described,where this is based on the algorithm scheme shown inFigure 1(b). Note that it is composed of both software andhardware platforms. The main difference between this andthe previous system is in perception system.

5.1. Perception System. The perception system is used toacquire visual information of the driver’s head duringnocturnal driving. It consists of three parts: a vision system,a synchronization system, and an illumination system; seeFigure 26.

The first part is composed of a miniature CCD camerawhere the IR filter has been removed. This camera generatesa composite video signal (PAL, phase alternating line). Thesecond part takes the signal from the camera and splits itup into odd and even fields using the LM1881N video syncseparator. In the third and last part, the illumination system,based on near-infrared (NIR) light (700–900 nm), containstwo sets of NIR light-emitting diode (LED) rings (inner andouter) which produce odd and even image fields where thepupil effect is highlighted [3, 4]. The inner led ring surroundsthe camera, while the outer leds are placed over a ruler in asymmetric position around the camera. Figure 27(a) showsthis system, and Figure 27(b) provides an example of anilluminated driver’s face from the NIR illuminator.

Each frame is deinterlaced in both odd and even fieldswhich contain the dark and bright pupil images, separately.Hence, the height of the odd and even image fields is amedium of the original image; this procedure can be seen inFigure 27(c): the top photograph is the even image field, andthe bottom is the odd image field. Even an odd image will beused later on for eye detection.

5.2. Eye Detection. The bright effect pupil is the mainprinciple behind locating the position of the eye. To do thissense, three images have been generated from the initialdriver image, these are the difference image (ID), the edgeimage (IG), and the bright part of the fast radial symmetrytransform (FRST) image [32] (IF).

The first image is computed using the absolute value ofthe pixel difference between the even and odd image fields;see (18). In this image most of the background and externallight illumination has been removed, and the pupils appearas the brightest part

ID(x, y

) = ∣∣IO(x, y)− IE

(x, y

)∣∣. (18)

The second and third images are obtained using the SobelOperator and the bright part of the FRST [32] over thedifference image, respectively; these may be seen in Figures28(a), 28(b), and 28(c)

IG =√‖I ∗ SX‖2 + ‖I ∗ SY‖2, (19)

where Sx and Sy are given by (5).Most researchers only make use of the difference image

for pupil detection; however, in real driving conditions, this


Figure 25: Different stages of the proposed algorithm during daytime conditions and for several time instants, driving conditions anddifferent drivers.

Synchronization system(split up in odd and even fields, LM1881N)

Vision system(CCD camera)

Illumination system(outer and inner rings of leds)

Figure 26: Perception system schema.

image deteriorates due to external illumination, vibrations,and so forth and is also very sensitive to lighting conditions.In such circumstances, it is necessary to incorporate morerobust information to improve the detection step. There-fore, in this paper, the edge and FRST images have beenimplemented to obtain enhanced results considering theaforementioned drawbacks.

Once all the images used to detect the eyes have beenspecified, the next step is to compute a binary thresholdfor the difference, edge, and FRST images. In the first ofthese, the threshold is obtained from a systematic analysisof its histogram, where two groups are formed. In thesecond case, the histogram is modelled using a Gammadistribution function where the 90% cumulative interval

provides the threshold. Finally, in the third image, themaximum histogram value produces the required thresholdlevel. This yields three binary images consisting of binaryblobs that may contain a pupil.

The pupils are detected by searching within the entireimage for the location of two blobs that satisfy a particularsize, shape, and distance constraints. To remove false blobsan unsupervised classifier has been implemented, in thiscase, the SMV classifier, which is based on statistical learningtheory and pattern classifiers. It uses a training set, S ={(xi, yi) : i = 1, . . . ,m}, where xi is the characteristicvector in Rn, yi ∈ {1, 2} represents the class, in this case1 for open eyes and 2 for noneyes, and m is the numberof elements of S [28]. Before training the SVM, it is crucialto preprocess each image, where this procedure involveshistogram equalization, filtering using the median filter,followed by the sharpen filter. The median filter is used toreduce noise in the image, and the sharpen filter is usedto enhance the borders. After this procedure a Gabor filter[31] is computed on each image of the eye database. Somecombinations of scales and orientations are more robustfor the classification between eyes and noneyes. One scaleand four orientations have been used in this research; theseare {0} and {0,π/4,π/2, 3π/4} which have been obtainedexperimentally from an image size of 30 by 20 (in pixels). Toperform this task, a training set has been built which consistsof open eyes and noneyes, where an example is shown inFigure 29. Once the response of the Gabor filter is obtained,


(a) (b) (c)

Figure 27: Perception system using infrared illumination (a) vision system, (b) drivers image, and (c) odd (top) and even (bottom) imagefields.

(a) (b) (c)

(d) (e) (f)

Figure 28: Difference (a), edge (b), and FRST (c) and their binary images (d), (e) and (f), respectively.

the eye characteristic vector is extracted and denoted by d ∈R600.

After several SVM training experiments, it was decidedto use the RBF kernel, that is, K(xi, xj) is exp(−γ‖xi −xj‖2), C = 100, and γ = 0.128; these parameters achieve ahigh training classification rate of about 93%.

This exhaustive search detects the pair of eyes; this isfollowed by an ellipse fitting which is applied to each pupil,and the center of the resulting ellipse is the position of thedetected pupil. This process is presented in Figure 30(b).These will be used later to initialize the eye tracker.

5.3. Face Detection. Once the eyes have been located, thesystem continues with driver face detection. To perform thistask, a human face model has been developed considering thepupil’s position and face anthropometric properties [19]. Letp1(x, y) and p2(x, y) be the center position of the right andleft eye, respectively, and dRL their distance (in pixels); thearea of the face is obtained from the following equations:

dRL =√

(x1 − x2)2 +(y1 − y2

)2,

θ = tan−1(y2 − y1

x2 − x1

),

R1 = 1.5dRL, (20)

R2 = dRL,

pc(x, y

) = (x1 + 0.5dRL, y1 + 0.3dRL

),

where pc(x, y) is the centre of the face, and R1 and R2 are theaxes of the face ellipse. Figure 30 depicts this model and itsresult.

5.4. Tracking. The tracking process has been developed usingthe Condensation algorithm for face and eye tracking.

5.4.1. Face Tracking. This system uses the neural networks[33] and models and parameters proposed previously forface tracking. Now that the elements for tracking have been


(1) (2)

(3)

(a)

(1) (2)

(3)

(b)

Figure 29: Gabor filter response for θ = {0} and φ = {0,π/4,π/2, 3π/4}.

dRL

p1(x, y)

pc(x, y)p2(x, y)

θ

R2

R1

Minor radius

Major radius

(a) (b)

Figure 30: (a) Human face model using the eye position, (b) face and eye detection.

identified, only the backpropagation neural network trainingis developed in this section. Before training, a preprocessingstep that consists of using a Gabor filter [18] with one scaleand two orientations has been implemented. After this, thecharacteristic vector which consists of gray-level values ofpixels coming from the face image is extracted. The rate ofclassification subsequent to the training is more than 92%.Figure 31 shows several face examples, and Table 4 presentsexperimental results. In this table, the true position has beenretrieved manually.

The density function of the initial state is p(x0) =N(z0,Σ0), where z0 is computed using the previous facedetection method, and Σ0 is given in [4]. Figure 32 shows thea posterior density function of the face center tracking.

5.4.2. Eye Tracking. For this task, the state of the eye ischaracterized by its position and velocity. These parametersare also described in the diurnal system. To evaluate theprobability observation density, a triangular density functionbased on the value of the difference image has been used(Figure 33). This function takes into account the gray-levelvalue for the intensity of the illumination system.

CA is initialized when the eyes are detected from themethod described in the previous section plus a white noise.In Table 4 the eye tracking results are presented, which havebeen carried out in several sequences of images. Again,the true position has been retrieved manually. Additionally,Figure 34 shows the a posterior density function of both eyes.

To evaluate the probability observation density, a trian-gular density function based on a value from the differenceimage has been used. CA is initialized when the eyes aredetected using the method described in the previous sectionplus a white noise. Table 4 shows the eye tracking resultswhich have been obtained from several image sequences.

5.5. Eye State Detection and Drowsiness Index. To identifydrowsiness from an eye analysis, knowledge of the eye’s stateis required, that is, open or closed, in time and to developan analysis over large periods of time, that is, to measure thetime spent in each state. Classification of the open and closedstate is complex due to changes in the shape of the eye, thechanging position, and face rotations, as well as variationsin twinkling and illumination, and so forth. All of thesefactors make it difficult to reliably analyze the eyes. However,when using the edge and FRST images, the eye state may becomputed satisfactorily.

The PERCLOS [7] has been implemented in this system.Figure 35 presents an instantaneous result of this systemobtained from a driver’s image, and in Figure 36 theevolution of the drowsiness index graph for a sequence ofdriver drowsiness is presented.

5.6. Distraction. This method is similar to the previous case,once the face is continuously located in time; a neuralnetwork is used to determine its orientation and to verify thedriver’s level of distraction. If the system detects that the face


(a) (b) (c)

(d) (e) (f)

Figure 31: Examples of the face database and its Gabor filter response.

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 50 100 150 200 250 300 350

Time

Pro

babi

lity

Figure 32: Estimated value of the a posteriori density of the face-center in a 350-frame sequence using the proposed tracker; the faceis detected in the fourth frame.

0 255

1(255,1)

Figure 33: Triangular density function.

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0 50 100 150 200 250 300 350

Time

Right eyeLeft eye

Pro

babi

lity

Figure 34: Estimated value of the a posteriori density of the eyecenter in a 350-frame for right and left eyes, the eyes are detected inthe fourth frame.

position is not facing forward, an alarm cue is issued to alertthe driver of a danger situation.

6. Conclusions

In this paper, a research project to develop a nonintrusiveand autonomous driver drowsiness system based on Com-puter Vision and Artificial Intelligence has been presented.This system uses advanced technologies which analyze and


CloseAwake

Drowsiness

Time: 9.011 (seg)Closed percentage: 39.286

Open percentage: 60.714Perclos: 30

Closed frames: 33Open frames: 51Total frames: 64

Figure 35: System instantaneous result.

Table 4: Results of face and eye tracking and eye state analysis.

Driver Total framesFace tracking Eye tracking Eye state

Tracking failure Correct rate Tracking failure Correct rate Eyes open Eyes closed Correct rate

D1 800 40 95.00% 14 98.25% 690/700 97/100 97.78%

D2 646 29 95.51% 29 95.51% 500/530 100/116 90.27%

D3 600 20 96.67% 24 96.00% 345/244 206/226 91.69%

020406080

1 43 85 127 169 211 253 295 337 379 421

Time

Perc

enta

ge

Figure 36: Drowsiness index graph in a 341-frame sequence of adrowsy driver.

monitor the state of the driver’s eye in real-time and forreal driving conditions; this is driving conditions for bothdaytime and nocturnal situations.

In the first case, based on the results presented in Tables1, 2, and 3, the algorithm proposed for eye detection,face tracking, and eye tracking is shown to be robust andaccurate for varying light, external illumination interference,vibrations, changing backgrounds, and facial orientations. Inthe second case, and as presented in the results of Table 4, thesystem is also observed to provide agreeable results.

To acquire the data required to develop and test thealgorithms presented in this paper, several drivers have beenrecruited and were exposed to a wide variety of difficultsituations commonly encountered on roadways, for bothdaytime and nocturnal conditions. This guarantees andconfirms that the experiments presented here are proven tobe robust and efficient for real traffic scenarios. The imageswere taken using two cameras within the IVVI (IntelligentVehicle based on Visual Information) vehicle (Figure 27(a)):a pin-hole analog camera connected to a frame-grabber forthe nocturnal illumination and a fire-wire camera for thediurnal use. Besides that, the hardware processes 4-5 frames

(a)

PC1 PC2 PC3

(b)

Figure 37: (a) IVVI vehicle, (b) processing system.

per second using an Intel Pentium D, with 3.2 GHz, 2 GB.RAM memory and MS Windows XP.

IVVI is an experimental platform used to develop thedriver assistance systems for real-life driving conditions. Themost up to date version is a Nissan-Note car; see Figure 37(a).


This vehicle is equipped with a processing system, whichprocesses the information coming from the cameras. Theprocessing system is composed of three personal computers(Figure 37(b)).

For future work, the objective will be to reduce thepercentage error, that is, reduce the amount of false alarms; toachieve this, additional experiments will be developed, usingadditional drivers and incorporating new analysis modules,for example, facial expressions.

Acknowledgments

This paper was supported in part by the Spanish Governmentthrough the CICYT projects VISVIA (Grant TRA2007-67786-C02-02) and POCIMA (Grant TRA2007-67374-C02-01).

References

[1] T. Brandt, R. Stemmer, B. Mertsching, and A. Rakotonirainy,“Affordable visual driver monitoring system for fatigue andmonotony,” in Proceedings of the IEEE International Conferenceon Systems, Man and Cybernetics (SMC ’04), vol. 7, pp. 6451–6456, October 2004.

[2] Z. Tian and H. Qin, “Real-time driver’s eye state detection,” inProceedings of the IEEE International Conference on VehicularElectronics and Safety, pp. 285–289, October 2005.

[3] W. Dong and X. Wu, “Driver fatigue detection based on thedistance of eyelid,” in Proceedings of the IEEE InternationalWorkshop on VLSI Design and Video Technology (IWVDVT’05), pp. 397–400, Suzhou-China, May 2005.

[4] Q. Ji and X. Yang, “Real-time eye, gaze, and face pose trackingfor monitoring driver vigilance,” Real-Time Imaging, vol. 8, no.5, pp. 357–377, 2002.

[5] L. M. Bergasa, J. Nuevo, M. A. Sotelo, and M. Vazquez, “Real-time system for monitoring driver vigilance,” in Proceedings ofthe IEEE Intelligent Vehicles Symposium, pp. 78–83, June 2004.

[6] L. Fletcher, L. Petersson, and A. Zelinsky, “Driver assistancesystems based on vision in and out of vehicles,” in Proceedingsof the IEEE Symposium on Intelligent Vehicles, pp. 322–327,2003.

[7] NHTSA, “Evaluation of techniques for ocular measurement asan index of fatigue and the basis for alertness management,”Tech. Rep. DOT HS 808762, National Highway Traffic SafetyAdministration, Washington, DC, USA, 1998.

[8] L. Hagenmeyer, Development of a multimodal, univer-sal human-machine-interface for hypovigilance-management-systems, Ph.D. thesis, University of Stuttgart, Stuttgart, Ger-many, 2007.

[9] G. Longhurst, Understanding Driver Visual Behaviour, SeeingMachine, Canberra, Australia.

[10] J. M. Armingol, A. de la Escalera, C. Hilario et al., “IVVI:intelligent vehicle based on visual information,” Robotics andAutonomous Systems, vol. 55, no. 12, pp. 904–916, 2007.

[11] T. D’Orazio, M. Leo, and A. Distante, “Eye detection in facesimages for a driver vigilante system,” in Proceedings of the IEEEIntelligent Vehicles Symposium, pp. 14–17, University of Parma,Parma, Italy, June 2004.

[12] W.-B. Horng, C.-Y. Chen, Y. Chang, and C.-H. Fan, “Driverfatigue detection based on eye tracking and dynamic templatematching,” in Proceedings of the IEEE International Conferenceon Networking, Sensing and Control, pp. 7–12, March 2004.

[13] P. Viola and M. Jones, “Rapid object detection using aboosted cascade of simple features,” in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition, pp. 511–518, December 2001.

[14] A. Branzan, B. Widsten, T. Wang, J. Lan, and J. Mah, “Acomputer vision-based system for real-time detection of sleeponset in fatigued drivers,” in Proceedings of the IEEE IntelligentVehicles Symposium (IV ’08), pp. 25–30, June 2008.

[15] Q. Ji, Z. Zhu, and P. Lan, “Real-time nonintrusive monitoringand prediction of driver fatigue,” IEEE Transactions on Vehicu-lar Technology, vol. 53, no. 4, pp. 1052–1068, 2004.

[16] J. G. Brookshear, Theory of Computation: Formal Languages,Automata and Complexity, Addison Wesley Iberoamericana,1993.

[17] R. Grace, “Drowsy driver monitor and warning system,”in Proceedings of Driving Assessment International DrivingSymposium on Human Factors in Driver Assessment, Training,and Vehicle Design, 2001.

[18] J. G. Daugman, “Uncertainty relation for resolution inspace, spatial frequency, and orientation optimized by two-dimensional visual cortical filters,” Journal of the OpticalSociety of America, vol. 2, no. 7, pp. 1160–1169, 1985.

[19] P. Gejgus and M. Sparka, Face Tracking in Color VideoSequences, The Association for Computing Machinery, 2003.

[20] N. Otsu, “A threshold selection method from gray-level his-tograms,” IEEE Transactions on Systems, Man and Cybernetics,vol. 9, pp. 62–66, 1979.

[21] I. Jafar and H. Ying, “A new method for image contrastenhancement based on automatic specification of localhistograms,” International Journal of Computer Science andNetwork Security, vol. 7, no. 7, 2007.

[22] Y. Wu, H. Liu, and H. Zha, “A new method of detecting humaneyelids based on deformable templates,” in Proceedings of theIEEE International Conference on Systems, Man and Cybernetics(SMC ’04), pp. 604–609, October 2004.

[23] G. J. McLachlan, The EM Algorithm and Extensions, John Wiley& Sons, New York, NY, USA, 1997.

[24] M. Isard and A. Blake, “Condensation: conditional densitypropagation for visual tracking,” International Journal ofComputer Vision, vol. 29, no. 1, pp. 5–28, 1998.

[25] M. A. Isard, Visual motion analysis by probabilistic propagationof conditional density, Ph.D. thesis, Oxford University, Oxford,UK, 1998.

[26] J. R. Parker, Practical Computer Vision Using C, John Wiley &Sons, New York, NY, USA, 1994.

[27] J. Satake and T. Shakunaga, “Multiple target tracking byappearance-based condensation tracker using structure infor-mation,” in Proceedings of the 17th International Conference onPattern Recognition (ICPR ’04), vol. 3, pp. 294–297, August2004.

[28] N. Cristianini and J. Shawe-Taylor, An Introduction to SupportVector Machines and Other Kernel-Based Learning Methods,Cambridge University Press, Cambridge, UK, 2000.

[29] C. Chang and C. Lin, “LIBSVM: a library for support vectormachine,” 2001, http://www.csie.ntu.edu.tw/∼cjlin/libsvm.

[30] I. Guyon, S. Gunn, M. Nikravesh, and L. A. Zadeh, FeatureExtraction: Foundations and Applications, Springer, Berlin,Germany, 2006.

[31] Y.-W. Chen and K. Kubo, “A robust eye detection andtracking technique using gabor filters,” in Proceedings of the 3rdInternational Conference on Intelligent Information Hiding andMultimedia Signal Processing (IIHMSP ’07), vol. 1, pp. 109–112, November 2007.


[32] G. Loy and A. Zelinsky, “Fast radial symmetry for detectingpoints of interest,” IEEE Transactions on Pattern Analysis andMachine Intelligence, vol. 25, no. 8, pp. 959–973, 2003.

[33] C. G. Looney, Pattern Recognition Using Neural Networks,Theory and Algorithms for Engineers and Scientists, OxfordUniversity Press, Oxford, UK, 1997.


Research Article

Multiobjective Reinforcement Learning forTraffic Signal Control Using Vehicular Ad Hoc Network

Duan Houli, Li Zhiheng, and Zhang Yi

Department of Automation, Tsinghua University, Beijing 100084, China

Correspondence should be addressed to Duan Houli, [email protected]

Received 1 December 2009; Accepted 5 September 2010


Copyright © 2010 Duan Houli et al. This is an open access article distributed under the Creative Commons Attribution License,which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

We propose a new multiobjective control algorithm based on reinforcement learning for urban traffic signal control, named multi-RL. A multiagent structure is used to describe the traffic system. A vehicular ad hoc network is used for the data exchange amongagents. A reinforcement learning algorithm is applied to predict the overall value of the optimization objective given vehicles’ states.The policy which minimizes the cumulative value of the optimization objective is regarded as the optimal one. In order to makethe method adaptive to various traffic conditions, we also introduce a multiobjective control scheme in which the optimizationobjective is selected adaptively to real-time traffic states. The optimization objectives include the vehicle stops, the average waitingtime, and the maximum queue length of the next intersection. In addition, we also accommodate a priority control to the busesand the emergency vehicles through our model. The simulation results indicated that our algorithm could perform more efficientlythan traditional traffic light control methods.

1. Introduction

Increasing traffic congestion over the road networks makesthe development of more intelligent and efficient trafficcontrol systems an urgent and important requirement. How-ever, traffic systems are typically complex large-scale systemsconsisting of a great number of interacting participants. Itis very difficult to use traditional control algorithms to getsatisfied control effect. Thus, various intelligent algorithmshave been used in attempts to build an efficient traffic controlsystem, such as fuzzy control technologies [1, 2], artificialneural networks [3, 4], and genetic algorithms [5, 6], whichgreatly improve the efficiency of urban traffic signal controlsystems.

Reinforcement learning is a category of machine learningalgorithms including Q learning, temporal difference, andSARSA algorithm [7–9]. Reinforcement learning is to learnthe optimal policy by a trial-and-error process includingperceiving states from the environment, choosing an actionaccording to current states and receiving rewards from theenvironment. The policy which maximizes the expected

long-term cumulative reward is considered as the optimalone. Reinforcement learning is a self-learning algorithmwhich does not need an explicit model of the environment.Thus, it can be applied in traffic signal control effectivelyto respond to the constant changes of traffic flow andoutperform traditional traffic control algorithms. Thorpestudied reinforcement learning for traffic light control in1997. He used a neural network to predict the waitingtime for all cars standing at the intersection and selectedthe best control policy using the SARSA algorithm [10].Abdulhai et al. presented a basic framework of applyingQ-learning to traffic signal control and got encouragingresults while applying it to an isolated intersection [11].Mikami and Kakazu combined the evolutionary algorithmand reinforcement learning for coordination traffic signalcontrol [12]. However, the above methods use traffic-light-based value functions, which means that the state space istoo large to handle. Therefore, these methods suffer fromthe “dimension curse” and achieve limited success whenapplied to large-scale road networks. Wiering et al. utilizeda car-based value function to solve this problem [13, 14].


They predicted each car’s total expected waiting time untilit arrived its destination given possible choices of relatedtraffic lights using reinforcement learning, and chose theaction which minimized the summed waiting time of allcars in the network. This method effectively reduces thestate space and thus can be applied to large-network control.Experiments in a network with 12 edge nodes and 16junctions proved the effectiveness of this method.

However, Wiering’s method uses the total waiting timeas the optimization goal which is mainly suitable for themedium traffic condition. In practical traffic systems, weshould consider different optimization objectives adaptive todifferent traffic situations, called the multiobjective controlscheme in this paper. Under the free traffic condition, theaverage vehicle speed is high and the average waiting timeis short, so the waiting time is not the focal point, whilethe vehicle stops will increase the vehicle emission and oilconsumption. Therefore, we should try to minimize theoverall vehicle stops in the network. Under the mediumtraffic condition, the overall waiting time is regarded as theoptimization goal because most drivers want to arrive attheir destinations as soon as possible. Under the congestedtraffic situation, queue spillovers must be avoided to keepthe network from large-scale congestion, thus, the queuelength must be regarded as the control goal [15]. Since themultiobjective control scheme can adapt to various trafficconditions and make a more intelligent control system, wepropose a multiobjective control strategy based on Wiering’smodel. In our model, data exchanges among vehicles androadside equipments are necessary. Thus, a vehicular ad hocnetwork is utilized to build a wireless traffic informationsystem.

This paper is organized as follows: in Section 2, we willintroduce how to model the road network with an agent-based structure; Section 3 describes how to exchange trafficdata using the ad hoc network; in Section 4, a multiagenttraffic control strategy using reinforcement learning is pro-posed; in Section 5, the proposed method is applied to a roadnetwork with 7 intersections to prove its effectiveness; finally,in Section 6, we draw the conclusion of this paper.

2. Agent-Based Model of Traffic System

We use an agent-based model to describe the practical trafficsystem. Vehicles and traffic signal controllers in the roadnetwork are regarded as two types of agents. Data will beexchanged among these agents. A typical road network isbuilt based on Wiering’s model [14] as shown in Figure 1.There are six possible settings for each traffic controllerto prevent accidents: two traffic lights from opposingdirections allow cars to go straight ahead or to turn right(2 possibilities), two traffic lights in the same direction ofthe intersection allow the cars from there to go straightahead, turn right, or turn left (4 possibilities). Road lanesare discretized into a number of cells at each traffic light.The capacity of each road lane is determined accordingto its practical length. At each time step, new cars withparticular destinations are generated and enter the network

from outside. After new cars have been added, traffic lightdecisions are made and each car moves to the subsequentcell if it is not occupied or the car’s predecessor is movedforward. All vehicles are assumed to have the same speedin this system. Thus, each car is at a specific traffic node(node), a direction at the node (dir), a position in the queue(place), and has a particular destination (des). Thus, wecan use [node, dir, place, des] ([n, d, p, des] for short) todenote the state of each vehicle [13]. Vehicles follow theshortest path through the road network to their destinations.As mentioned before, a multiobjective control scheme isadopted in this method. The optimization objectives includethe total waiting time, vehicle stops, and the queue length,which will be chosen adaptively to the traffic condition. Weuse Q([n, d, p, des], action) to denote the total expected valueof the optimization objective for each car until it arrives atthe destination given its current node, direction, place andthe decision of the light. The optimal action of a node j isdetermined by the following formulation:

Aoptj = arg max

Aj

∑i∈Aj

∑(n,d,p,des) ∈ queuei

Q([

n, d, p, des], red

)

−Q([

n, d, p, des], green

).

(1)

It should be noticed that Q([n, d, p, des], action) here doesnot only refer to the total waiting time but also refer tovehicle stops or queue lengths, according to the real-timetraffic states. This is the most important difference betweenour model and Wiering’s model, which will be explained indetail in Section 4.

3. Traffic Information Exchange System UsingVehicular Ad Hoc Network

We need to exchange a lot of information during the signalcontrol process. Thus, a wireless traffic information exchangesystem based on a vehicular ad hoc network is built toexchange data among the vehicles and signal controllers.An illustration of such information exchange system isshowed in Figure 2. It is assumed that all vehicles in thenetwork are intelligent ones equipped with Vehicular AdHoc Network communication devices, so that they havethe ability of communicating with other vehicles and theroadside controllers. Thus, all necessary information can becollected through the intercommunication of vehicles andcontrollers. The data to be collected include the followings:

(a) traffic flow through each intersection within eachtime step;

(b) queue length at each traffic light within each timestep;

(c) type of each vehicle (car, bus, or emergent vehicle);

(d) destination of each vehicle;

(e) node where each vehicle stands at;

(f) direction each vehicle moving towards;


Figure 1: Agent-based traffic model illustration.

Wirelessnetwork

ControllerTraffic control

center

Figure 2: Illustration of traffic information exchange system.

(g) position in the queue where each vehicle stands at;

(h) total waiting time each vehicle used to pass throughthe network;

(i) total number of stops each vehicle used to passthrough the network.

4. Multiobjective Control Algorithm Based onReinforcement Learning (Multi-RL)

We extend Wiering’s algorithm to a multiobjective schemeby selecting the optimization objective according to the real-time traffic condition. In addition, it is assumed that somespecial vehicles such as buses and ambulances need a prioritycontrol, and thus they should be considered separately.

The multiobjective control algorithm considers threetypes of traffic conditions as follows. The method to estimate

traffic conditions should be defined carefully according to theactual situation of the road network.

4.1. Free Traffic Condition. Under this condition, we aim tominimize the number of stops, in other words, we expect tohave the vehicles pass through the network with the feweststops. Thus, the cumulative number of stops is selected asthe optimization objective.

The number of stops will increase when a vehiclemoving to a green light at current time step meets a redlight at the next time step. Therefore, we denote Q([node,dir, pos, des],L) as the expected cumulative number of stopswhile V([node, dir, pos, des]) denotes the number of stops(without knowing the traffic light decision) for a car at[node, dir, pos] until it reaches its destination. The iterativeformulation of Q([node, dir, pos, des],L) is shown as follows:

Q([

node, dir, pos, des],L)

=∑

(node′, dir′, pos′,L,L′)

P(L′ | [node, dir, pos, des

],L,

[node′, dir′, pos′, des

])× (R([node, dir, pos, des

],[node′, dir′, pos′, des

])+γV

([node′, dir′, pos′, des′

])),

V([

node, dir, pos, des])

=∑L

P(L | [node, dir, pos, des

])Q([

node, dir, pos, des],L),

(2)


where [node′, dir′, pos′, des] means the state of a vehicle atthe next time step; L is the action of the traffic light atthe current time step, while L′ is the action of the trafficlight at the next time step. P(L′ | [node, dir, pos, des], L,[node′, dir′, pos′, des]) gives the probability that the trafficlight turns L′ at the next time step given the current stateand the next state of this vehicle; R([node, dir, pos, des],[node′, dir′, pos′, des]) is a reward function as follows: if L =Green, L′ = Red, which means the vehicle moving to a greenlight at the current time step meets a red light at the next timestep, then the number of vehicle stops will increase, R = 1;otherwise, R = 0; γ is the discount factor (0 < γ < 1) whichensures that the Q-values are bounded. The probability thata traffic light turns red is calculated as follows:

P(L′ | [node, dir, pos, des

],L,


])

= C([

node, dir, pos, des],L,


],L′)

C([



]) ,

(3)

where C([node, dir, pos, des], L, [node′, dir′, pos′, des])means the number of times a car in the state of [node, dir,pos, des] transiting to the state of [node′, dir′, pos′, des]and the transiting light is L, C([node, dir, pos, des], L,[node′, dir′, pos′, des],L′) is the number of times the lightturns L′ after such a transiting procedure.

4.2. Medium Traffic Condition. Under this medium trafficcondition, we focus on the overall waiting time of vehi-cles, which is the same as in Wiering’s model [13, 14].Q([node, dir, pos, des], action) is used to denote the totalwaiting time before all traffic lights for each car until itarrives at the destination given its current state and theaction of the light. V([node, dir, pos, des]) denotes the totalwaiting time (without knowing the traffic light decision)for a car at [node, dir, pos]until it reaches its destination.Q([node, dir, pos, des], action) and V([node, dir, pos, des])are iteratively updated as follows:

V([

node, dir, pos, des])

=∑L

P(L | [node, dir, pos, des

])Q([

node, dir, pos, des],L),

(4)

Q([


=∑

(node′, dir′, pos′)

P([



])

× (R([node, dir, pos, des],[node′, dir′, pos′, des

])

+γV([

node′, dir′, pos′, des′]))

,

(5)

where L is the traffic light state (red or green), P(L |[node, dir, pos, des]) is calculated in the same way as (3),R([node, dir, pos, des], [node′, dir′, pos′, des]) is defined asfollows: if a car stays at the same place, then R = 1, otherwise,R = 0 (the car can move forward).

4.3. Congested Traffic Condition. Under the congested trafficcondition, we must do our best to avoid the queue spillovers,which will seriously degrade the traffic control effect andprobably cause large-scale traffic congestion [15]. Therefore,the queue length is taken into consideration when we designthe Q learning procedure. Denote the maximum queuelength at the next traffic light tl′ as Ktl′ , shortly written asK . When the traffic light is red, no vehicle can pass throughto the next light. Thus, the equations at a red light do notchange, we focus on the function when light is green. Then(5) can be rewritten as follows:

Q([

node, dir, pos, des], Green

)=

∑(node′, dir′, pos′)

P([

node, dir, pos, des], Green,


])× (

R([

node, dir, pos, des],[node′, dir′, pos′, des

])+ αR′

([node, dir, pos, des


])+γV


])),

(6)

Q([

node, dir, pos, des], Red

)=

∑(node′, dir′, pos′)

P([

node, dir, pos, des], Red,


])× (

R([

node, dir, pos, des],[node′, dir′, pos′, des

])+γV


])),

(7)

where Q([node, dir, pos, des],L) and V([node, dir, pos, des])have the same meanings as under the medium trafficcondition. Compared (6) with (5), another reward functionR′([node, dir, pos, des], [node′, dir′, pos′, des]) is added toindicate the influence from traffic condition at the next light.R([node, dir, pos, des], [node′, dir′, pos′, des]) is the rewardof vehicles’ waiting time while R′([node, dir, pos, des],[node′, dir′, pos′, des]) indicates the reward from the queuelength increasing at the next traffic light. The parameter α isan adjusting factor.

R([node, dir, pos, des], [node′, dir′, pos′, des]) is definedas follows: if a car stays at the same place, then R = 1,otherwise, R = 0 (the car can move forward).

R′([node, dir, pos, des], [node′, dir′, pos′, des]) is definedas follows: if a car passes through the current intersection tothe next traffic light, which means that the queue length at


the next traffic light will increase by 1 in a short time, thenR = 1, otherwise, R = 0.

Given the capacity of the lane of next traffic light is L,then the adjusting factor α is determined by the queue lengthKtl′ as follows. Note when queue spillovers happen, Ktl′ willbe larger than L [15]

α =

⎧⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎩

0, if Ktl′ ≤ 0.8L,

10(Ktl′

L− 0.8

), if 0.8L < Ktl′ ≤ L,

2, if Ktl′ > L.

(8)

Through the definition we can find that α will increasesharply when the queue length approaches the capacity ofthe lane, which means that queue spillovers would liketo happen. Thus, under such a situation, Q([node, dir,pos, des], Green) will increase sharply and make the gainof this policy decrease. Therefore, the green phase lengthand the number of vehicles allowed to pass through will bedecreased until the queue at the next light has been dispersed.The largest value of α is set to 2 in this paper, but you canadjust its value according to the practical traffic condition.

4.4. Priority Control for Buses and Emergency Vehicles. Whenbuses or emergency vehicles (fire trucks or ambulances)enter the road network, they should have a priority to passthrough. It is necessary to realize the priority control of thesespecial vehicles with least disturbance to the regular trafficorder. Thus, we revise (5) as follows. A priority factor βis added to describe the emergency degree of these specialvehicles, which needs to be determined separately by thetraffic management department

Q([


=∑

(node′, dir′, pos′)

P([



])× (βR([node, dir, pos, des


])+γV


])).

(9)

5. Case Studies

We have done some case studies to prove the effectivenessof our model. Since it is very hard to apply a model tothe real traffic system management, traffic simulation ischosen to do the case studies. Paramics V6.3 was selectedas the simulation platform because it is a professional trafficsimulation tool which is recognized by traffic engineers allover the world. A practical road network within BeijingSecond Ring Road was modeled in Paramics as shownin Figure 3. This is a network with 7 intersections (N1–N7) and 8 OD zones (Zone1–Zone8). Intersections N1–N7correspond to the real intersections Xiaoweihutong, Dong-dansantiao, Jingyuhutong, Dengshidongkou, Dengshikou,Wangfujingbeikou, and Taiwanfandian.

N5

N4

N6

N3N2

N1 N7

Zone1

Zone2 Zone3

Zone4

Zone5

Zone6Zone7

Zone8

Figure 3: Sketch diagram of a practical road network in Beijing.

The simulation ran for 10000 time steps, the first 4000steps made up the learning process, and the latter 6000 stepswas used to collect the simulation results. Factor γ is set tobe 0.9 and β is set to be 3. The lanes in the network aredivided into cells with length of 7.5 m. The capacity of thelanes equals to the number of the cells.

We compared our method with the fixed control, theactuated control and also Wiering’s method. The setting offixed control is as follows, the cycle is 2 minutes and the greentime is equally assigned to all phases. In the actuated controlstrategy, the minimum green time is 10 s, the maximumgreen time is 50 s, and the extension of green time is set to 4 s.Parameters of Wiering’s method are the same as our modelunder the medium traffic condition.

We wanted to estimate the effectiveness of the mul-tiobjective scheme, thus, we estimated the control effectsof these four algorithms under different traffic conditions.We changed the traffic volume entering the network everyminute from 30 to 270 and estimated the average waitingtime, the number of stops, and maximum queue length ofthese four methods.

In our model, when the traffic volume entering thenetwork in a minute is less than 90, it is regarded as thefree traffic; when the volume is larger than 90 but less than180, it is regarded as the medium traffic; when the trafficvolume is larger than 180, it is regarded as the congestedtraffic condition.

5.1. Comparison of the Number of Stops. The comparison ofthe number of stops with respect to the increasing of trafficvolume is shown in Figure 4. Fixed means the fixed controlstrategy, actuated means the vehicle actuated method, RLmeans the algorithm proposed by Wiering [13, 14], andmulti-RL means the model proposed in this paper.

It is obvious that when the traffic volume is less than90, which means that the traffic state is free. The numberof stops under the multi-RL control is less than those underother control strategies. This is because the multi-RL is


0 50 100 150 200 250 3001

2

3

4

5

6

7

Traffic volume

Ave

rage

stop

s

FixedActuated

RLMulti-RL

Figure 4: Control effects comparison estimated by average stops.

the only one that aims to minimize the number of stops.However, with the increase of traffic volume, the multi-RLmethod changes its objective, and the actuated control getsthe minimum stops.

5.2. Comparison of the Average Waiting Time. The com-parison of the average waiting time with respect to theincreasing of traffic volume is shown in Figure 5. Sincethe multi-RL is the same as the RL method under themedium traffic condition, they have almost the same averagewaiting time in the middle. Under the free traffic state,the RL gets the minimum waiting time because this is itsoptimization objective. It should be noticed the multi-RLgets the minimum waiting time when the traffic is congested.This indicates that although the RL aims to minimize thewaiting time, the queue spillover which is not considered willdecrease the traffic efficiency and increase the waiting time.

5.3. Comparison of Maximum Queue Length. The compari-son of the average waiting time with respect to the increasingof traffic volume is shown in Figure 6. The maximum queuelength exceeds 40 under the fixed control, which indicatesthat there must be some queue spillovers. This is taken intoconsideration in the multi-RL, thus, we get a short queueunder the congested traffic condition.

6. Conclusion

In this paper, a multiobjective control algorithm based onreinforcement learning is proposed. The simulation resultsindicate that the multi-RL gets the minimum stops underthe free traffic, though not the minimum waiting time;the multi-RL has almost the same performance with the

0 50 100 150 200 250 300

Traffic volume

FixedActuated

RLMulti-RL

150

200

250

300

350

400

Ave

rage

wai

tin

gti

me

Figure 5: Control effects comparison estimated by average waitingtime.

0 50 100 150 200 250 300

Traffic volume

FixedActuated

RLMulti-RL

Max

imu

mqu

eue

len

gth

0

5

10

15

20

25

30

35

40

45

Figure 6: Control effects comparison estimated by maximumqueue length.

RL method under the medium traffic, which is better thanthe fixed control and the actuated control; under congestedcondition, the multi-RL can effectively prevent the queuespillovers to avoid large-scale traffic jams. It should be alsonoticed that multi-RL is a car-based algorithm. Therefore,it is less time consuming than the light-based reinforcementlearning algorithms [13].


However, there are still some system parameters thatshould be carefully determined by hand, for example, theadjusting factor α indicating the influence of the queue atnext traffic light to the waiting time of vehicles at currentlight under the congested traffic condition. This is a veryimportant parameter, which we should further research itsdetermining way based on the traffic flow theory. In addition,some phenomena in real traffic system such as the lanechanging and overtaking of cars will influence their traveltime. The assumption that all vehicles run at the samespeed is also not so reasonable. We would take these intoconsideration and build a model closer to the real trafficsystem in future work. Besides, the communications betweentraffic signal controllers will help to observe the network-wide traffic states and predict future traffic conditions, whichwill improve the traffic control effect and should be furtherresearched in the future.

Acknowledgments

This work is supported by the National High TechnologyResearch and Development Program (“863” Program) ofChina, Contract no.s 2006AA11Z229, 2007AA11Z215; by theKey Project of Chinese National Programs for FundamentalResearch and Development (973 program), Contract no.2006CB705506; by Chinese National Natural Science Foun-dation, Contract nos. 60834001, 60774034.

References

[1] C. P. Pappis and E. H. Mamdani, “Fuzzy logic controller fora traffic junction,” IEEE Transactions on Systems, Man andCybernetics, vol. 7, no. 10, pp. 707–717, 1977.

[2] M. B. Trabia, M. S. Kaseko, and M. Ande, “A two-stage fuzzylogic controller for traffic signals,” Transportation ResearchPart C, vol. 7, no. 6, pp. 353–367, 1999.

[3] J. C. Spall and D. C. Chin, “Traffic-responsive signal timingfor system-wide traffic control,” Transportation Research PartC, vol. 5, no. 3-4, pp. 153–163, 1997.

[4] Z. Liu, “Hierarchical fuzzy neural network control for largescale urban traffic systems,” Information and Control, vol. 26,no. 6, pp. 441–448, 1997.

[5] M. D. Foy et al., “Signal timing determination using geneticalgorithms,” Transportation Research Record 1365, NationalResearch Council, Washington, DC, USA, 1992.

[6] B. Park et al., “Enhanced genetic algorithm for signal timingoptimization of oversaturated intersections,” TransportationResearch Record 1727, National Research Council, Washing-ton, DC, USA, 2000.

[7] R. S. Sutton, “Learning to predict by the methods of temporaldifferences,” Machine Learning, vol. 3, no. 1, pp. 9–44, 1988.

[8] C. Watkins, Learning from delayed rewards, Ph.D. thesis, King’sCollege, Cambridge, UK, 1989.

[9] L. P. Kaelbling, M. L. Littman, and A. W. Moore, “Rein-forcement learning: a survey,” Journal of Artificial IntelligenceResearch, vol. 4, pp. 237–285, 1996.

[10] T. Thorpe, Vehicle traffic light control using SARSA, M.S. thesis,Colorado State University, 1997.

[11] B. Abdulhai, R. Pringle, and G. J. Karakoulas, “Reinforcementlearning for true adaptive traffic signal control,” Journal ofTransportation Engineering, vol. 129, no. 3, pp. 278–285, 2003.

[12] S. Mikami and Y. Kakazu, “Genetic reinforcement learning forcooperative traffic signal control,” in Proceedings of the 1st IEEEConference on Evolutionary Computation, vol. 1, pp. 223–228,Orlando, Fla, USA, June 1994.

[13] M. Wiering et al., “Intelligent traffic light control,” Tech. Rep.UU-CS-2004-029, University Utrecht, 2004.

[14] M. Wiering, “Multi-agent reinforcement learning for trafficlight control,” in Proceedings of the 17th International Confer-ence on Machine Learning (ICML’ 2000), pp. 1151–1158, 2000.

[15] C. F. Daganzo, “Queue spillovers in transportation networkswith a route choice,” Transportation Science, vol. 32, no. 1, pp.3–11, 1998.


Research Article

Design and Experimental Evaluation of a Vehicular NetworkBased on NEMO and MANET

Manabu Tsukada,1 Jose Santa,2 Olivier Mehani,1 Yacine Khaled,1 and Thierry Ernst1

1 INRIA Paris, Rocquencourt Domaine de Voluceau Rocquencourt, B.P. 105, 78153 Le Chesnay Cedex, France2 Department of Information and Communications Engineering, University of Murcia, Campus de Espinardo, 30100 Murcia, Spain

Correspondence should be addressed to Manabu Tsukada, [email protected]

Received 1 December 2009; Revised 19 July 2010; Accepted 5 September 2010


Copyright © 2010 Manabu Tsukada et al. This is an open access article distributed under the Creative Commons AttributionLicense, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properlycited.

Mobile Ad hoc Network (MANET) routing protocols and Network Mobility (NEMO) Basic Support are considered keytechnologies for vehicular networks. MANEMO, that is, the combination of MANET (for infrastructureless communications) andNEMO (for infrastructure-based communications) offers a number of benefits, such as route optimization or multihoming. Withthe aim of assessing the benefits of this synergy, this paper presents a policy-based solution to distribute traffic among multiplepaths to improve the overall performance of a vehicular network. An integral vehicular communication testbed has been developedto carry out field trials. First, the performance of the Optimized Link State Routing protocol (OLSR) is evaluated in a vehicularnetwork with up to four vehicles. To analyze the impact of the vehicles’ position and movement on network performances, anintegrated evaluation environment called AnaVANET has been developed. Performance results have been geolocated using GPSinformation. Second, by switching from NEMO to MANET, routes between vehicles are optimized, and the final performance isimproved in terms of latency and bandwidth. Our experimental results show that the network operation is further improved withsimultaneous usage of NEMO and MANET.

1. Introduction

Terrestrial transportation is one of the most importantservices that support humans’ daily life. Intelligent Trans-portation Systems (ITS) aim at enhancing road trafficsafety and efficiency as well as optimizing social costs andimproving drivers’ comfort by providing services such asfleet management, route guidance, billing, or infotainment.These days, communication technologies are more and moreconsidered as a key factor for ITS deployment however, newapproaches are needed to integrate mobile networks in thevehicle field.

IPv6 can be a good base technology to fulfill severalITS communication requirements, thanks to its extendedaddressing space, embedded security, enhanced mobilitysupport, and autoconfiguration advances. Moreover, futurevehicles will embed a number of sensors and other IPv6-enabled devices [1]. A number of ITS applications can be

conceived when sensors deployed in vehicles are connectedto the Internet and data collected from them is shared amongvehicles and infrastructure. Since the speed and position ofvehicles can be shared in real time, valuable informationabout traffic conditions can be inferred. For example, byreporting brake events, vehicles driving towards the affectedroad segment can be warned in advance and authorities canbe ready for possible fatalities.

In order to deal with communication requirements ofITS applications [2], on-the-move and uninterrupted Inter-net connectivity is necessary. Network Mobility (NEMO)Basic Support has been specified by the IETF (InternetEngineering Task Force) NEMO Working Group [3] to pro-vide on-the-move IP connectivity maintaining addressingconfiguration. NEMO is an essential part of the Commu-nication Architecture for Communications Access for LandMobiles (CALM)) (http://www.calm.hu/), currently beingstandardized at ISO [4]. The European ITS Communication


Architecture defined by COMeSafety [5] and ETSI [6] alsointegrates NEMO and IPv6 to provide and maintain Internetconnectivity to vehicles.

Additionally, Mobile Ad hoc Networks (MANETs) canbe used for vehicular communications without dependingon any third-party infrastructure. Several MANET protocolshave been specified by the IETF MANET Working Group.These routing protocols are classified as reactive or proac-tive [7], depending on whether communication routes arecreated when needed or they are continuously maintained.The Optimized Link State Routing (OLSR) protocol hasbeen specified at IETF as a proactive protocol [8]. Thisprotocol has been chosen in the present research to createa Vehicular Ad hoc Network (VANET), since it is a well-know implemented, tested, and standardized protocol in theMANET literature.

This paper describes the work done to combine NEMOand MANET/VANET in a design that distributes trafficamong multiple paths to improve the overall performanceof the vehicular network. A complete testbed has beendeveloped and used to experimentally evaluate the system.The rest of the paper is organized as follows. Networktechnologies related to vehicle communications are summa-rized in Section 2. Section 3 outlines scenarios and objec-tives of our network platform. Our integrated evaluationenvironment for vehicular networks and the Linux-basedimplementation are described in Section 4. Experimentalresults are covered in the following two parts: Section 5deals with the performance of the VANET subsystem, whileSection 6 evaluates the integrated MANEMO performance,both indoor and outdoor, considering field trials of theIPv6 mobility testbed of the Anemone project [9]. Finally,Section 7 concludes the paper summarizing main results andaddressing future works.

2. Network Technologies inVehicular Communications

This section presents a brief overview of relevant networkingtechnologies in vehicular communications. Several researchfields highly related to the work described in this paper,regarding NEMO and MANET, are also introduced, such asMultihoming, Route Optimization, and MANEMO.

2.1. VANET. Vehicular Ad hoc Networks (VANET) are aparticular case of MANET, but they are characterized bybattery constraints free, high speed, GPS-equipped nodes,and regular distribution and movement. First, vehicles havebatteries better than the ones integrated in mobile terminalsor sensor devices. Moreover, they are recharged while thevehicle’s engine is on. Hence, it is not necessary to takespecific measures to save energy resources (e.g., avoid signal-ing traffic). Second, mobility conditions of road vehicles aredifferent from the ones given in common portable terminals.The relative speed between two vehicles driving in oppositedirection can reach 300 km/h. Thus, in some scenarios, thelifetime of routing entries can be extremely short. Third, GPSinformation can be assumed to be available in many cases,

since an increasing number of vehicles are equipped withnavigation systems. Position and movement information canbe used to improve network performances. Additionally, themovement and density of vehicular nodes are not random,since vehicles drive along roads. This makes the position ofnodes somehow predictable.

Although there are many works related to VANETapplications, as well as basic research at the physical linkand network layers in vehicular communications, there isan important lack of real evaluation analysis. Many VANETsolutions and protocols could be considered as nonpracticaldesigns if they were tested in real scenarios, as it hasbeen proved for MANETs [10]. Performance of VANETprotocols based on a pure broadcast approach can be moreor less predictable in simple configurations, even if notexperimentally evaluated. However, the number of issuesconcerning real performances of multhop designs is muchlarger. There are works related to real evaluations of VANETdesigns [11, 12], and a limited literature for concrete casesof multi-hop transmissions [13], but there is an importantlack on works evaluating routing protocols on VANETs.This paper details the works carried out towards easingthe experimental evaluation of a multi-hop and IPv6-basedvehicular network. The design covers the integration ofvarious communication technologies to overcome commonproblems in VANETs, such as penetration rate or the need ofInternet connectivity.

OLSR is a well-known protocol in the MANET literature.Since the application of MANET concepts in the particularVANET case is a common procedure, the results given inthis paper assess how a common ad hoc proactive protocoloperates under vehicular conditions. Because vehicles arenot constrained by battery restrictions, one may think thata proactive protocol tuned for highly dynamic topologiescould be suitable in the vehicular domain. Evaluatingthis idea is an interesting point in the work. Moreover,the existence of stable implementations of OLSR and itspopularity among real ad hoc deployments have encouragedus to create a reference point in the VANET literature withreal multi-hop experiments based on this protocol. Thetestbed platform presented in next sections is prepared tochange the routing protocol, thus it will be extended withfuture implementations of pure-VANET protocols in theframe of our research on georouting [14].

2.2. NEMO. The NEMO Basic Support functionalitiesinvolve a router on the Internet to allow mobile computersto communicate with mobile or static remote nodes. Theapplication of NEMO in ITS is direct and it is done as follows.A Mobile Router (MR) located in the vehicle acts as a gatewayfor the in-vehicle Mobile Network and manages mobility onbehalf of its attached nodes (Mobile Network Nodes, or MNNsfor short). MR and a fixed router in the Internet, its HomeAgent (HA), establish a bidirectional tunnel to each otherwhich is used to transmit packets between the MNNs andtheir Correspondent Nodes (CN).

The possible configurations offered by NEMO have beenclassified in [15], according to three parameters: the number


Internet

Mobile network nodes

MR1

GPRS/UMTS WiMAX IEEE802.11x

HA1 HA2

MANET

NEMO

MR2

A B

1) Integrated evaluation environment

2) Simultaneous usage of NEMO and MANET

Figure 1: Generic intervehicle communication scenario. Network nodes inside vehicles communicate with their peers via the VANET orthrough the Internet using NEMO.

Figure 2: Prototype vehicles used in the field experiments.

x of MRs in the mobile network, the number y of HAsserving the mobile network and the number z of MNPs(Mobile Network Prefixes) advertised in the mobile network.In this paper, we focus on the “single MR, single HA andsingle MNP” configuration, commonly called (x, y, z) =(1, 1, 1).

2.3. Multihoming. Mobile Routers can be shipped withmultiple network interfaces such as Wi-Fi (IEEE 802.11 a/b/gand more recently 802.11 p), WiMAX (IEEE 802.16-2004/e-2005) or GPRS/UMTS. When an MR simultaneously main-tains several of these interfaces up and thus has multiplepaths to the Internet, it is said to be multihomed. In mobileenvironments, MRs often suffer from scarce bandwidth,frequent link failures and limited coverage. Multihomingbrings the benefits of alleviating these issues.

NEMO Basic Support establishes a tunnel between theHome Agent’s address and one Care-of Address (CoA) ofthe MR, even if the MR is equipped with several interfaces.In [16], it is proposed the Multiple Care-of AddressesRegistration (MCoA), an extension of both Mobile IPv6 andNEMO Basic Support, to establish multiple tunnels betweenMRs and HAs. Each tunnel is identified by its BindingIdentification Number (BID). In other words, MCoA dealswith simultaneous usage of multiple interfaces.

2.4. Route Optimization. Route Optimization allows to sortthe communication path between a mobile router (or a host)and a correspondent node that is not connected to the HomeAgent at a concrete moment. In NEMO, all the packets to andfrom MNNs must be encapsulated within the tunnel betweenMR and HA. Thus, all packets to and from CNs must gothrough HA. This causes various problems and performancedegradations. One could imagine the delay of using the HAtunnel when both nodes could (in the worst case) be in thesame transiting network. A standardized solution for RouteOptimization is still missing for NEMO Basic Support, whilethere exists one for Mobile IPv6 [17]. Main drawbacks ofsuch NEMO behavior can be classified as follows.

(1) Suboptimal routes are caused by packets being forcedto pass by HA. This leads to an increased delaywhich is undesirable for applications such as real-time multimedia streaming.

(2) Encapsulation with an additional 40-bytes headerincreases the size of packets and the risk of frag-mentation. This results in a longer processing timefor every packet being encapsulated and decapsulatedboth at MR and HA.

(3) Bottlenecks in HA is an important problem, since asignificant amount of traffic for MNNs is aggregatedat HA, particularly when it supports several MRsacting as gateways for several MNNs. This may causecongestion which would in turn lead to additionalpacket delays or even packet losses.

(4) Nested Mobility which occurs when a Mobile Routerget attached to other Mobile Routers. This couldarise, for example, when passengers carry a Per-sonal Area Network or in scenarios where the sameoutbound MR is used by several vehicles. NestedMobility further amplifies the aforementioned routesuboptimality.


Vehicle network

MNN2

MR1

3G

Ethernet

MNN1

NEMO1

IPv4 internetIPv6 over IPv4 tunnel

NEMO2

MR2MR4

Ethernet

HA1

HA2

IEEE 802.11bmanaged mode

Inria IPv6network (France)

Infrastructure network

MR3

SFR 3GIPv4 network

IPv6 over IPv4 tunnel

Irisa IPv6network (France)

IEEE 802.11b ad hoc mode

Figure 3: Topology of the vehicular network and Internet connectivity.

The previous route optimization issues of NEMO areidentified in [18] by the IETF whereas the solution space isanalyzed in [19]. Requirements for Route Optimization invarious scenarios have been described for vehicle networksin [20] and for aeronautic environments in [21].

2.5. MANEMO. Both MANET and NEMO are layer-threetechnologies. NEMO is designed to provide global con-nectivity, while MANET provides direct routes in wirelesslocal area networks. MANEMO combines both concepts toprovide several benefits related to route optimization.

Since direct routes are available in MANETs, they canprovide direct paths between vehicles. These paths can beoptimal and free from NEMO tunnel overhead [22, 23].Possible topology configurations with MANEMO have beendescribed in [24], while issues and requirements have beensummarized in [25]. In addition, MANEMO has alreadybeen suggested for vehicular communications. For example,VARON [26] focuses on NEMO route optimization usingMANET. It also provides the same level of security as thecurrent Internet, even if communications are done via theMANET route.

3. Scenario and Objectives

This paper focuses on the scenario of intervehicle communi-cation shown in Figure 1. Sensors installed in the vehicle areconnected to the Internet to share real-time information, andon-board computers or mobile terminals (i.e., MNNs) areconnected to the mobile network within the vehicle. Vehiclesare connected to the Internet everywhere and anytime withmultiple interfaces using NEMO. Each MR, acting as agateway for the mobile network, supports both NEMO andMANET connectivity.

In this paper, the focus is on investigating the operationand performance of the simultaneous usage of VANET andNEMO routes. An initial set up of a field testbed based

on four-wheeled electric vehicles was carried out, calledCyCabs [27], to identify issues and requirements of realenvironments. This testbed helped us to prepare a feasiblestudy considering issues such as wireless links features, con-nectivity changes or vehicles’ movement. The experimentspresented in the following sections were conducted usingup to four common commercial vehicles (Citroen C3s) asdepicted in Figure 2.

Among the different advantages of the developed testbed,three main capabilities can be remarked. First, apart fromstudying traffic flows sent through the fixed network, it ispossible to evaluate VANET performances in detail using anintegrated testing environment. Second, the testbed is opento develop and validate any ITS application. Third, a numberof different scenarios can be tested to analyze the operationof all network layers working together.

In order to measure the network performance of aVANET, various metrics should be considered. The band-width, round-trip time (RTT), jitter, packet delivery ratio(PDR), and number of hops are measured for various com-munication types (e.g., UDP, TCP, or ICMPv6). Geographicmetrics, such as speed, position and distance between carsare also collected and linked with the previous network mea-suraments. As far as authors know, there are no integratedtools that perform all this tasks at once.

Several issues arise when the previous performance mea-surements are collected and linked. These can be grouped inthe next three classes.

(1) Path awareness. This comprises the problem of deter-mining the route followed by packets from source todestination in a dynamic topology.

(2) Performance measurements hop-by-hop. Perfor-mance data is usually collected in an aggregate end-to-end manner by classical network analysis tools(e.g., ping6 or IPerf), but is not accessible on aper-hop basis. Hence, it is not easy to identify wherepackets are lost, for instance.


−6

−4

−2

0

2

4

6

Nu

mbe

rof

hop

s

90

80

70

60

50

40

30

20

10

00

Dis

tan

ce(m

)

50 100 150 200 250 300 350 400

Time (seconds)

Car 3

Sender(UDP, TCP, ICMPv6

traffic generation)

Ethernet(in-vehiclenetwork)

Ethernet(in-vehiclenetwork)

MR3

Senderscript

Iperf andping6 logs

MR script MR scriptMR script

Tcpdumplog

Tcpdumplog

Tcpdumplog

GPS log GPS logGPS log

Receiverscript

Iperf log

(Optional) (Optional)

Car 4/2/∗ Car 1

MR4/2/∗

Receiver

Wi-Fi(VANET)

Wi-Fi(VANET)

Processing

XMLstatistics

Packettraces

Graphicgenerator

Analysis

Web front-end (google maps)

AnaVANET

Experiments

Distance between MR3 and MR2

Distance between MR2 and MR1

Hops

Figure 4: Experimental setup and data processing units.

(3) Movement awareness. The route followed by vehiclesin the physical world is also an important issue tofurther identify the cause of network problems dueto real mobility conditions.

Moreover, in preceding works [28], switching from aNEMO to a MANET route gave benefits regarding routeoptimization in terms of bandwidth and delay. In this paper,we also propose to distribute traffic into multiple paths toimprove the global network performance. This simultaneoususage of NEMO and MANET has been experimentallyevaluated within our testbed.

4. Vehicular Network Design and TestbedArchitecture

Our network architecture setup is detailed in this section.First, the global architecture is introduced in Section 4.1.Sections 4.2 and 4.3 focus on describing the evaluationenvironment used to analyze the VANET performance andthe general MANEMO architecture, respectively.

4.1. Vehicular Network Architecture. The testbed comprises acombination of vehicle-to-vehicle and infrastructure-based


::0 (NEMO route)::/64 (MANET route)::/64 (other route)::128 (other route)

PacketIF

IF

Routing table

Figure 5: Classic routing. A single routing table is used, and packets are forwarded along the route with the longest matching prefix.

Packet IF

IF

Packet mark

Src addressDst address

Src portDst port

Flow type

Routing tables

NEMO route (BID1)

Routingpolicy

database

NEMO route (BID2)

MANET route

Other routes IF

Figure 6: Policy routing. Depending on several criteria, each packet is routed according to one of several routing tables.

networks, as Figure 3 depicts. Each vehicle is equipped with amobile router, with at least two interfaces: an Ethernet linkand an 802.11b adapter in ad hoc mode. MNNs connectto the in-vehicle network via its Ethernet interface (aninternal managed Wi-Fi network could also be used for thispurpose), while the ad hoc Wi-Fi interface is used for theinter-vehicle connections. In Figure 3, MR1 and MR2 arealso connected to an infrastructure network using another802.11 interface in managed mode. MR1 has an additional3G modem to establish a second link to the Internet (PPPlink provided by SFR (SFR is a french mobile telephonyoperator partially owned by Vodafone) ). MR1 is supportedby HA1 at INRIA Rocquencourt and MR2 is supported byHA2 inside Irisa’s network. Both networks are located inFrance and interconnected via Renater (French backbone foreducation and research) using a direct 6in4 tunnel to workaround some IPv6 firewalling problems (the testbed sites are12 IPv4 hops apart).

4.2. VANET Experimentation Subsystem. An experimenta-tion tool has been designed to overcome the issues relatedto VANET evaluation described in Section 3. This softwarecovers the VANET part of the testbed architecture (i.e.,bottom part of Figure 3).

4.2.1. Data Acquisition and Postprocessing Fusion with Ana-VANET. An overview of the experimental evaluation processis presented in Figure 4. The four vehicles previouslydescribed are considered here, although the system can sup-port any number of vehicles. A sender terminal (MNN), con-nected to one of the in-vehicle networks, is in charge of gen-erating data traffic towards a receiver terminal (MNN) insideanother vehicle. Both sender and receiver save a high levelperformance log according to the applications used to gener-ate network traffic. All MRs keep track of sent or forwardeddata packets using tcpdum (http://www.tcpdump.org/) andlog the vehicles’ position. All these data are then postpro-cessed by the AnaVANET software.

AnaVANET is a Java application which traces all datapackets transmitted or forwarded by mobile routers. It thus

detects packet losses and can generate both end-to-end andper-hop statistics, as well as join these measurements withtransport level statistics from the traffic generation tool.AnaVANET generates XML files with statistics at a onesecond granularity, and packet trace files listing the pathfollowed by each data packet.

The XML statistics file is uploaded to a Web server, whichuses the Google Maps API to graphically replay the tests andshow performance measurements in a friendly way, as canbe seen in Figure 4. A screenshot of this web applicationis available on Figure 10 in Section 5. All experimentswhich have been performed up to now can be replayed andmain performance metrics can be monitored at any time,by using the control buttons on the left side of the webpage. The replay speed can be adjusted and a step-by-stepmode has been implemented. On the map, the positionsand movements of the vehicles are depicted along withtheir speed and distance to the rest of cars. The amountof transferred data, throughput, packet loss rate, round-trip time, and jitter, both per-hop and end-to-end, are alsodisplayed. Main network performances can be graphicallychecked looking at the width and color of the link linesamong vehicles.

The Graphic Generator module gives another view of thenetwork performance. It processes both the XML statisticsand packet traces to generate several types of graphs usingGNUPlot (http://www.gnuplot.info/).

4.2.2. Traffic Analysis and Performance Metrics. Three dif-ferent types of traffic have been considered over the IPv6VANET in the tests.

UDP: A unidirectional transmission of UDP packets, fromthe sender to the receiver terminal has been generatedusing IPerf (http://iperf.sourceforge.net/). The packetlength is 1450 bytes to avoid IP fragmentation, andthey are sent at a rate of 1 Mbps.

TCP: A TCP connection is established between the senderand receiver terminals without any bandwidth limit.


IF

MR

MNN

Packetmark

NEMO routing table (BID2)

MAIN routing table

MANET routing table

NEMOD

OLSRD

OLSR node

OLSR node

IF

IF

IF

HA

BID1

BID2

Egress interface

Userpolicy

Ingress interface

Packets transmissionEntries addtion

IP tunnel

NEMO routing table (BID1)

Rule add/del

Route add/del

Route add/del

Routingpolicy

database

Rule add/del

Ad hoc

Figure 7: Internal route updating and selection mechanisms. NEMO and OLSR routes are stored in completely independent routing tables.

Web server

3G

MNN1

HTTP request(2 seconds)

IPerf serverWeb server

MR1IEEE 802.11b

infrastructure mode MNN2

IPerf client

MR2

HA2 IEEE 802.11binfrastructure mode

HA1

XML

IEEE 802.11b ad hoc mode

Figure 8: Network topology of the MANEMO demonstration system.

< markers >

<

markers >

<? xml version =′′ 1.0′′ encoding =′′ utf-8′′? >

marker interval =′′ 2.0′′ transfer =′′ 649′′ bandwidth =′′ 2660′′

lat =′′ 48.8375′′ lng=′′ 2.1010′′ offset lat=′′ 6.59′′

offset lng =′′ 7.21′′ distance =′′ 9.77′′ time =′′ 1195225195′′/ >

< /

Figure 9: XML data file generated based on IPerf measurements.

IPerf is again used as the traffic generator. Thesegment size is 1440 bytes.

ICMP: The Windows XP ping6 utility is used to generateIPv6 ICMP echo request packets at the sendernode and to receive echo reply packets from theremote note.

These three traffic types have been used to analyze hop-by-hop and end-to-end network performances, consideringthe most extended metrics for MANET evaluations [7].In the TCP case, statistics from IPerf with a 0.5 secondgranularity, such as the current throughput, are directly usedby AnaVANET for performance analysis. Each ICMPv6 andUDP packet is, however, traced across nodes. Since there is

Figure 10: AnaVANET replaying a VANET experiment. Buildingsavoid a direct line-of-sight communication, thus forcing the usageof multihop routes.

no fragmentation for UDP packets, a direct correspondenceexists between MAC and IP packets. At this level, thepacket delivery ratio (PDR), number of hops and jitter arecalculated. For ICMPv6 tests, the RTT is logged to analyzethe network delay.


20

60

100

2060

100

2060

100

2060

100

100 200 300 400 500 600

PDR from MR3 to MR1

0

20

40

60

80

100

PDR from MR3 to MR2

PDR from MR2 to MR1

0

100 200 300 400 500 6000

100 200 300 400 500 6000

100 200 300 400 500 6000

PDR

JitterHops

100 200 300 400 500 6000

Time (seconds)

Time (seconds)

Time (seconds)

Time (seconds)

Time (seconds)

Pack

etde

liver

yra

tio

(%)

Pack

etde

liver

yra

tio

(%)

Pack

etde

liver

yra

tio

(%)

Pack

etde

liver

yra

tio

(%)

Jitt

er(m

s)

Nu

mbe

rof

hop

s

543210−1−2−3−4−5

Figure 11: UDP performances over a multihop VANET of three cars moving in an urban environment.

4.3. MANEMO Implementation. A policy routing algorithmhas been developed and integrated in the architecture toallow simultaneous usage of NEMO and MANET. Thissubsystem allows vehicles to communicate with each otherover both the fixed and VANET networks at the same time,as was illustrated in Figure 3.

4.3.1. Policy Routing. The system has been implementedon GNU/Linux (kernel 2.6.21.3). To distribute packets tomultiple paths simultaneously from a MR, a policy routingscheme has been designed. Classic routing mechanisms arenot suitable because of the “longest match” principle. Asshown in Figure 5, packets arriving to the MR are forwardedto the routing table entry which has the longest prefix incommon with the destination address. In the MANEMOcase, MANET routes typically have longer prefix lengths than

NEMO ones. The formers are thus used in priority when theyare available in the routing table. NEMO routes then havethe least preference and are used as default routes. A singlerouting table can be used for switching between routes butnot for simultaneous usage of NEMO and MANET.

To solve the previous problem, we propose multiplerouting tables using a Route Policy Database (RPDB), asshown in Figure 6. To achieve this goal, the Netfilter(http://www.netfilter.org/) framework is used. The RPDBallows to maintain several independent routing tables inthe kernel. Each packet can then be routed according toany of these tables. The determination of which routingtable that should be used in a particular case is up tothe implementation. It is usual to route depending on thetype of flow that is being treated. This mechanism allowsdistributing packets to multiple concurrent routes at thesame time.


−10 −5 0 5 10 −10 −5 0 5 10 −10 −5 0 5 10

−10 −5 0 5 10 −10 −5 0 5 10

−5 0 5 10 −10 −5 0 5 10 −10 −5 0 5 10

Ban

dwid

th(K

bits

/sec

)B

andw

idth

(Kbi

ts/s

ec)

Ban

dwid

th(K

bits

/sec

)

Ban

dwid

th(K

bits

/sec

)B

andw

idth

(Kbi

ts/s

ec)

Ban

dwid

th(K

bits

/sec

)

Ban

dwid

th(K

bits

/sec

)

Ban

dwid

th(K

bits

/sec

)

Average1st2nd

Average1st

2nd

Average1st

2nd Average1st

2nd

Average1st2nd

Average1st2nd

Average1st2nd

Average1st2nd

NW176, 344,477, 594

W192, 361,

489,

E111, 272,432, 551

SW216, 388,

504

S231, 403,

519

SE97, 257,419, 537

Time

N146, 322,459, 580

NE127, 304,447, 566

4th3rd

4th3rd

4th3rd

4th3rd

4th3rd

3rd3rd

3rd

Time (seconds)Time (seconds) Time (seconds)

Time (seconds)

Time (seconds) Time (seconds) Time (seconds)

Time (seconds)

0

200

400

600

800

1000

1200

0

200

400

600

800

1000

1200

0

200

400

600

800

1000

1200

0

200

400

600

800

1000

1200

0

200

400

600

800

1000

1200

0

200

400

600

800

1000

1200

0

200

400

600

800

1000

1200

0

200

400

600

800

1000

1200

Figure 12: Network throughput at corners and straight roads for the UDP multihop test.


0

100

200

300

400

500

600

0 50 100 150 200 250 300

Always 3G3G or managedAny interface

11b managed is available11b ad-hocis available

Loss

Time (seconds)

Rou

nd

trip

tim

e(m

s)

Figure 13: Impact of route changes on the RTT, measured using ICMPv6 packets in the absence of background traffic.

4.3.2. Implementation Details. NEPL (NEMO Platformfor Linux) (http://software.nautilus6.org/NEPL-UMIP/) ver-sion 20070716 has been installed on MRs along witholsrd (OLSR Daemon) (http://www.olsr.org/) version0.5.3. NEPL is developed and distributed freely by Nau-tilus6 (http://www.nautilus6.org/) within the WIDE project(http://www.wide.ad.jp/). NEPL is based on MIPL (MobileIPv6 for Linux) (http://www.mobile-ipv6.org), developedby the Go-Core (Helsinki University of Technology) andNautilus6 projects.

The OLSR daemon has been adapted to the routingscheme proposed in Section 4.3.1. In this way, OLSR routingentries are maintained in one of the multiple routing tablesof the kernel. The NEMO daemon already handles policyrouting when patched for MCoA support (http://software.nautilus6.org/MCoA/).

NEMO and OLSR daemons operate independently inMRs. The NEMO one maintains its binding update listand NEMO routes, while the OLSR daemon takes care ofMANET routes. As shown in Figure 7, both NEMO andMANET routing entries are kept up-to-date in separatetables.

When started, both daemons add rule entries that specifywhich packets should be routed according to which routingtable (these are removed at the execution end). MRs havemultiple routing tables, which save NEMO and MANETroutes, and the default one (depicted as MAIN in Figure 7),which saves the rest of routes. There is the same numberof NEMO routing tables than egress interfaces on the MR.Each of these routing tables has a specific BID. The MANETrouting table is used for traffic that should be routed directlyto neighboring vehicles, and the MAIN table is mostly usedto route OLSR signaling.

Packets from MNNs arrive at the MR containing thesource and destination addresses and ports, as well as theflow type information. Packets are distributed according to

the latter mark to either the NEMO or MANET routingtables. Packets matched with a NEMO routing table aretransmitted to the tunnel bound to the HA, while packetsmatched with the MANET table are transferred to otherOLSR nodes directly.

4.3.3. Demonstration Platform. As a demonstration of thepolicy-based MANEMO system, the performance measuredin a communication between two vehicles is shown ona website (http://fylvestre.inria.fr/∼tsukada/experiments/),mapped to their geographical positions. The data havebeen collected during field trials on the Promotion Daysof the Anemone Project (12–14th December 2007). Thisproject aims at developing a large-scale testbed for mobilecommunication technologies. Our demonstration was anexample of a third party system using the mobility testbed.

Measurements were made with a GPS-enabled IPerf(http://gforge.inria.fr/frs/?group id=620&release id=915) ina topology as shown in Figure 8. This diagram illustrates indetail the MANEMO part of the general vehicular networkdescribed at the beginning of this section in Figure 3. MNN1works as an IPerf server and MNN2 is the client. IPerfreports the amount of transferred data and used bandwidth.Additionally, the GPS patch appends location information(latitude and longitude) as well as the offset and distancefrom the starting point. Only a regular GPS receiver isneeded.

The demonstration can be performed either in real-time or log mode. The former shows network performancesmapped with position in real time on the website, while thelatter saves them on the MNN’s local disk to be displayed later(see Figure 18).

In real-time mode, XML files are generated from mea-sured metrics and positions every two seconds by MNN1.An example XML output is shown in Figure 9. The remoteweb server periodically gets the data file from MNN1 using


0

10

20

30

40

50

60

110 115 120 125 130Loss

Up

(122.5, 21.27)

170 175 180 185 190

3G or managedAny interface

Down

Rou

nd

trip

tim

e(m

s)

Time (seconds)Time (seconds)

Figure 14: Closer look at the RTT values collected when the adhoc interface is turned on and off.

0

1000

2000

3000

4000

5000

0 50 100 150 200 250 300

Any interface3G or managedAlways 3G


Time (seconds)

Th

rou

ghpu

t(k

bps)

Figure 15: Evolution of the throughput of three TCP flows betweenMNNs using routing policies.

wget. The real-time mode has the advantage of everyone canimmediately check the network operation. Measurementscan, however, be slightly affected by the XML file transfers, asthey are carried over the NEMO route. By contrast, this effectis not present in log mode. Main results of these experimentsare later analyzed in the paper using the log mode.

5. VANET-Only Performance Evaluation

This section presents an experimental evaluation of ourVANET testbed. Here, we only consider the lower part ofthe architecture shown in Figure 3. Then, field trials havebeen performed using the integrated evaluation environmentdescribed in Section 4.1. Seven different scenarios with

0

200

400

600

800

1000

0 50 100 150 200 250 300

Always 3G3G or managedAny interface


Loss

Rou

nd

trip

tim

e(m

s)

Time (seconds)

Figure 16: RTT between MNNs with three background TCP flows.

various mobility patterns have been evaluated, using thethree types of traffic previously described (UDP, TCP, andICMPv6). This section analyzes one of our urban scenariosas reference. See [29] for a complete description of all theexperiments.

5.1. Experimental Setup Details. In the VANET evaluationshown in Figure 4, mobile routers use only the routingtable given by OLSR to forward data packets. MNN1 isa Mac OS X 10.4 laptop and MNN2 is a Windows XPProfessional PC. An embedded computer is used as MRin each car. It consists of a Soekris net4521 board witha Texas Instruments ACX111 Mini-PCI 802.11 b/g wirelesstransceiver and a compact flash memory card. The wirelessinterface has been setup for 11 Mbps operation, emulating an


0

500

1000

1500

2000

2500

3000

3500

0 20 40 60 80 100 120 140

Any interface3G or managedAlways 3G

11b managedis available

11b managedis available

Time (seconds)

Th

rou

ghpu

t(k

bps)

11b ad hoc is available

Figure 17: Throughputs of three TCP streams in a field experiment.Performances are less stable than in the indoor case.

Table 1: OLSR Parameters for the VANET-only experiments.

Parameter Value (default)

HELLO interval 0.5 sec (2.0 sec)

HELLO validity 6.0 sec (6.0 sec)

HNA interval 3.0 sec (5.0 sec)

HNA validity 9.0 sec (15.0 sec)

802.11 b device. This computer is also connected via a serialport to a Trimble AgGPS 323 GPS receiver. Both the Wi-Fiand the GPS antennas are affixed on the roof of the car.

The timing parameters of the OLSR daemon installed inMRs have been modified as showed in Table 1, to accom-modate mobility conditions of a vehicular network. Thesemodifications enable MRs to discover topology changesmore quickly.

5.2. Non-Line-of-Sight Multihop Communication. This sce-nario considers a typical urban environment where buildingsprevent a direct line of sight between the source anddestination cars. A multi-hop network is better suited toprovide a robust connectivity under these conditions.

During 600 seconds of test, a unidirectional transmissionof UDP packets is generated from MNN1, in vehicle 3, toMNN2, in vehicle 1. As was explained in Section 4.2, thepacket size is 1450 Bytes to avoid IP fragmentation and theyare sent at a rate of 1 Mbps.

The results of this experiment (along with the restof performed trials) is available on a public website(http://fylvestre.inria.fr/∼tsukada/experiments/vanet-jose/),and can be replayed to graphically show the performance ofthe network during the tests (see Figure 10).

The experiment was performed in the Rocquencourtcampus of INRIA. This area contains a set of small buildingssurrounded by streets, as can be seen in Figure 10. The

four streets showed in the image, which round three of thebuildings, have been chosen for this scenario. They standin a 100 × 100 m square area. Three vehicles have beendriven around the buildings, trying to block the direct linkbetween cars one and three. The speed of the vehicles waskept between 15 and 30 km/h. The right and left roads visiblein Figure 10 are very narrow and some communicationproblems were experienced when approaching the corners.

The results collected in the UDP tests are plotted inFigure 11. The several graphs show the results collectedduring four tests around the buildings. The upper plot showsthe number of hops used in the paths followed by UDPpackets whereas the lower graphs show the end-to-end andper-link PDR. PDR is computed every second, while thenumber of hops is plotted for each packet transmitted fromthe sender node. When no hops are drawn, the route to thedestination vehicle is not available. Zero hops means that thepacket was sent by the first MR, but was not received by anyother. Negative values represent those packets that did notarrive to their destination, but reached some intermediatehops.

As can be seen, a direct relation exists between PDR andnumber of hops. When the number of hops is equal to orlower than zero, PDR decreases. When the vehicles drivealong the same street, some direct paths (one hop) appear.On the contrary, when the distance between the sender andthe receiver cars is large enough, the two-hop routes areused. These different types of paths can also be observedwith the per-link PDR. Whereas the direct link (MR3-MR1)gives intermediate PDR values, the PDR between consecutivevehicles is almost identical and close to 100% when the two-hop link is used, due to the lower distances between nodes.

The performance obtained in the scenario has beenanalyzed according to the location of vehicles: corner andstraight road. As can be seen in Figure 12, each corner iscalled SE, NE, NW, SW, and according to its position (i.e.,South-East for SE). In the same way, straight roads havebeen assigned the names E, N, W, and S. Numbers belowcorner and road names indicate the time in which car 2(vehicle in the middle) passes these points. For example, forSE, numbers 97, 257, 419, and 537 mean that car 2 passesthe SE corner at times 97s, 257s, 419s, and 537s. At thesetimes, the sender vehicle is at the next road (E in this case)and the receiver vehicle is at the previous one (S). On theother hand, when the middle vehicle is at a straight road, thesender vehicle reaches the next corner and the receiver vehicleis at the previous corner. The driving order is SE - E - NE -N - NW - W - SW - S, and three and a half complete roundshave been considered. The analysis starts at time 97s at SEcorner, and it ends at 594 s at NW.

As can be seen in Figure 12, the throughput obtained hasbeen mapped with corner and straight road segments, and ithas been analyzed for each round at periods of ±10 seconds.A dotted line shows the result of each trial considered in thesegment, and a bold line shows the average bandwidth. Onecan notice the two different bandwidth patterns obtainedat corners and straight roads. Communication performanceincreases in corner scenarios, while it decreases at straightroads. When the intermediate vehicle reaches a corner, the


direct path between the source and the destination vehicle isblocked, thus a multi-hop route is established in the network.At this moment, a good bandwidth is noticeable. When theheading vehicle turns again in the next corner, a multi-hoproute still maintains connectivity but, as soon as car 3 and2 cannot maintain the communication link, the networkperformance falls. This can be seen in the last ten secondsof straight road graphs in Figure 12. The effect of loosingthe link between vehicles 2 and 1 is also present when theintermediate vehicle left the corner (last seconds of cornergraphs and first seconds of straight road graphs), but it is lessnoticeable, since these two vehicles were arbitrarily drivenmore closely during tests.

Results obtained for segment W shows a different behav-ior than for the rest of straight roads. This is explained bythe special physical conditions of the environment. First, thisstretch comprises a narrow street surrounded by buildingson both sides. As can be seen in Figure 10, these conditionsare only present in this segment, since the rest of roads havea clear space on one side. This fact enables the reflection ofsignals on the various walls. Moreover, the second interestingcondition identified in road W is the greater altitude of thesender car with regard to the receiver car, when these arelocated near corners NW and SW, respectively. About fivemeters of altitude difference increases the packet receptionprobability, and a direct path between cars 1 and 3 is evennoticeable at this segment. This can be checked at time367s in Figure 11. It is interesting to note that buildingson this INRIA area are quite low, about 3.5 meters, whatcomplements the altitude effect. The rest of direct pathscollected in the trials belong to segments S and E, which dohave open areas on one of the sides.

6. MANEMO Performance Evaluation

For the case of the MANEMO subsystem described inSection 4.2, measurements of latency and throughput havebeen collected using both the VANET and the infrastructuresegments of the testbed (Figure 3). A set of indoor andoudoor experiments have been conducted also at the Roc-quencourt campus of INRIA and this section presents andanalyzes most interesting results.

6.1. Experimental Setup Details and Initial Tests. Attendingto the global testbed setup in Figure 3, tests have beencarried out generating traffic from MNN1 towards MNN2using the best available communication route (i.e., NEMOor MANET). MNN1 is a Mac OSX 10.4 laptop and MNN2is a Windows XP tablet PC. As was done for VANET-only experiments, OLSR settings have been adjusted withthe values shown in Table 2. These have been chosento maintain a tradeoff between the delay experimentedwhen a topology change occurs and the network overloadthat implies control messages. Signaling traffic, apart fromreducing the effective bandwidth of the network, it consumescomputation resources on the nodes. For these experiments,OLSR settings have been adjusted to be aware of topologychanges faster than in VANET-only tests in Section 5, with

Figure 18: Website screen shot of the MANEMO experiment.

0

500

1000

1500

2000

2500

3000

Throughput

Th

rou

ghpu

t(k

bps)

Distance (meters)Average throughput

∼5 ∼10

∼15

∼20

∼25

∼30

∼35

∼40

∼45

∼50

∼55

∼60

60∼

Figure 19: Relation between distance and bandwidth using theMANEMO system.

the aim of performing fast route changes between NEMOand MANET.

Some initial tests were performed to check the oper-ation of the network. One issue that had to be solvedwas radio interferences between 802.11b managed and adhoc networks. Even when channels were chosen with agood distribution the problem persisted. To overcome thisdrawback, the bandwidth of MR1 interfaces were limitedto 2 Mbps using a Linux’ QoS system based on tc (TrafficControl) (http://www.linux-foundation.org/en/Net:Iputils).Network performance measurements under static condi-tions, and without any policy, between MNN1 and MNN2are summarized in Table 3, including three different routes.RTT results is the average of 100 packets of ICMPv6 betweenMNNs and throughput results have been obtained averagingresults obtained during a total of ten minutes of TCP tests.


Table 2: OLSR parameters for the MANEMO experiments.

Parameter Value

HELLO interval 0.5 sec

HELLO validity 1.5 sec

HNA interval 1.0 sec

HNA validity 3.0 sec

Table 3: Performance of the MANEMO system under staticconditions and depending on the route type.

Route & interface RTT Throughput

NEMO over 3G 279.43 ms 416 kbps

NEMO over 802.11b managed 32.74 ms 1977 kbps

OLSR over 802.11b ad-hoc 8.58 ms 1987 kbps

As can be seen, the results of the UMTS link offer the worseresults, due to the delay of the operator’s network and thebandwidth limited by the radio coverage and the availableresources in the cell. The 802.11b link improves these results,offering bandwidth capabilities equivalent to the ad hoc case.However, the delay induced by the managed mode and therelay access point impact on the RTT results.

6.2. Indoor Test Scenario. The policy-based MANEMO sys-tem has been firstly evaluated in an indoor testbed, to avoidinterferences of other equipments and difficulties to trace themovement of MRs. The following experiments have beenperformed without any vehicle. Neither MRs nor MNNshave moved during a reference experiment of 300 seconds.It clearly demonstrates the performance expected for longertimes or subsequent trials.

MNN1 has three addresses (A, B, and C) in the MNP,and MR1 distributes traffic from the mobile network viamultiple paths depending on the source address. Packetsfrom source address A or to port number 5102 are alwaysforwarded via the 3G interface. Those from source addressB or to destination port 5101 are routed via the Wi-Fimanaged interface when it is available. Otherwise, they areforwarded over the 3G interface. Traffic from source addressC or to destination port number 5009 is transmitted viawhatever available interface, prioritizing the most efficient(i.e., prefer ad hoc to managed Wi-Fi and only use 3 G ifno other link is available). Table 4 summarizes these policiesand indicates priorities in case several routes can be chosen.MR2 distributes returning flows into its managed and ad hocinterfaces according to the third policy, since it does not haveany 3 G interface. The Home Agents distribute flows to matchthese policies and avoid asymmetric routes.

For this indoor experiment, connection and discon-nection events have been created using a shell script andcommon system tools. From t = 0 to t = 60, both Wi-Fimanaged and ad hoc interfaces of MR1 are down. At t = 60,the managed interface comes up. At t = 120, the ad-hoc oneis made available. From t = 120 to t = 180, all the interfacesare up and running. At t = 180, the ad-hoc link is turned off.

At t = 240, the managed one is also switched off. The 3 Ginterface is always available throughout the test.

6.2.1. Latency Measurements. To measure the RTT betweenMNNs, MNN1 sends 56 Bytes ICMPv6 echo requestpackets from all addresses (A, B, and C) to MNN2 once every0.5 sec. There is no other traffic. These packets are distributedaccording to the policies described above. Results are showedin Figure 13. The average RTT on the NEMO route over3 G has been 261.9 ms. Changing paths to the NEMO routeover the managed Wi-Fi interface, has reduced the RTT to anaverage of 34.72 ms, which represents an 87% improvement.During the time the ad hoc mode has been available, theaverage RTT collected on the OLSR route (ad hoc link)has been 7.93 ms. In this way, route optimization usingMANEMO has further reduced the latency by 26.79 ms, whatrepresents an extra improvement of 77%.

For the two periods where the three ICMPv6 flows arecarried over the 3G network (from t = 0 to t = 60 and fromt = 240 to t = 300) an offset of 20 ms of delay between themis noticeable. It has been checked that the transmission of thethree echo request packets in a consecutive way results ina first-in first-out problem due to transmission and receptiontimes needed by the 3G driver. The extra overload incurredby the NEMO and tunnel and the L2TP (Layer-2 TunnelingProtocol) tunnel, necessary to support IPv6 traffic in the 3Gnetwork, increases the impact of this effect.

Figure 14 gives a closer look at the RTT results when thead hoc interface goes up/down and routes thus change. Att = 120, the ad hoc interface comes up, and then direct routeinformation of both MNPs are exchanged. At t = 122.5,the RTT obtained for the marked packet is 21.27 ms, whichcomprises an intermediate value between NEMO and OLSRmodes. This is because the ICMPv6 echo request hasused the NEMO route, while the echo reply has returnedthrough the ad hoc one. It takes 2.5 seconds for OLSRrouting entries to be added to MR1’s table after the ad hoclink has been connected. By contrast, the route is changedback from OLSR to NEMO 1.5 sec after the ad hoc linkis disconnected. During this switching phase, three packetshave been lost (From t = 180 to t = 181.5), due to the suddendisconnection of the ad hoc interface.

6.2.2. Throughput Measurements. To measure the through-put between MNNs, MNN1 sends three TCP streams toMNN2 by means of IPerf, with destination port numbers5102, 5101 and 5009, in the same routing scenario usedabove. At the same time, MNN1 also sends 56 Bytes ICMPv6echo request packets as in the previous section. IPerf givesa report once every two seconds and ping6 gives it every 0.5seconds. A reference test has been chosen among the set ofperformed tests. Figure 15 shows the achieved throughputwith stacked area graph and Figure 16 shows the observedRTT when the TCP flows are active.

A summary of throughput results is given in Table 5.The average total throughput on the NEMO route over 3G is455 kbps from t = 0 to t = 60. Since an 802.11b managednetwork is available from t = 60 to t = 120, the flows


Table 4: Flow distribution policies for MR1. Smaller numbers reflect higher priorities.

Policy Targets 3G Managed Wi-Fi Adhoc Wi-Fi

Always 3GSource address A or

1 × ×destination port 5102

3G or managedSource address B or

2 1 ×destination port 5101

Any interfaceSource address C or

3 2 1destination port 5009

are distributed in two paths and the average throughputincreases up to 1913 kbps, which represents an improvementof 76% (1458 kbps). From t = 120 to t = 180, ad hocconnectivity is also available. The average total throughputincreases again up to 3752 kbps when the three TCP flowsare distributed through the three paths, which represents anew improvement of 49% (1837 kbps).

The average RTT between MNNs is also listed in Table 5.The RTT on the NEMO route is about 400 ms when the threeTCP streams are transmitted using the 3G link. When twoTCP streams are diverted to the 802.11b managed interface,from t = 60 to t = 120, the RTT over the 3G link decreasesby about 280 ms, which represents an improvement of 30%.The RTT also decreases from 400 ms to about 130 ms forpolicies “3G or managed” and “Any interface” when Wi-Fimanaged is available, which comprises an improvement of68%. In addition, a further 50% (approx.) of improvement isobserved for policies “3G or managed” and “Any interface”when all the interfaces on MR1 are available, since eachcommunication technology is used by only one flow.

6.3. Field Experiment. The system has been evaluated with aset of field trials performed on the Telecom Bretagne/INRIARennes campus. 40 access points have been installed inthis area. The test has been performed in a straight roadsurrounded by buildings, where two access points have beeninstalled at two far away locations. The source vehicle startsmoving at a speed of 10 km/h from a position before thefirst access point, while the destination vehicle with MR2has been parked next the two access points. Both MRswere mounted inside the vehicles. Three TCP flows weretransmitted from MNN1 to MNN2 as in the previous tests.The flow distribution policies of MR1, MR2, HA1 andHA2 are also identical to those of the indoor testbed but,obviously, periods of 802.11 connectivity are not simulatednow.

The switch between access media and/or networks hasa clear impact on the available bandwidth, as can be seenin Figure 17. From t = 0 to t = 60, the path betweenMNNs is only via the NEMO route over 3G. The averagetotal throughput of the TCP flows is 344 kbps during thisperiod. The throughput in this field experiment is 111 kbpsless than in the indoor experiment. This is mostly due toobstacles and movements of the vehicle equipped with MR1.From t = 62 to t = 86 and from t = 106 to t = 116,the NEMO route through managed Wi-Fi is available, sincethe moving vehicle is near one access point each time. The

average total throughput of the TCP stream at these twoperiods is 1430.83 kbps and 957.34 kbps, respectively. Fromt = 124 to t = 130, the OLSR route over the VANET isavailable. The average throughput increases until 2408.4 kbpsduring this period.

In the evaluation, the NEMO route on the 802.11bmanaged interface has been used for 24 seconds for thefirst access point, and then an additional 10 seconds for thesecond one. As the speed of the vehicle was 10 km/h, thecoverage of the access points can be estimated to be between30 and 65 meters. The ad hoc interface has been availableduring six seconds, thus the VANET range can be estimatedto be 17 meters. In this case, the antennas of both MRswere located inside the vehicles. 802.11 performance couldtherefore be improved by mounting external and/or morepowerful antennas.

6.4. Impact of Geographical Location on Network Performance.In the previous section, the range of the available accesspoints and VANET links are estimated considering a sim-plification of the driving speed and the time of connection.This section presents more thorough range measurements.These have been collected by maintaining MR1 (and thusMNN1) moving in a 65 meters radius around the positionof MR2, and reporting the achievable throughputs. None ofthe MRs have gone out of the access points’ coverage and abuilding sometimes block the VANET route. As the wirelessaccess points are quite close to the test site, the managedinterface has been forcibly limited to 1 Mbps to account formore distant APs and highlight which network path MR1uses. By contrast, the ad hoc interface was not limited andthe average throughput between MRs using this interface was2685 kbps.

The position-mapped throughputs were measured atINRIA Rocquencourt in France, using the GPS-patchedversion of IPerf. To obtain a high density of throughput dataaround MR2, the evaluation was actually performed withoutvehicles. MR1 has been carried by a human at an averagespeed of 4 km/h. It starts moving from the position of MR2and comes back to the same position within 250 seconds. Theexperiments were run eight times. All the results are publiclyavailable (http://fylvestre.inria.fr/∼tsukada/experiments/).

A screenshot of the Web application can be seen inFigure 18. The website displays the throughput betweenMNNs by varying the size of the blue circles at each measurepoint. All tests can be displayed by selecting the Log option.Clicking on one of the circles reveals additional information,


Table 5: Total throughput of three TCP flows and RTT between MNNs.

PolicyAvailable interfaces

3G only 3G and managed All interfaces

Throughput

Always 3G 156 kbps 262 kbps 276 kbps

3G or managed 184 kbps 733 kbps 1612 kbps

Any interface 114 kbps 918 kbps 1863 kbps

Total 455 kbps 1913 kbps 3752 kbps

Round-Trip Time

Always 3G 389 ms 277 ms 275 ms

3G or managed 411 ms 127 ms 64 ms

Any interface 432 ms 130 ms 64 ms

including the test time, location and offset from the startingpoint, transferred data size and bandwidth in the last twoseconds. All the data can be shown at once with the Show Logbutton. Users can also analyze the results by changing datadensity and see the trajectory of MR1 (and thus MNN1).

Throughput results depending on the distance betweenMNNs can be seen in Figure 19. Data points show thethroughput obtained at the current distance, while the bargraph represents an average for five meters. Values over1 Mbps (the arbitrary limitation on the managed Wi-Fi link)are those recorded when the VANET route was available. Onecan see that it is available up to 40 meters. Between (approx.)20 and 40 meters, throughput measurements spread overa wide range from 100 kbps to 2700 kbps, because mediahandovers between the managed and ad hoc interfaces areperformed in these zones.

An asymmetrical tendency of the ad hoc link ranges hasbeen observed. From the collected results, it turns out thatthe OLSR route is usable over a longer distance when twovehicles are getting further from one another than when theyare getting closer. This hysteresis behavior is due to OLSR’sinitial delay caused by the period of sharing HELLO packets.This fact is further analyzed in [29].

7. Conclusions and Future Works

A proposal to distribute data traffic in vehicular communi-cations combining NEMO and VANET has been presented.This comprises an integral communication platform forthe ITS frame, which has been experimentally evaluated bymeans of a complete and open testbed. In a first stage, anintegrated evaluation environment for VANET enabled us toanalyze the network performance in detail in both per-hopand end-to-end manners, also considering the movementof vehicles. The evaluation environment provides novelperformance metrics for VANET, according to the currentliterature, such as the number of hops used to deliver a packetor the per-link PDR, in addition to typical statistics, such asend-to-end packet delivery ratio, round-trip delay time, jitteror bandwidth. Although it has been tuned to dynamic con-ditions, the OLSR protocol shows limitations to efficientlyupdate routing tables under stressful conditions, as it hasbeen seen above all in the MANEMO evaluation. A more

VANET-oriented protocol developed at INRIA, in frames ofthe GeoNet Project (http://www.geonet-project.eu/) will beevaluated through new field trials, using the presented test-bed. This is located inside the geographic-based routing pro-posals, which are demonstrating to be the correct directionin vehicular network research.

The MANEMO proposal has been evaluated usingcommon vehicles in real environments. Up to four vehicleshave been setup to carry out a number of experiments. Inour system, mobile routers use multiple egress interfacessimultaneously with NEMO and OLSR. The latter couldthus mitigate the sub-optimality caused by NEMO routes.Previous experiments results showed that MANEMO withroute switching from NEMO to MANET improved networkperformance in terms of latency and bandwidth. It cannow be stated that MANEMO with simultaneous usage ofNEMO and MANET can achieve further improvements on aintegrated vehicular network. Experimental results show thatthe achievable throughput and delay are improved when a setof interfaces (3G, 802.11 b managed and 802.11 b ad hoc) areavailable.

Among the different research lines that are now activeregarding this work, we plan to extend the MANEMO systemas follows. First, evaluation results show that network perfor-mances such as latency and bandwidth dynamically changeaccording to available interfaces, mobility or obstacles. Adap-tive applications are thus desirable in these environments.Second, traffic flows have to be allocated to appropriatepaths depending on the application demands and networkperformances. Since real-time applications are sensitive tohandovers, an intelligent path allocation is required. Third,traffic between MNNs has been distributed according topolicies manually specified by the administrator. As an MRcan only control its outbound traffic, policy changes on anMR may create asymmetric routes. By introducing commonfilter rules and a exchanging procedure among MRs and HAs[30, 31], policies on each entity can be synchronized. Fourth,currently, the position-mapped reports for the MANEMOcase are focused on bandwidth statistics. Network metricssuch as latency, packet loss rate or layer-two information willbe considered in further analysis for the whole MANEMOsystem, in the line of the work carried out for the VANETsegment.


Acknowledgments

The authors would like to thank the partners of Europeanproject Anemone, and more specifically Telecom Bretagneand INRIA Rennes who built the IPv6 mobility testbed fromwhich the MANEMO experiments exposed in this paper havelargely benefited.

References

[1] T. Ernst, “The information technology era of the vehic-ular industry,” ACM SIGCOMM Computer CommunicationReview, vol. 36, no. 2, pp. 49–52, 2006.

[2] Y. Khaled, M. Tsukada, J. Santa, J. Choi, and T. Ernst, “A usageoriented analysis of vehicular networks: from technologies toapplications,” Journal of Communications, vol. 4, no. 5, pp.357–368, 2009.

[3] V. Devarapalli, R. Wakikawa, A. Petrescu, and P. Thubert,“Network mobility (NEMO) basic support protocol,” RFC3963 (Proposed Standard), January 2005.

[4] ISO/TC204 WG16. ISO/WD 21210, “CALM—medium andLong Range, High Speed, Air Interfaces parameters andprotocols for broadcast, point to point, vehicle to vehicle,and vehicle to point communication in the ITS sector—IPv6Networking,” Working draft, February 2009.

[5] R. Bossom, R. Brignolo, T. Ernst et al., “European ITS commu-nication architecture—overall framework—proof of conceptimplementation,” Tech. Rep. Version 2.0, EC “InformationSociety Technologies” Programme, March 2009.

[6] T. Kosch, I. Kulp, M. Bechler, M. Strassberger, B. Weyl, andR. Lasowski, “Communication architecture for cooperativesystems in Europe,” IEEE Communications Magazine, vol. 47,no. 5, pp. 116–125, 2009.

[7] Z. Chang, G. Gaydadjiev, and S. Vassiliadis, “Routing proto-cols for mobile ad-hoc networks: current development andevaluations,” in Proceedings of the Annual Workshop on Cir-cuits, Systems and Signal Processing, pp. 489–494, Veldhoven,The Netherlands, November 2005.

[8] T. Clausen and P. Jacquet, “Optimized link state routingprotocol (OLSR),” RFC 3626 (Experimental), October 2003.

[9] L. Bokor, N. Montavont, P. Di Francesco, T. Ernst, T. Hof,and J. Korva, “ANEMONE: a pan-European testbed to validateIPv6 mobility technologies,” in Proceedings of the InternationalSymposium on Applications and the Internet, Workshop onNetwork Mobility (SAINT WONEMO ’07), Hiroshima, Japan,2007.

[10] C. Tschudin, H. Lundgren, and E. Nordstrom, “EmbeddingMANETs in the real world,” in Proceedings of the 8th IFIPConference on Personal Wireless Communications, vol. 2775 ofLecture Notes in Computer Science, pp. 578–589, September2003.

[11] V. Gonzalez, A. Los Santos, C. Pinart, and F. Milagro,“Experimental demonstration of the viability of IEEE 802.11bbased inter-vehicle communications,” in Proceedings of the4th International Conference on Testbeds and Research Infras-tructures for the Development of Networks & Communities(TridentCom ’08), pp. 1–7, Innsbruck, Austria, 2008.

[12] J. P. Singh, N. Bambos, B. Srinivasan, and D. Clawin,“Wireless LAN performance under varied stress conditionsin vehicular traffic scenarios,” in Proceedings of the 56th IEEEVehicular Technology Conference (VTC ’02), vol. 2, pp. 743–747, Vancouver, Canada, September 2002.

[13] M. Jerbi, S.-M. Senouci, and M. Al Haj, “Extensive experi-mental characterization of communications in vehicular ad

hoc networks within different environments,” in Proceedingsof the 65th IEEE Vehicular Technology Conference (VTC ’07),pp. 2590–2594, Dublin, Ireland, 2007.

[14] M. Tsukada, I. Ben-Jemaa, H. Menouar, W. Zhang, M. Goleva,and T. Ernst, “Experimental evaluation for IPv6 over VANETgeographical routing,” in Proceedings of the 6th InternationalWireless Communications and Mobile Computing Conference,pp. 736–741, Caen, France, June 2010.

[15] C. Ng, T. Ernst, E. Paik, and M. Bagnulo, “Analysis ofmultihoming in network mobility support,” RFC 4980 (Infor-mational), October 2007.

[16] R. Wakikawa, V. Devarapalli, G. Tsirtsis, T. Ernst, and K.Nagami, “Multiple Care-of Addresses Registration,” RFC 5648(Proposed Standard), October 2009.

[17] D. Johnson, C. Perkins, and J. Arkko, “Mobility support inIPv6,” RFC 3775 (Proposed Standard), June 2004.

[18] C. Ng, P. Thubert, M. Watari, and F. Zhao, “Networkmobility route optimization problem statement,” RFC 4888(Informational), July 2007.

[19] C. Ng, F. Zhao, M. Watari, and P. Thubert, “Networkmobility route optimization solution space analysis,” RFC4889 (Informational), July 2007.

[20] R. Baldessari, T. Ernst, A. Festag, and M. Lenardi, “Automotiveindustry requirements for NEMO route optimization,” Jan-uary 2009.

[21] W. Eddy, W. Ivancic, and T. Davis, “Network MobilityRoute Optimization Requirements for Operational Use inAeronautics and Space Exploration Mobile Networks,” RFC5522 (Informational), October 2009.

[22] R. Wakikawa, K. Okada, R. Koodli, A. Nilsson, and J. Murai,“Design of vehicle network: mobile gateway for MANET andNEMO converged communication,” in Proceedings of the 2ndACM International Workshop on Vehicular Ad Hoc Networks(VANET ’05), pp. 81–82, September 2005.

[23] J. Lorchat and K. Uehara, “Optimized inter-vehicle communi-cations using NEMO and MANET,” in Proceedings of the 2ndInternational Workshop on Vehicle-to-Vehicle Communications(V2VCOM ’06), July 2006.

[24] R. Wakikawa, T. Clausen, B. McCarthy, and A. Petrescu,“MANEMO topology and addressing architecture,” July 2007.

[25] R. Wakikawa, P. Thubert, T. Boot, J. Bound, and B. McCarthy,“Problem statement and requirements for MANEMO,” July2007.

[26] C. J. Bernardos, I. Soto, M. Calderon, F. Boavida, and A.Azcorra, “VARON: vehicular ad hoc route optimisation forNEMO,” Computer Communications, vol. 30, no. 8, pp. 1765–1784, 2007.

[27] C. Pradalier, J. Hermosillo, C. Koike, C. Braillon, P. Bessiere,and C. Laugier, “The CyCab: a car-like robot navigatingautonomously and safely among pedestrians,” Robotics andAutonomous Systems, vol. 50, no. 1, pp. 51–67, 2005.

[28] M. Tsukada and T. Ernst, “Vehicle communication experimentenvironment with MANET and NEMO,” in Proceedings ofthe International Symposium on Applications and the Internet,Workshop on Network Mobility (SAINT WONEMO ’07), p. 45,Hiroshima, Japan, January 2007.

[29] J. Santa, M. Tsukadat, T. Emstt, and A. F. Gomez-Skarmeta,“Experimental analysis of multi-hop routing in vehicular ad-hoc networks,” in Proceedings of the 2nd Workshop on Exper-imental Evaluation and Deployment Experiences on VehicularNetworks in Conjunction with the International Conference onTestbeds and Research Infrastructures for the Development ofNetworks and Communities and Workshops (TridentCom ’09),Washington, DC, USA, April 2009.


[30] C. Larsson, M. Eriksson, K. Mitsuya, K. Tasaka, and R. Kuntz,“Flow Distribution Rule Language for Multi-Access Nodes,”IETF, Work in progress, draft-larsson-mext-flow-distribution-rules-02, February 2009.

[31] H. Soliman, G. Tsirtsis, N. Montavont, G. Giaretta, andK. Kuladinithi, “Flow Bindings in Mobile IPv6 and NemoBasic Support,” IETF, Work in progress, draft-ietf-mext-flow-binding-04, November 2009.


Research Article

Traffic Data Collection for Floating Car Data Enhancement inV2I Networks

D. F. Llorca, M. A. Sotelo, S. Sanchez, M. Ocana, J. M. Rodrıguez-Ascariz,and M. A. Garcıa-Garrido

Department of Electronics, University of Alcala, Ctra. N-II Km. 33, Alcala de Henares, C.P. 28871 Madrid, Spain

Correspondence should be addressed to D. F. Llorca, [email protected]

Received 19 November 2009; Accepted 5 July 2010

Academic Editor: Shahrokh Valaee

Copyright © 2010 D. F. Llorca et al. This is an open access article distributed under the Creative Commons Attribution License,which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

This paper presents a complete vision-based vehicle detection system for floating car data (FCD) enhancement in the context ofvehicular ad hoc networks (VANETs). Three cameras (side-, forward- and rear-looking cameras) are installed onboard a vehiclein a fleet of public buses. Thus, a more representative local description of the traffic conditions (extended FCD) can be obtained.Specifically, the vision modules detect the number of vehicles contained in the local area of the host vehicle (traffic load) andtheir relative velocities. Absolute velocities (average road speed) and global positioning are obtained after combining the outputsprovided by the vision modules with the data supplied by the CAN Bus and the GPS sensor. This information is transmitted bymeans of a GPRS/UMTS data connection to a central unit which merges the extended FCD in order to maintain an updatedmap of the traffic conditions (traffic load and average road speed). The presented experiments are promising in terms of detectionperformance and computational costs. However, significant effort is further necessary before deploying a system for large-scalereal applications.

1. Introduction

Floating car data (FCD) refers to technology that collectstraffic state information from a set of individual vehicleswhich float in the current traffic. Each vehicle, which canbe seen as a moving sensor that operates in a distributednetwork, is equipped with positioning (GPS) and commu-nication (GSM, GPRS, UMTS, etc.) systems, transmittingits location, speed, and direction to a central control unitthat integrates the information provided by each one of thevehicles.

FCD systems are being increasingly used in a variety ofimportant applications since they overcome the limitationsof fixed traffic monitoring technologies (installation andmaintenance costs, lack of flexibility, static nature of theinformation, etc.). We refer to [1] for general backgroundconcerning the most representative FCD activities in Japan,Europe, and the United States.

FCD can be used by the public sector to collect roadtraffic statistics and to carry out real-time road traffic control.The information provided by FCD systems can be supplied to

individual drivers via dynamic message signs, PDA devices,satellite navigation systems, or mobile phones, includingdynamic rerouting information. Thus, drivers would beable to make more informed choices, spending less time incongested traffic. In addition, the knowledge of the currenttraffic situation can be also used to estimate time of arrivalof a fleet of public transport vehicles and, furthermore,to plan and coordinate the movements of the fleet (fleetmanagement) so that driving assignments can be carried outmore efficiently. Besides previous applications, the use ofFCD entails environmental benefits since it can be used toreduce fuel consumption and emissions.

The basic data provided by FCD systems (vehicle loca-tion, speed, and direction) can be enriched using newonboard sensors (ambient temperature, humidity and light,windshield wiper status, fog light status, fuel consump-tion, emissions, tire pressure, suspension, emergency brake,etc.) which are centralized by means of the controller-area-network (CAN) bus. Such data can be exploited toextend the information horizon including traffic, weather,road management, and safety applications [1]. In addition,


computer vision systems can be included in order toimprove the automatic detection of potentially interestingevents and to document them by sending extended data[2].

In order to provide ubiquitous coverage of the entire roadnetwork, a minimum representation of the total passengercar fleet has to be used, since each moving sensor (eachvehicle) only supplies information about its status. Thefact that everyday road users have to be asked to shareinformation regarding their movements and speeds arisesprivacy issues that have to be addressed. Many potential roadtravellers may be reluctant to join FCD projects because ofviolations of their privacy due to permanent traceability orpossible liability in case of speed limit violations. Thus, thefundamental concept for FCD systems calls for no identi-fication information to be sent with the basic data, whichcan be easily implemented from a technical perspective. Forexample, in [3] a general method for anonymization of FCDby deriving pseudonyms for trips is presented.

Another approach consists of using the information sup-plied by a specific fleet of vehicles, rather than informationcoming from individual road users. Taxis or public transportbuses can be used due to the extended periods of timethey spend on the urban road network. Although taxis andbuses provide a major source of innercity traffic informationbecause of the time they spend mobile, they have limitations.Problems arise if the taxi drivers, through detailed knowledgeof the local road network, take steps to avoid congested areaswhich will not be reported [4]. Traffic load perception maybe lower than the actual one if reserved taxis or buses lanesare used. On the contrary, privacy issues are not as criticalas before, especially when using a fleet of public transportbuses.

This paper presents a complete vision-based vehicledetection system onboard a fleet of public transport buseswith the aim of improving the data collected in FCDapplications. The proposed system has been developed in theframework of the GUIADE project. Three cameras coveringthe local environment of the vehicle are used: forward-rear-and side-looking cameras. The system obtains under certainconstraints, such as good weather and daytime conditions,the number of vehicles in the local range of the bus as wellas their relative position and velocity. This information iscombined with the data provided by regular FCD systems(global location, speed, and direction), obtaining a moredetailed description of the local traffic load and the averagespeed. The communication system between the vehicles andthe central control unit is based on wireless technologyvia GPRS/UMTS cellular protocols. Finally, the central unitintegrates the data collected by the fleet in order to generateupdated traffic status maps.

The remainder of this paper is organized as follows: thedescription of the system including the wireless communica-tion scheme is summarized in Section 2. Section 3 describesthe vision-based vehicle detection system as well as the spatialand temporal integration of the collected data. Experimentalresults that validate the proposed approach are presented inSection 4. Finally, conclusions and future works are discussedin Section 5.

2. System Descripction

The proposed FCD architecture can be seen in Figure 1.Floating car data is supplied by a fleet of public transportbuses which corresponds to an inner-city bus line. Each vehi-cle is equipped with a global positioning system (GPS), wire-less communication interfaces (GPRS/UMTS and WLANIEEE 802.11) and a complete vision-based vehicle detectionsystem.

The vehicle-to-infrastructure (V2I) communication sys-tem is based on the geographic coverage provided by cellularnetworks. General packet radio service (GPRS) and universalmobile telecommunications system (UMTS) are used toconnect each vehicle with the central control unit. Eachvehicle provides information that can be divided in threemain groups.

(1) Standard FCD information: vehicle identifier (2bytes), timestamp (11 bytes), GPS position (8 bytes),speed (2 bytes), and direction (2 bytes).

(2) Vehicle status information: ambient temperature (2bytes), humidity (2 bytes), light (2 bytes), windshieldwiper status (1 byte), fog light status (1 byte), fuelconsumption (4 bytes), and emissions (4 bytes).

(3) Extended FCD information: globally referenced aver-age traffic load (2 bytes) and average road speed for ameasured segment travel time (2 bytes).

As can be observed, the total message size per vehicle is45 bytes. The extended FCD information is supplied to thecentral unit at a frequency of 1 Hz. Accordingly, the band-width currently demanded by vehicular communication inthe communication channel, that is, the vehicle throughput,is 360 bps without overheads. This value can be considerednegligible taking into account the available bandwidth andthe proposed FCD architecture.

The central control unit integrates the informationprovided by each one of the vehicles in order to computeupdated traffic and weather maps which will be used for fleetmanagement tasks as well as to estimate the time of arrival.

The vehicle-to-vehicle (V2V) communication systemis defined as a backup communication system based ona wireless-fidelity (WiFi) IEEE 802.11a/b/g interface. Insituations where the cellular network is not working, in-range vehicles will exchange the most updated informationavailable.

One of the main advantages of the proposed approachis that it does not need to deal with privacy issues since thefloating vehicles correspond to a fleet of public transportbuses.

3. Vision-Based Traffic Detection System

In this section, we present the main contribution of thiswork: a complete vision-based traffic detection system whichenhances the data supplied by standard FCD systems.The benefits of using computer vision instead of othertechnologies such as radar-based systems can be summarizedas follows. Computer vision systems can compensate for


FLEET

V2V WLANbackup

communication

Globalpositioning

system

3G/GPRS/UMTScellular

communications

Central control unit

Figure 1: Overview of the proposed FCD architecture.

the lower angular resolution of the low-cost radar and theincreased appearance of ghost radar targets (guard-rails,railings, lamp posts, reflections, etc.). These false positivesare relevant and they cannot simply be ignored. The camerahas very good angular resolution and can be used todetermine height, width, and lateral speed of the target.Pattern recognition can be used to classify the object andeven weakly reflective targets such as pedestrians can bedetected. Moreover, the cost of a vision system is significantlylower than the cost saved by using the simpler radar. A visionsystem, in addition to overcoming cost reduction problems,can contribute to the system features such as road analysisand scene understanding.

Each individual vehicle is equipped with three FireWirecameras (forward-, rear- and side-looking cameras) thatcover the local environment of the bus (see Figure 2). A com-mon hardware trigger synchronizes the image acquisition ofthe three cameras and an onboard PC houses the computervision software.

Each individual vehicle detection system provides infor-mation about the number of detected vehicles and boththeir relative position and speed. These results are combinedwith the GPS measurements and the data provided by theCAN bus in order to provide globally referenced trafficinformation. This scheme is described in Figure 3.

The layers of the proposed architecture of the three visionmodules are conceptually the same: lane detection, vehiclecandidates selection, vehicle recognition, and tracking. The first

step of each one of the vision systems consists of reducing thesearching space in the image plane in an intelligent mannerin order to increase the performance of the vehicle detectionmodule. Accordingly, road lane markings are detected andused as the guidelines that drive the vehicle searching process(see Figure 4). The area contained by the limits of the lanes isscanned in order to find vehicle candidates that are passedon to the vehicle recognition modules. Thus, the rate offalse positives is reduced. In case that no lane markings aredetected, a basic region of interest is used instead covering thefront, rear, and side parts of the vehicle. Finally, a trackingstage is implemented using Kalman filtering techniques.

3.1. Lane Detection. An attention mechanism is necessary inorder to filter out inappropriate candidate windows basedon the lack of distinctive features, such as horizontal edgesand vertical symmetrical structures, which are essentialcharacteristics of road vehicles. This has the positive effectof decreasing both the total computation time and therate of false positive detections. Lane markings are detectedusing gradient information in combination with a localthresholding method which is adapted to the width of theprojected lane markings. Then, clothoid curves are fittedto the detected markings. The algorithm scans up to 25lines in the candidates searching area, from 2 meters infront of the camera position to the maximum range inorder to find the lane marking measurements. The proposedmethod implements a nonuniform spacing search that


GPSForwardlookingcamera

Rearlookingcamera

Side looking camera

Figure 2: Main vehicle sensors: three cameras (forward, rear, and side looking cameras) and a global positioning system.

Forwardvehicle

detection

Rearvehicle

detection

Sidevehicle

detection

GPSHardwaretriggering

CANbus

Data fusion(spatial and temporal integration)

Figure 3: Stages of the vision-based traffic detection system.

Forward lanedetection

Forward lanedetection

Side lanedetection

Rear lanedetection

Rear lanedetection

Figure 4: Rear, side, and forward lane detection.

reduces certain instabilities in the fitted curve. The final statevector is composed of 6 variables [5] for each lane on theroad

x = [coh, c1h, cov, c1v, xo, θo,wo]T , (1)

where coh and c1h represent the clothoid horizontal curvatureparameters, cov and c1v stand for the clothoid vertical curva-ture parameters, while xo, θo, and wo are the lateral error andorientation error with regard to the centre of the lane andthe width of the lane, respectively. The clothoid curves arethen estimated based on lane marking measurements using aKalman filter for each lane.

Apart from the detected road lanes additional virtuallanes have been considered so as to cope with situationsin which a vehicle is located between two lanes (e.g., ifit is performing a change lane manoeuvre). Virtual lanesprovide the necessary overlap between lanes, avoiding bothmisdetections and double detections caused by the twohalves of a vehicle being separately detected as two potentialvehicles. A virtual lane is located to provide overlap betweentwo adjoining lanes. Figure 5 provides some examples of lanemarkings detection in real outdoor scenarios. Detected lanesdetermine the vehicle searching area and help reduce falsepositive detections. In case no lane markings are detectedby the system, fixed lanes corresponding to a straight roadmodel are assumed instead.

3.2. Side Vehicle Detection. Side vehicle detection module [6]relies on the computation of optical flow. In order to reducecomputational time, optical flow is computed only on Cannypoints in the image. Canny edge pixels are consequentlymatched and grouped together in order to detect clustersof pixels that can be considered as candidate vehicles in theimage. Classical clustering techniques are used to determinegroups of pixels, as well as their likelihood to form a singleobject. Even after pixels clustering, some clusters can stillbe clearly regarded as belonging to the same real object.A second grouping stage (double-stage) is then carried outamong different clusters in order to determine which ofthem can be further merged into a single blob. For thispurpose, simple distance criteria are considered. Two objectsthat are very close to each other are finally grouped togetherin the same cluster. The reason for computing a two-stageclustering process relies on the fact that by selecting a smalldistance parameter in the first stage, interesting informationabout clusters in the scene can be obtained. Otherwise, usinga large distance parameter in the single clustering process,highly gross clusters would have been achieved, losing allinformation about the granular content of the points thatprovide optical flow in the image.

The selected clusters constitute the starting point forlocating candidate vehicles in the image. For that purpose,the detected positions of clusters are used as a seed pointto search for a collection of horizontal edges that couldpotentially represent the lower part of a car. The candidateis located on the detected horizontal edges that meetcertain conditions of entropy and vertical symmetry. Someof the most critical aspects in side vehicle detection are


(a) (b) (c)

Figure 5: Vehicle searching area as a result of the lane markings analysis for forward, rear and side modules.

HOG/SVMclassification

Yes No

WarningYesNoNo

No

No

YesYes

YesVehiclebehaviour

Pre-detector

Frontalpart

detection

Masscenter

detection

Tracking(relative position

and speed)

Figure 6: Side vehicle detection flow diagram.

subsequently listed: (1) shadows on the asphalt due tolampposts, other artefacts or a large vehicle overtaking theego-vehicle on the right lane; (2) self-shadow reflected on theasphalt (especially problematic in sharp turns like in round-about points), or self-shadow reflected on road protectionfences; (3) robust performance in tunnels; and (4) avoidingfalse alarms due to vehicles on the third lane.

The flow diagram of the two-stage detection algorithmis depicted in Figure 6. As can be observed, there is a pre-detector that discriminates whether the detected object isbehaving like a vehicle or not. If so, the frontal part of thevehicle is located in the region of interest. In addition, thevehicle mass centre is computed. In case the frontal part ofthe vehicle is properly detected and its mass centre can alsobe computed, a final warning message is issued. After beinglocated, vehicle candidates are classified by using a linearSVM classifier [7] with HOG features [8] previously trained

with the samples obtained from real road images, and at thatpoint vehicle tracking starts. Tracking is stopped when thevehicle gets out of the image. Sometimes, the shadow of thevehicle remains in the image for a while after the vehicledisappears from the scene, provoking the warning alarm tohold on for 1 or 2 seconds. This is not a problem, however,since the overtaking car is running in parallel with the ego-vehicle during that time although it is out of the image scene.Thus, maintaining the alarm in such cases turns out to be adesirable side effect.

Figure 7 shows an example of blind spot detection in asequence of images. The indicator depicted in the upper-right part of the figure toggles from green to blue whena vehicle enters the blind spot area (indicated by a greenpolygon). A blue bounding box depicts the position of thedetected vehicle.

3.3. Forward and Rear Vehicle Detection. Forward- and rear-looking vehicle detection systems share the same algorithmiccore. The attention mechanism sequentially scans each roadlane from the bottom to the maximum range looking fora set of features that might represent a potential vehicle.Firstly, the vehicle contact point is searched by meansof the top-hat transformation. This operator allows thedetection of contrasted objects on nonuniform backgrounds[9]. There are two different types of top-hat transformations:white hat and black hat. The white hat transformationis defined as the residue between the original image andits opening (◦ operator). The black hat transformation isdefined as the residue between the closing (• operator) andthe original image. The white and black hat transformationsare analytically defined as follows:

WHT

(x, y

) = (f − f ◦ b)(x, y

)White Hat, (2)

BHT

(x, y

) = (f • b − f

)(x, y

)Black Hat. (3)

The opening operator (◦) is defined as the dilation of theerosion and the closing operator (•) is defined as the erosionof the dilation (for more details see [10]). In our case we usethe white hat operator (2) since it enhances the boundarybetween the vehicles and the road [11]. Horizontal contactpoints are preselected if the number of white top-hat featuresis greater than a configurable threshold. Then, candidates arepreselected if the entropy of Canny points is high enough


(a) (b) (c) (d)

Figure 7: Example of the side-vehicle detection module (also called blind spot detection) in a sequence of images. The indicator in theupper-right part of the figure toggles from green to blue when a car is detected in the blind spot.

(a) (b) (c)

Figure 8: From left to right: original image; contact point detection on white top-hat image; candidate preselected with high entropy ofCanny points.

(a)

(b)

Figure 9: Canny images after adaptive thresholding.


(a)

(b)

(c)

Figure 10: Upper row: gray level symmetry; Middle row: vertical edges symmetry; Lower row: horizontal edges symmetry.

(a) (b)

Figure 11: (a) Overlapped candidates. (b) Nonmaximum suppression results.

for a region defined by means of perspective constraints andprior knowledge of target objects (see Figure 8).

Before computing the Canny features, an adaptivethresholding method is applied. This process is based onan iterative algorithm that gradually increases the contrastof the image, and compares the number of Canny pointsobtained in the contrast increased image with the number

of edges obtained in the current image. If the number ofCanny features in the actual image is higher than in thecontrast increased image the algorithm stops. Otherwise,the contrast is gradually increased and the process resumed.This adaptive thresholding method permits to obtain ro-bust image edges, as depicted in the examples provided inFigure 9.


(a)

(b)

Figure 12: Forward data set. (a) positive samples (vehicles). (b) negative samples.

(a)

(b)

Figure 13: Rear data set. (a) positive samples (vehicles). (b) negative samples.

(a) (b)

Figure 14: Linear SVM with HOG features classification examples: nonvehicle (red) and vehicle (green).


(a) (b)

(c) (d)

Figure 15: Data association by features matching. (a, b) Harris features on image t. (c, d) matched Harris features on image t + 1.

0

1

2

3

4

5

6

Nu

mbe

rof

veh

icle

sin

ran

ge

0 100 200 300 400 500 600 700 800

Frames

Ground truthDetected vechicles (Side/forward/rear)

Figure 16: Number of vehicles detected by the three vision modules compared with the manually labeled ground truth in a real sequence.

In a second step, vertical edges (Sv), horizontal edges(Sh), and grey level (Sg) symmetries are obtained, so that,candidates will only pass to the next stage if their symmetriesvalues are greater than a threshold. The vertical and hori-zontal edges symmetries are computed as listed inAlgorithm1. The grey level symmetry computation procedure isshown inAlgorithm 2. Some examples of the three types ofsymmetries are depicted in Figure 10.

Symmetry axes are linearly combined to obtain thefinal position of the candidate. Finally, a weighted vari-able is defined as a function of the entropy of Cannypoints, the three symmetry values and the distance to thehost vehicle. We use this variable to apply a nonmaxi-mum suppression process per lane which removes over-lapped candidates. An example of this process is shown inFigure 11.


(a) (b)

Figure 17: Examples with strong illumination changes after passing beneath a bridge.

(1) Initialize Acc0,...,ROIWIDTH = 0(2) For i = 0, ...,ROIHEIGHT

(3) For each pair of vertical/horizontal edge pixels (x1, i) and (x2, i)(4) Acc(x1+x2)/2 + +(5) Sv,h = arg

i

max(Acci/Sv,h,MAX)

Algorithm 1: Vertical and horizontal edges symmetries computation procedure.

The selected candidates are classified by means of alinear SVM classifier [7], in combination with histograms oforiented gradients features [8]. We have developed and testedtwo different classifiers depending on the module (forwardand rear classifiers). All candidates are resized to a fixedsize of 64 × 64 pixels to facilitate the features extractionprocess. The rear-SVM classifier is trained with 2000 samplesand tested with 1000 samples (1/1 positive/negative ratio)whereas the forward-SVM classifier is trained with 3000samples and tested with 2000 samples (1/1 positive/negativeratio). Figures 12 and 13 depict some positive and negativesamples of the forward and rear training and test data sets,respectively. Figure 14 shows a couple of examples of vehicledetection after linear SVM classification with HOG features.

After detecting consecutively an object classified asvehicle a predefined number of times (empirically set to3 in this work), data association and tracking stages aretriggered. The data association problem is addressed by usingfeature matching techniques. Harris features are detectedand matched between two consecutive frames as depicted inFigure 15.

Tracking is implemented using Kalman filtering tech-niques [12]. For this purpose, a dynamic state model anda measurement model must be defined. The proposeddynamic state model is simple. Let us consider the statevector xn, defined as follows:

xn =[u, v,w,h, u, v, w, h

]T. (4)

In the state vector x and y are the respective horizontal andvertical image coordinates for the top left corner of everyobject, and w and h are the respective width and height inthe image plane, a dynamical model equation can be writtenlike this

xn+1 = A · xn + ωn

=

⎛⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎝

1 0 0 0 Δt 0 0 00 1 0 0 0 Δt 0 00 0 1 0 0 0 Δt 00 0 0 1 0 0 0 Δt0 0 0 0 1 0 0 00 0 0 0 0 1 0 00 0 0 0 0 0 1 00 0 0 0 0 0 0 1

⎞⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎠

⎛⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎝

uvwhuvw

h

⎞⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎠n

+ ωn.(5)

In the model, Δt is the simple time, A represents the systemdynamics matrix and ωn is the noise associated to the model.Although the definition of A is simple, it proves to behighly effective in practice since the real time operationof the system permits to assure that there will not begreat differences in distance for the same vehicle betweenconsecutive frames. The model noise has been modelled as afunction of distance and camera resolution. The state modelequation is used for prediction in the first step of the Kalmanfilter. The next step is to define the measurement model. The


(1) For each possible symmetry axis xi initializes Si = 0(2) For j = 0, ...,ROIHEIGHT

(3) For k = 0, ...,ROIWIDTH/2(4) If abs(image[ j][xi+k] − image[ j][xi−k]) < Δ(5) Si + +(6) Sg = arg

i

max(Si/(areaROI /2))

Algorithm 2: Gray level symmetry computation procedure.

00.10.20.30.40.50.60.70.80.9

1

Traffi

clo

ad(L

i)

0 100 200 300 400 500 600 700 800

Frame

Figure 18: Traffic load Li at every frame in a real sequence.

0

20

40

60

80

100

120

140

Ave

rage

road

spee

dat

fram

ei

(km

/h)

0 100 200 300 400 500 600 700 800

Frame

Figure 19: Average road speed vi at every frame in a real sequence.

measurement vector is defined as zn = [u, v,w,h]T . Then,the measurement model equation is established as follows:

zn+1 = H · xn + vn

=

⎛⎜⎜⎜⎝

1 0 0 0 0 0 0 00 1 0 0 0 0 0 00 0 1 0 0 0 0 00 0 0 1 0 0 0 0

⎞⎟⎟⎟⎠

⎛⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎝

uvwhuvw

h

⎞⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎠n

+ vn.(6)

In last equation, H represents the measurement matrix andvn is the noise associated to the measurement process. Thepurpose of the Kalman filtering is to obtain a more stableposition of the detected vehicles. Besides, oscillations invehicles position due to the unevenness of the road makesv coordinate of the detected vehicles change several pixels up

00.10.20.30.40.50.60.70.80.9

1

Ave

rage

traffi

clo

ad

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31

Time (seconds)

Figure 20: Average traffic load at every second in a real sequence.

0

20

40

60

80

100

120

140

Ave

rage

road

spee

d(k

m/h

)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31

Time (secconds)

Figure 21: Average road speed at every second in a real sequence.

or down. This effect makes the distance detection unstable,so a Kalman filter is necessary for minimizing these kinds ofoscillations.

3.4. FCD Integration. As depicted in Figure 3, the FCDintegration or Data Fusion module uses three sources of data:the measurements provided by the GPS, the data suppliedby the CAN bus, and the output obtained from the threevision-based vehicle detection modules. Whereas the GPSand the CAN bus sample frequency is 1 Hz, the vision-basedsystem operates in real-time at 25 frames per second (25 Hz).The proposed data fusion scheme provides information atthe lowest sample frequency (1 Hz) covering two consecutiveGPS measurements, the vehicle speed vhi (via CAN bus) andthe outputs of the vision module.

The outputs of the side, forward, and rear vehicledetection systems at frame i are the number of detectedvehicles Ni and their corresponding distances to the host

vehicle d(k)i (note that d is used here as a distance/range


Figure 22: GPS trajectory and the corresponding traffic load computed at the central unit (the aerial image has been obtained from GoogleEarth).

measurement). These outputs are combined to cover thewhole local environment of the vehicle. The traffic load atframe i is given by next expression

Li = Ni + 1NMAX

, (7)

where NMAX is the maximum number of vehicles in rangethat can be detected by the three systems (in our case NMAX isdefined as 9 or 13 for two lanes and three lanes roads, resp.).The average road speed at frame i is computed as follows:

vi = 1Ni + 1

⎛⎝Ni−1∑

k=0

⎛⎝(d

(k)i − d

(k)i−1

)Δt

+ vhi

⎞⎠ + vhi

⎞⎠

= 1(Ni + 1)Δt

Ni−1∑k=0

(d

(k)i − d

(k)i−1

)+ vhi ,

(8)

where d(k)i and d

(k)i−1 represent the distance between the host

vehicle and vehicle k at frames i and i − 1, respectively, Δtcorresponds to the sample time, vhi is the host vehicle speedprovided by the CAN bus, and Ni is the number of detectedvehicles. Note that the distance values correspond to filteredmeasurements since they are obtained from the first twoelements of the Kalman filter state vector (u and v) usingknown camera geometry and ground-plane constraints.

Two consecutive GPS measurements define both a spatialand a temporal segment. The temporal segment correspondsto the GPS sample time (1 second), and the spatial segmentwill be defined as the globally referenced trajectory betweenthe two GPS measurements. In order to obtain the extendedFCD information (i.e., the road traffic load and the roadspeed) for this spatio/temporal segment we integrate thevalues supplied by the vision modules during 25 consecutiveframes. With this approach a dense coverage of the roadtraffic load and the road speed can be assured for host vehiclespeeds up to 180 km/h since the total range of the vision

module covers more than 50 m (25 meters for both the rearand the forward looking modules; the side range covers upto two third parts of the bus length in the adjacent lane).Obviously this maximum speed will never be exceeded by apublic bus. This approach facilitates further map-matchingtasks since the extended FCD information between twoconsecutive points will always be globally referenced.

4. Experimental Results

The system was implemented on a PC Core 2 Duo at 3.0 GHzand tested in real traffic conditions using CMOS cameraswith low-resolution images (320 × 240). After training andtest, a tradeoff point has been chosen at detection rate (DR)of 95% and false positive rate (FPR) of 5% for the rear-SVM classifier and at DR of 90% and FPR of 6% for theforward-SVM classifier. We have to note that these numbersare obtained in an offline single-frame fashion, so that, theywill be improved in subsequently stages. In addition, the lanedetection system reduces the searching area and the numberof false candidates passed to further stages.

In order to validate the proposed vision-based vehicledetection system as an extended source for FCD applicationswe have recorded several video sequences in real trafficconditions, and we have manually labeled the number ofvehicles in range at every frame (a total of 800 frames). Thespeed of the host vehicle was around 90 km/h so the lengthof the traveled route was 1 km approximately. Both the trafficload Li and the average road speed vi are computed at everyframe using (7) and (8). Figure 16 shows the ground truthand the number of vehicles detected in range. Most of theerrors take places in cases where the host vehicle is passingbeneath a bridge due to strong illumination changes (seeFigure 17) and in curves or cases where there are strongchanges in the vehicle pitch, roll or camera height.

The traffic load Li and the average road speed vi atevery frame are depicted in Figures 18 and 19, respectively.


These values are provided by the vision modules at afrequency of 25 Hz. As the extended FCD information issupplied to the central unit at a frequency of 1 Hz thetraffic load and the average road speed are finally integratedduring 25 consecutive frames. These results are shown inFigures 20 and 21. We use a colour code to describe thelevel of traffic load and the road speed: green indicatesthat there is good flow/high speeds; yellow indicates thatthere is semi-dense traffic/medium speeds, and red showsdense traffic/slow speeds (traffic jams). After combining theresults with the GPS measurements we can obtain the trafficload in universal transverse mercator (UTM) coordinates, asdepicted in Figure 22 (note that map-matching is not carriedout; the aerial image has been obtained from Google Earth).

5. Concluding Remarks

This paper presented a complete vision-based vehicle detec-tion system that enhances the data supplied by FCD systemsin the context of vehicular ad hoc networks. The systemis composed of three vision subsystems (side, forward andrear subsystems) that detect the traffic load and the relativevelocities of the vehicles contained in the local area ofthe host vehicle. Under certain constraints, such as goodweather and daytime conditions, absolute velocities, andglobal positioning are obtained after combining the outputsprovided by the vision modules with the outputs suppliedby the CAN Bus and the GPS sensor. Standard FCD systemsprovide the vehicle position, speed, and direction. Theproposed approach extends this information by includingmore representative measurements corresponding to thetraffic load and the average road speed.

In order to cover the entire road network, the proposedvision-based system is defined for being installed onboarda fleet of public buses where privacy is a minor issue.The extended packets collected by each moving vehicle aretransmitted to the central unit by means of a GPRS/UMTSdata connection. The central unit merges the extended FCDin order to maintain an updated map of the traffic conditions(traffic load and average road speed).

The presented experiments are promising in terms ofdetection performance and computational costs. However,significant effort is further necessary before deploying asystem for large-scale real applications. For this purpose, newexperiments will be carried out merging the data collected bymore than one vehicle, including map-matching techniquesand further analysis on V2I and V2V communications (e.g.,using repetition based MAC protocols [13]). In addition,the proposed vision-based vehicle detection system will beextended to deal with complex weather conditions (e.g., wetor snowy roads) as well as night-time conditions.

Acknowledgments

This work has been supported by the Spanish Ministryof Science and Innovation by means of Research GrantTRANSITO TRA2008-06602-C03 and Spanish Ministry ofDevelopment by means of Research Grant GUIADE P9/08.

References

[1] R. Bishop, Intelligent Vehicle Technologies and Trends, ArtechHouse, Boston, Mass, USA, 2005.

[2] S. Messelodi, C. M. Modena, M. Zanin et al., “Intelligentextended floating car data collection,” Expert Systems withApplications, vol. 36, no. 3, part 1, pp. 4213–4227, 2009.

[3] S. Rass, S. Fuchs, M. Schaffer, and K. Kyamakya, “How toprotect privacy in floating car data systems,” in Proceedingsof the 5th ACM International Workshop on Vehicular Inter-Networking (VANET ’08), pp. 17–22, September 2008.

[4] P. Day, J. Wu, and N. Poulton, Briefing Note on Floating CarData, ITS Internacional, 2006.

[5] M. A. Sotelo, J. Nuevo, L. M. Bergasa, M. Ocana, I. Parra,and D. Fernandez, “Road vehicle recognition in monocularimages,” in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE ’05), pp. 1471–1476, June 2005.

[6] M. A. Sotelo and J. Barriga, “Blind spot detection using visionfor automotive applications,” Journal of Zhejiang University:Science A, vol. 9, no. 10, pp. 1369–1372, 2008.

[7] C. J. Christopher Burges, “A tutorial on support vectormachines for pattern recognition,” Data Mining and Knowl-edge Discovery, vol. 2, no. 2, pp. 121–167, 1998.

[8] N. Dalal and B. Triggs, “Histograms of oriented gradientsfor human detection,” in Proceedings of the IEEE ComputerSociety Conference on Computer Vision and Pattern Recognition(CVPR ’05), pp. 886–893, June 2005.

[9] Y. Derong, Z. Yuanyuan, and L. Dongguo, “Fast computationof multiscale morphological operations for local contrastenhancement,” in Proceedings of the 27th Annual InternationalConference of the Engineering in Medicine and Biology Society(IEEE-EMBS ’05), pp. 3090–3092, September 2005.

[10] E. R. Dougherty, An Introduction to Morphological ImageProcessing, SPIE Optical Engineering Press, 1992.

[11] D. Balcones, D. F. Llorca, M. A. Sotelo, et al., “Real-time vision-based vehicle detection for rear-end collisionmitigation systems,” in Proceedings of the 12th InternationalConference on Computer Aided Systems Theory (EUROCAST’09), vol. 5717 of Lecture Notes in Computer Science, pp. 320–325, 2009.

[12] R. E. Kalman, “A new approach to linear filtering and predic-tion problems,” Journal of Basic Engineering, Series D, vol. 82,pp. 35–45, 1960.

[13] B. Hassanabadi, L. Zhang, and S. Valaee, “Index codedrepetition-based MAC in vehicular ad-hoc networks,” inProceedings of the 6th IEEE Conference on Consumer Commu-nications and Networking Conference, 2009.


Research Article

Improvement of Adaptive Cruise Control Performance

Shigeharu Miyata,1 Takashi Nakagami,2 Sei Kobayashi,2 Tomoji Izumi,2 Hisayoshi Naito,2

Akira Yanou,3 Hitomi Nakamura,1 and Shin Takehara1

1 Department of Intelligent Mechanical Engineering, Faculty of Engineering, Kinki University, 1, Takaya-Umenobe,Higashihiroshima, Hiroshima 739-2116, Japan

2 Vehicle Development Division, Vehicle System Development Department, Mazda Motor Corporation, 3-1,Shinchi, Aki-Fuchu, Hiroshima 730-8670, Japan

3 Department of Intelligent Mechanical Systems, Graduate School of Natural Science and Technology,Okayama University, 3-1-1, Tsushima-naka, Kita-ku, Okayama 700-8530, Japan

Correspondence should be addressed to Shigeharu Miyata, [email protected]

Received 14 October 2009; Revised 7 June 2010; Accepted 24 August 2010


Copyright © 2010 Shigeharu Miyata et al. This is an open access article distributed under the Creative Commons AttributionLicense, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properlycited.

This paper describes the Adaptive Cruise Control system (ACC), a system which reduces the driving burden on the driver. TheACC system primarily supports four driving modes on the road and controls the acceleration and deceleration of the vehicle inorder to maintain a set speed or to avoid a crash. This paper proposes more accurate methods of detecting the preceding vehicle byradar while cornering, with consideration for the vehicle sideslip angle, and also of controlling the distance between vehicles. Bymaking full use of the proposed identification logic for preceding vehicles and path estimation logic, an improvement in drivingstability was achieved.

1. Introduction

The number of traffic accidents and injuries continues toincrease year by year, and annual traffic fatalities in Japanremain at over 7000. Under these conditions, there is anurgent need for technologies which can mitigate the seriousdamage caused by car accidents, as well as prevent theaccidents themselves. Some members of this study haveplayed active roles in a study group for the Advanced SafetyVehicle (ASV), a project conducted by the Japanese Ministryof Land, Infrastructure, Transport and Tourism, and areinvolved with ASV research and development. This studygroup has proposed and studied safety technologies such asAdaptive Cruise Control (ACC) and Precrash Safety Systems[1–7]. Some of these technologies have already been put intopractical use.

An ACC system maintains the vehicle at the speed set bythe driver, and when it detects a preceding vehicle travellingat a slower speed than the driver’s vehicle, it controls thevehicle speed to match the speed of the preceding vehicle.It also performs following control to maintain the level of

distance between vehicles which was set by the driver (adistance proportional to the vehicle speed). There has beenpast research for ACC systems aimed at designing a vehiclefollowing distance control system using linear approximationand linear control logic [8]. However this research has notsucceeded in producing natural vehicle behavior that meetsthe drivers’ expectations. In addition, while a method forfollowing a preceding vehicle on curves has been proposed[9] based on the following distance measured by stereo imageprocessing and on the relative vehicle speed, this method islimited to following distances in the range of 4 m–22 m, andit is not suitable for maintaining the following distance ofapproximately 100 m that is needed for expressway driving.

When a vehicle is driving at high speed such as on anexpressway, sideslip occurs when the vehicle corners. It isknown that this sideslip operates toward the outside of theturn when the vehicle is at slow speed, and toward theinside of the turn when the vehicle is at high speed [10–12]. Research and development of 4 WS systems that steerthe rear wheels in order to reduce this sideslip and improvedriving performance have been carried out. However, no


attempts other than this paper have been found at creatingan ACC system with following control that incorporatesconsideration of the vehicle sideslip. Our research team hasalready submitted a patent application for this concept [13].

This paper describes the Adaptive Cruise Control system(ACC), a system which reduces the driving burden on thedriver. The ACC system primarily supports the four drivingmodes on the road that are described in Section 2.1, andcontrols the acceleration and deceleration of the vehicle inorder to maintain a set speed or to avoid a collision. The keyto achieving intelligent ACC control is the method used todetect and follow the preceding vehicle.

The use of obstacle detection equipment, which detectsthe preceding vehicle by means of millimeter wave radar, isalready well established [14–18]. When the obstacle detectionequipment detects a vehicle (obstacle) in the path ahead ofthe driver’s vehicle, the turning radius of the driver’s vehicle(calculated based on the values from the yaw rate sensorand steering angle sensor) is used to estimate the position ofthe preceding vehicle after the next radar scan. The systemthen judges whether or not the estimated position of thepreceding vehicle matches the position detected after theradar scan. As long as the turning radius of the driver’svehicle does not change significantly between the radarscans, then calculating the amount of offset of the detectedpreceding vehicle from its path center line allows the futureposition of the preceding vehicle to be easily estimated. Inparticular when the speeds of the driver’s vehicle and thepreceding vehicle are approximately the same, the amountof this offset can be assumed to be unchanging betweenradar scans, making it possible to easily predict the positionof the preceding vehicle. However as described above, onexpressways where the vehicles are travelling at high speed,if there is a curve where the turning radius of the driver’svehicle changes between radar scans, the sideslip that occurswhen the vehicle corners causes the amount of offset fromthe preceding vehicle path center line to change. As a result,it becomes impossible to accurately estimate the position ofthe preceding vehicle, and the driver’s vehicle accelerates,producing a potentially dangerous situation.

This ACC system that incorporates consideration of thesideslip angle makes it possible for one vehicle to reliablyfollow the preceding vehicle by performing the following twooperations. First, the system estimates the path (the radiusof curvature of the path center line) by determining thecornering radius of the vehicle based on the detected yawrate, steering wheel angle, and vehicle velocity. Second, thesystem judges whether or not the current detected vehicleis the same as the previous detected vehicle by comparingthe position of the detected vehicle with a position thatis estimated based on the calculated offset from the centerline of the path each time a radar scan occurs. Usually, theradius of curvature of the path center line is assumed to beunchanged during the period between radar scans. However,in cases of curves near an entrance or exit, the radius ofcurvature of the path actually does change during the periodbetween radar scans. In these cases, if the offset is assumedto be unchanged, the radar fails to lock onto the precedingvehicle. Therefore, because the position of the preceding

vehicle relative to the driver’s own vehicle is influenced bythe sideslip of vehicle, the offset needs to be corrected ateach radar scan based on the most recent detected corneringradius.

For the following operation, ACC performs control inorder to maintain a constant distance. However, for example,when the preceding vehicle is traveling close to the secondvehicle ahead, and if the preceding vehicle brakes andaccelerates suddenly and repeatedly (jerky motions), thedriver’s own vehicle is forced to perform the same jerkymotions. As a result, such a system cannot be expected toprovide good ride quality and a feeling of safety. Therefore, itis necessary to change from velocity control which maintainsa constant distance between the driver’s vehicle and thepreceding vehicle to velocity control which maintains aconstant distance between the driver’s vehicle and the secondvehicle ahead. In this way, even if the preceding vehicleperforms repeated jerky motions, the driver’s own vehicle isnot forced to perform sudden braking and acceleration. Thisresults in a definite improvement to the ride quality and thefeeling of safety.

From the above points of view, this paper proposes amore accurate method of detecting the preceding vehicle byradar, and of controlling the distance between vehicles. Thisproposal also for the first time includes consideration of thevehicle sideslip angle. These methods are expected to resultin improved driving stability.

Section 2 of this paper describes the ACC systemconfiguration. Section 3 explains the functions related tovelocity control. Evaluation for the ACC system based onthe experiment results and conclusions are presented inSection 4.

2. ACC System Configuration

2.1. Primary ACC Modes. The ACC system supports fourcontrol modes, which are described below and shown inFigure 1.

(1) Constant velocity control: when there are no vehiclesstraight ahead, or when there is a large distancebetween the driver’s vehicle and the preceding vehi-cle, the system maintains a constant vehicle velocity.

(2) Deceleration control: when a vehicle traveling ahead ata slower speed is detected, the system uses the throttleto decelerate the driver’s vehicle. If this decelerationis insufficient, the system uses the brake to deceleratethe vehicle.

(3) Following control: when the driver’s vehicle is fol-lowing behind the preceding vehicle, the systemcontrols the throttle and brake so that the timeinterval between the vehicles (which corresponds to adistance between the vehicles that is proportional tothe velocity of the driver’s vehicle) is the time whichwas set by the driver.

(4) Acceleration control: when, due to a lane change, thereis no longer a vehicle ahead of the driver’s vehicle, thesystem accelerates the vehicle up to the velocity set


60 km/h↓

↓

↓

↓

80 km/h

80 km/h

60 km/h

60 km/h 60 km/h

50 km/h

Set velocity

80 km/h

60 km/h

50 km/h

(1)

Con

stan

tve

loci

ty

oper

atio

n(2

)D

ecel

erat

ion

con

trol

(3)

Follo

win

gop

erat

ion

(4)

Acc

eler

atio

nco

ntr

ol

Figure 1: Four modes supported by the ACC system.

by the driver, and then maintains a constant velocity.Furthermore, when the driver’s vehicle approaches avehicle ahead of it without slowing down enough, analert buzzer and display prompt the driver to applythe brakes or take other appropriate action.

2.2. System Configuration. Figure 2 shows the ACC systemconfiguration and unit layout in a vehicle. This configurationincludes the existing Auto Speed Control (ASC), DynamicStability Control (DSC), Millimeter Wave Radar used todetect objects ahead, ACC Electrical Control Unit (ACCECU) used to calculate control values for the engine andbrake, Distance Setting Switch used to set a time intervalfor following the preceding vehicle (in order to maintain aconstant distance), and Indicator/Display which informs thedriver of the control mode. Figure 3 shows the part layout forthe sensors, ECU, and other components. For brake control,the DSC brake actuator function is extended to performfeedback control, so that the deceleration demanded by thedriver is delivered. For engine control, the vehicle velocity setby ACC is transmitted to the vehicle velocity control section

Millimeter wave radar

Distance setting switch

Brake lamp

Cruise control switch

Electronic throttle valve

Wheel speed sensor

Brake pressure sensor

Brake actuator

Steering angle sensor

Indicator / display Meter

BCM ECU

DSC ECU

PCM ECU

DSC

ASC

CANCAN

Yaw rate/lateral G sensor

ACC ECU

Figure 2: System configuration.

ACC ECU

Brake pressuresensor

Wheel speedsensor

Steeringangle sensor

Yaw rate/lateral G sensorDSC

Brakelamp

PCM Cruise control switch

Distance setting switchMillimeterwave radar

Speedmeter

BCM

Figure 3: ACC unit layout.

of the existing ASC, and the system then controls the vehiclevelocity so as to follow the vehicle preceding it.

Values from the wheel speed sensor and yaw rate/lateralG sensor are transmitted to the DSC ECU, and values forthe vehicle velocity and other items set from the operationswitches are transmitted to the PCM ECU. Furthermore,the signals of both ECUs are collected at the ACC ECU viaCAN (Control Area Network). The transmissions betweenthe Millimeter Wave Radar and the ACC ECU are carriedout via CAN. In addition, when brake control is performed,ACC controls the relay between the brake pedal switch andthe brake lamp in order to turn on the brake lamps.

As described above, rather than developing a new actu-ator control for the engine and brake, instead the functionsof the existing system are extended, ensuring that this systemcan be efficiently developed.


Figure 4: View of the Millimeter Wave Radar.

Table 1: Millimeter Wave Radar performance.

Detection rangeMax 150 m

Min 2 m

Resolution 0.1 m

Range rateMax 200 km/h

Min −200 km/h

Resolution 0.36 km/h

Azimuth angleArea ±7.5 deg

Resolution 0.1 deg

Data rate 100 msec

2.3. Millimeter Wave Radar. Figure 4 shows an external viewof the Millimeter Wave Radar used in this system. Thefrequency of the Millimeter Wave Radar is 76 to 77 GHz.An FMCW system is used, allowing the distance betweenvehicles and the relative velocity to be simultaneouslymeasured with very high precision. In order to detect thehorizontal angle, a mechanical scan system is used, in whichthe antenna and millimeter wave transmitter/receiver arepanned back and forth to the right and left sides by amotor. This system yields a relatively high angle resolution.In addition, the thickness of this radar has been reducedto 70 mm, which makes it possible to install it inside thebumper, despite the mechanical scanning. The specificationsof the millimeter wave radar are shown in Table 1.

3. ACC ECU

The main functions of the ACC system are selection of thepreceding vehicle and control of the distance between thevehicles. Figure 5 shows a block diagram of the control logic.

The ACC ECU selects the preceding vehicle to followby utilizing the information (distance and relative velocity)transmitted by the Millimeter Wave Radar. The system thencontrols the acceleration and deceleration of the vehiclebased on control values such as the target vehicle velocity andthe target acceleration/deceleration.

3.1. Selection of the Preceding Vehicle during Cornering. TheMillimeter Wave Radar is used to detect the vehicles travelingahead of the driver’s vehicle. When the road is straight,the preceding vehicle can be easily identified as the vehicle

traveling ahead on the same path as the driver’s vehicle.However, it becomes more difficult to identify the precedingvehicle when there are curves in the road. As shown inFigure 6, when three vehicles are traveling ahead of thedriver’s vehicle at a curve in the road, it is first necessaryto determine which vehicle is on same path as the driver’svehicle and which vehicles are not on the same path. Thenthe vehicle to be followed can be correctly identified.

The system judges whether the object detected by theradar is a relative static object or moving object by comparingthe velocities of the driver’s vehicle and detected vehicle.If the preceding vehicle is traveling at a speed that isapproximately the same, that vehicle is considered to be arelative static object.

At the same time, the path (with radius of curvature R) isestimated by determining the cornering radius of the driver’svehicle based on the detected yaw rate, steering wheel angle,and vehicle velocity.

The system judges whether or not the current detectedvehicle is the same as the previous detected vehicle eachtime a radar scan occurs. This judgment is performed bycomparing the position of the detected vehicle with theestimated position. Therefore the offset from the center lineof the path is recalculated at each radar scan in order toestimate a new position for the detected vehicle after it hasmoved.

Usually, the radius of curvature R of the path center lineis assumed to be unchanged during the period between radarscans. However, in cases when the vehicle is on a curve nearan entrance or exit, the path radius of curvature actually doeschange between radar scans. If the sideslip is assumed to beunchanged, the system fails to lock onto the target vehicle.Therefore, because the relative position of the precedingvehicle to the driver’s vehicle is influenced by the sideslip ofthe vehicle, the offset needs to be corrected at each radar scanbased on the most recent detected cornering radius.

3.2. Basic Logic. The system estimates the radius of curvaturefor the path of the driver’s vehicle based on the yaw rateand vehicle velocity. When the vehicle is cornering, therelationship among the velocity V , the yaw rate r, and thecornering radius R is obtained as follows:

R = V

r. (1)

Figure 7 shows the geometric relationships between thedriver’s vehicle and the vehicle immediately preceding it. Thearc of radius R is the traffic lane of the driver’s vehicle. If theoffset ε, which is the distance between the traffic lane andthe preceding vehicle, is within a certain range, the precedingvehicle is determined to be traveling on the same path. Here,if the distance to the vehicle immediately preceding is d,and the irradiation angle of the radar is θ, the followingrelationship can be derived:

{R− (ε − dθ)}2 = R2 − d2. (2)


dr, vrVt

vr

vL

Targetheadwaydistance

generator

Selectionleadingvehicle

Targetacceleration/deceleration

generator

Vehicle model PlantFeedback

compensator

Feedbackcompensator

Vr

VrdL

dt

+ − GcGt

Gt

Gr

Vc

PCMDSC

ACC C/M

Figure 5: Control block diagram.

FOV of radar

Predicted lane

Mazda MPV

Detected vehicleOther objectsout side of road

- Detect several objects by radar

- Predict own lane by Yaw ratesensor and vehicle velocity.

- Decide target vehicle based onthe Predicted lane and positionmoving object

Obstacle detection and path estimation

Leading vehicle

Figure 6: Detection of preceding vehicle and estimation of path.

Assuming that (ε− dθ) is sufficiently smaller than the radiusR, the offset ε can be derived as follows:

ε = dθ +d2

2R= d

(θ +

d

2R

). (3)

3.3. Introduction of the Sideslip Angle. In actual cases when avehicle is traveling at a certain velocity on the road, a sideslipangle occurs at the vehicle during cornering. When the trafficlane curves, this sideslip result in a deviation in the rangeof irradiation. Correcting the range of irradiation can beexpected to improve the offset accuracy.

A dynamic model of the vehicle shown in Figure 8 can beexpressed as follows [12]:

2(Kf + Kr

)β +

[mV +

2V

(l f K f − lrKr

)]r = 2Kf δ,

2(l f K f − lrKr

)β +

2V

(l2f K f + l2r Kr

)= 2l f K f δ.

(4)

R

R

θ

ε

dθ

d

Figure 7: Position of preceding vehicle during cornering withoutconsidering the sideslip angle.

Using (4), the sideslip angle can be calculated as follows:

β =(

1− m

2l

l f

lrKrV 2

)lrRl. (5)


Rear tirer: Yaw rateFront tire

C.G.

δ f β

lrl f

Figure 8: Two-wheel model of a vehicle.

Table 2: Vehicle system parameters.

m Vehicle mass

V Vehicle speed

C.G. Center of gravity

l f Distance between front axle and C.G.

lr Distance between rear axle and C.G.

Kf Front cornering power

Kr Rear cornering power

r Yaw rate

β Sideslip angle

δ f Tire angle

From the relationship between the sideslip and thecornering radius shown in Figure 9, the amount of centertravel L can be obtained as follows:

L = Rβ =(

1− m

2l

l f

lrKrV 2

)lrl. (6)

Based on the relationships shown in Figure 9 and consideringthe sideslip element, the following relationship can beexpressed in nearly the same way as in Figure 7:

{R− (ε − d

(θ + β

))}2 = R2 − d2. (7)

Assuming that (ε − d(θ + β)) is sufficiently smaller thanthe radius R, the offset ε can be obtained as follows:

ε = d(θ + β

)+

d2

2R= d

(θ +

d + 2L2R

).

(8)

The absolute amount of center travel L is proportional tothe square of the vehicle velocity as shown in (6). Therefore,here L can be expressed as follows:

L = aV 2 + b. (9)

Parameters a and b are then identified by using the actualexperiment data. Figure 10 shows the relationship betweenthe vehicle velocity V and the amount of center travel L.

Figure 11 shows the relationship between the amount ofcorrection dβ and the cornering radius R. It can be seen that

L

d(θ + β)

R

R

ε

d

θ

β

β

Figure 9: Position of preceding vehicle during cornering, withconsideration for the sideslip angle.

−2

−1

0

1

2

3

4

0 20 40 60 80 100

Vehicle velocity (km/h)

Cen

ter

trav

elL

(m)

Figure 10: Relationship between amount of center travel L and thevehicle velocity.

the smaller the cornering radius is, the larger the amount ofcorrection becomes.

The system judges whether the objects detected by themillimeter wave radar are relative static objects, whichinclude a preceding vehicle traveling at the same speed, orrelative moving objects. This judgment is performed by usingthe velocity of the objects relative to the velocity of thedriver’s vehicle. The driving path for the driver’s vehicle isestimated from the current vehicle conditions such as thevehicle velocity, steering angle, and yaw rate. The targetpreceding vehicle on the estimated path is selected based onthe position of the preceding vehicle relative to the estimateddriving path.

In Figure 12, the green lines indicate the estimated path.Despite the fact that there are many reflectors at the side ofroad and a vehicle traveling in the adjacent lane, the systemhas reliably locked onto the preceding vehicle. Figure 13shows an actual scene in which the preceding vehicle can beselected quickly by using the estimated path and the positionof the target vehicle.

3.4. Control of the Distance between Vehicles. The logicdescribed in Section 3.3 makes faster lock-on times possible,allowing the ACC ECU to smoothly control the velocity of


0

0.02

0.04

0.06

0.08

0.10

0.12

0.14

0.16

0.18

100 300 500 700 900

Cor

rect

ion

dβ(m

)

Cornering radius R (m)

Figure 11: Relationship between amount of correction dβ andcornering radius R (vehicle velocity: 70–75 km/h).

Preceding vehicle

(m)

(m)

Estimated path

Vehicle in adjacent lane

Scanning area

Reflectors

−20 −15 −10 −5 0 5 10 15 2020

40

60

80

100

Figure 12: Estimation of traffic lane and recognition of thepreceding vehicle.

the driver’s vehicle in order to maintain a constant distancebetween the vehicles. Furthermore, it allows the radar tomaintain a continuous lock on the preceding vehicle.

The target vehicle velocity and target acceleration/decel-eration are calculated using the time interval which wasset by the driver, the current distance between the driver’svehicle and the preceding vehicle, and the current relativevelocity of the vehicles. Based on the difference betweenthe target velocity and the current velocity, the velocitycan be controlled so that the distance between vehiclesgently converges on the target distance. Acceleration anddeceleration are performed in the same way as when thevehicle is operated by the driver. As a result, the driver doesnot experience any discomfort. Figure 14 shows the changesin the relative velocity when the driver’s vehicle is travelingat a speed of 85 km/h and approaches a preceding vehicletraveling at a constant speed of 60 km/h. It can be seen thatthe control of acceleration and deceleration by the ACC ECUis almost the same as when the vehicle is operated by thedriver. In fact, the ACC ECU control can been seen to besmoother than driver control.

View from the millimeter wave radar

Predicted path

Leadingvehicle

Vehicle

Frontal view and detected data overlay

The leading vehicle can be selected quickly using path andposition of the moving objects.

Figure 13: Actual case in which the preceding vehicle can beselected quickly by using the path and the position of the targetvehicle.

−30

−22

−14

−6

2

10

−16 −8 0 8 16 24 32 40 48

Target headway distance deviation (m)

Rel

ativ

eve

loci

ty(k

m/h

)

by ACCby driver

Figure 14: Results of driving test.

4. Conclusions

Improving the preceding vehicle lock-on performance byimproving the Millimeter Wave Radar unit and making fulluse of object identification logic and path estimation logicresulted in improved driving stability. This performance wasachieved even on an expressway with continuous sharp turnsin a mountainous region, or an expressway with multiplelanes and heavy traffic in an urban area. The precedingvehicle identification performance was approximately thesame as or better than the performance of other such systemsin the industry.

For the acceleration/deceleration performance, whichhas a large effect on occupant comfort, when the lane waschanged during tracking and the preceding vehicle sped up,the acceleration performance satisfied the need for a smoothfeeling of acceleration with a short response delay.

Conversely, when the driver’s vehicle caught up withthe preceding vehicle, a smooth and comfortable feeling ofdeceleration was achieved.


This Adaptive Cruise Control system was developed forthe purposes of driving safety and comfort. It reduces thenumber of brake and switch operations that are requiredof the driver. As a result, the system reduces the drivingburden so that the driver can drive in comfort. The systemdemonstrated sufficient lock-on, tracking, and accelera-tion/deceleration performance, and the system was able toprovide a satisfactory driving experience for the driver.

References

[1] Y. Yamamoto, T. Terano, and T. Nakagami, “Development ofMazda radar cruise control system,” Tech. Rep. 24, MAZDA,2006.

[2] Y. Yamamura, Y. Seto, H. Nishira, and T. Kawabe, “An ACCdesign method for achieving both string stability and ridecomfort,” Journal of System Design and Dynamics, vol. 2, no.4, pp. 979–990, 2008.

[3] H. Soma, Y. Shiraishi, T. Watanabe, Y. Takada, and Y. Takae,“Trust in low-speed adaptive cruise control systems—analysisof trust structure,” Review of Automotive Engineering, vol. 26,no. 2, pp. 211–212, 2005.

[4] Y.-S. Kim and K.-S. Hong, “An IMM algorithm for trackingmaneuvering vehicles in an adaptive cruise control envi-ronment,” International Journal of Control, Automation andSystems, vol. 2, no. 3, pp. 310–318, 2004.

[5] H. Fukuoka, Y. Shirai, and K. Kihei, “Driving support systemadaptive to the driver state of surrounding vehicles: simulationstudy on a rear-end precrash safety system,” Transactions of thesociety of Automotive Engineering of Japan, vol. 40, no. 3, pp.933–938, 2009.

[6] K. Fujita and S. Tokoro, “Pre-crash safety,” Journal of Societyof Automotive Engineerings of Japan, vol. 59, no. 12, pp. 85–90,2005.

[7] Z. Sun, G. Bebis, and R. Miller, “Monocular precrash vehicledetection: features and classifiers,” IEEE Transactions on ImageProcessing, vol. 15, no. 7, pp. 2019–2034, 2006.

[8] K. Adachi, “Proposal of a target headway distance methodand applying this method to a car for adaptive cruisecontrol system,” Proceedings of the Japan Society of MechanicalEngineers, no. 06-52, pp. 23–28, 2006.

[9] N. Shimomura, “A study on preceding vehicle tracking oncurved roads using stereo vision,” Tech. Rep. PRMU97-27,IEICE, 1997.

[10] T. Hiraoka, H. Kumamoto, and O. Nishihara, “Sideslip angleestimation and active front steering system based on lateralacceleration data at centers of percussion with respect tofront/rear wheels,” JSAE Review, vol. 25, no. 1, pp. 37–42, 2004.

[11] X. Gao, Z. Yu, J. Neubeck, and J. Wiedemann, “Sideslip angleestimation based on input-output linearization with tire-roadfriction adaptation,” Journal of JSAE, vol. 48, no. 2, pp. 217–234, 2010.

[12] B.-C. Chen and F.-C. Hsieh, “Sideslip angle estimation usingextended Kalman filter,” Vehicle System Dynamics, vol. 46, no.1, pp. 353–364, 2008.

[13] H. Okazaki, H. Omura, T. Seto, and T. Nakagami, “Obstacledetection equipment of vehicle,” Patent Application 2006-106549(P2006-106549).

[14] J. Takizawa, K. Sakagami, M. Masuda, K. Yamada, and T.Kyuma, “A development of pre-crash safety system for a minisized vehicle,” Transactions of Society of Automotive Engineeringof Japan, vol. 39, no. 2, pp. 15–20, 2008.

[15] N. Shimomura, A. Nakamura, T. Goto, K. Fujimoto, andH. Shitsu, “A method of tracking a forward vehicle using ascanning laser radar and a camera,” Transactions of the Instituteof Electrical Engineers of Japan C, vol. 123, no. 8, pp. 1427–1438, 2003.

[16] H. Miyazaki, S. Arita, and W. Ishio, “Development of laserradar for automotive application,” Journal of Society of Auto-motive Engineering of Japan, vol. 60, no. 5, pp. 45–48, 2006.

[17] K. Osugi, K. Miyauchi, N. Furui, and H. Miyakoshi, “Devel-opment of the scanning laser radar for ACC system,” DENSOTechnical Review, vol. 6, no. 1, pp. 43–48, 2001.

[18] M. ABE, Motion and Control of Vehicle, Sankaido, Tokyo,Japan, 2nd edition, 2007.


Research Article

Reducing Congestion in Obstructed Highways with Traffic DataDissemination Using Ad hoc Vehicular Networks

Thomas D. Hewer,1, 2 Maziar Nekovee,1, 3 and Peter V. Coveney2

1 Department of Computer Science, UCL, Gower Street, London WC1E 6BT, UK2 Centre for Computational Science, UCL, 20 Gordon Street, London WC1H 0AJ, UK3 Mobility Research Group, BT Research, Polaris 134, Adastral Park, Martlesham IP5 3RE, UK

Correspondence should be addressed to Thomas D. Hewer, [email protected]

Received 1 December 2009; Revised 14 June 2010; Accepted 5 July 2010

Academic Editor: Shahrokh Valaee

Copyright © 2010 Thomas D. Hewer et al. This is an open access article distributed under the Creative Commons AttributionLicense, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properlycited.

Vehicle-to-vehicle communications can be used effectively for intelligent transport systems (ITSs) and location-aware services.The ability to disseminate information in an ad hoc fashion allows pertinent information to propagate faster through a network.In the realm of ITS, the ability to spread warning information faster and further is of great advantage to receivers. In this paperwe propose and present a message-dissemination procedure that uses vehicular wireless protocols to influence vehicular flow,reducing congestion in road networks. The computational experiments we present show how a car-following model and lane-change algorithm can be adapted to “react” to the reception of information. This model also illustrates the advantages of couplingtogether with vehicular flow modelling tools and network simulation tools.

1. Introduction

In the realm of vehicle-to-vehicle (V2V) communicationsthere are several methods for the dissemination of data thatare being actively researched. The use of satellite commu-nication such as those linked to global positioning services(GPS) offers global communication but requires expensiveequipment, large antennae (for two-ray transmission) and,due to the large distances the signal must travel, they willhave a high latency. Cellular telephone networks offer alower latency over large distances but are still slow whencommunicating with nearby vehicles, due to a centralisedapproach, and require cellular contracts to use the network.The scenarios presented in this paper require high-speedcommunication disseminated from source, which is difficultto achieve using either cellular or satellite communications.

Ad hoc networks offer a good method to spreadinformation outwards from an origin quickly and efficiently.It has been shown by Nekovee [1] that in ad hoc networksworms spread in an epidemic pattern that can be modelled.Using such modelling techniques we can develop algorithmsthat allow for a change to be made to the speed, position and

route of a vehicle. A further advantage of ad hoc networkingis the unlicensed use of the radio-spectrum and the recentreduction in cost for the equipment for communication.A separate strand of the research being undertaken onwireless fidelity (under IEEE standard 802.11) has beendeveloped in the past few years specifically for vehicularad hoc networks (VANETs). The 802.11p WAVE standardspecifies network protocols which address the difficultiesassociated with vehicular networks. These difficulties involveshort link times, delay tolerance, and the inefficiency ofwired-network paradigms that have been inherited into thewireless standards.

The following simulation experiments and algorithmswere developed with particular scenarios in mind. Thesescenarios operate in a dual-carriageway environment with nojunctions together with an obstacle or danger that is presentat some point in the field. The simulations use both vehicularflow modelling and message propagation to advise thevehicles of the obstacle at a greater distance than line-of-sightprovides. The results show that by spreading informationquickly and efficiently through the network we can developsystems to reduce congestion and limit negative vehicular


flow effects. By coupling the telecommunications and drivermodel in one tool we can perform these simulations in thesame runtime. This method of simulation can also be usefulfor intelligent transport systems, where simulations run inparallel can sweep the parameter space for a desired outcome.The impact of applications such as this is to reduce thetime spent in congested vehicular flow and also to increasethe safety interval and reduce collisions. Many possibleapplications of VANET technology are designed to improvesafety of life while driving, and in this work we show that theapplication of our novel algorithms will realise this goal.

The paper is organised as follows. Section 2 shows somerelated work in VANET and ITS applications. Section 3presents the simulation system we have used for this workand the models that have been incorporated. Section 4presents the algorithms for the intelligent driver modeland the lane-changing decision, and also the epidemicpropagation modelling for dissemination of information. Wethen present the simulation scenario in Section 5. Section 6shows the studies we have explored and the results onvehicular flow through the simulated system. The paper endswith conclusions in Section 7.

2. Related Work

There are many examples of work using vehicular networksimulation to test the applications of wireless technology.Yin et al. [2] use simulation to evaluate the performance ofdedicated short-range communication (DSRC) and Eichleret al. [3] analyse the impact of V2V messaging on vehicularflow. Torrent-Moreno [4] has produced a great deal ofwork on safety applications for vehicular communicationsnetworks, providing empirical comparison of results underdifferent network settings and providing mechanisms for thedelivery and dissemination of vital data.

There are tools available for performing coupled simu-lations of vehicular mobility and network communicationincluding VGrid [5] which has been designed for usemainly with V2I systems, and MobiReal [6] which producesscenario data for accurate mobile ad hoc network (MANET)simulations, similar to the work of Wellnitz et al. [7]. Thesesystems are useful but we found that the ability to takethe best from both aspects and combine them in a highlyconfigurable and extensible simulation system, was best metby the vehicular flow application of Treiber. The work ofChen et al. [8] is similar to ours for a different vehicular flowand accident scenario. We explore our data deeply so that wecan validate against this established work.

3. Simulation System

The simulation tool we use is adapted from the dynamicvehicular flow simulator by Treiber et al. [9]. This tooluses a simple model of a two-lane roadway, but containsan advanced driver model and lane-changing algorithm,MOBIL [10]. We have added telecommunications function-ality to the original simulator for V2V communication. Inthis section we show the underlying models that operatewithin the simulation system.

3.1. Vehicular Modelling. To accurately simulate vehicularbehaviour, there are several key components: a driver modelto develop how real people will drive under certain circum-stances, a lane-changing model to make realistic decisions onwhen it would be advantageous to change lane and a roadwaywith rules (e.g., drive on the left in the UK).

A car following algorithm will contain at least a desiredvelocity, a safe time separation when following other vehiclesand acceleration and braking criteria [9]. At each simulationtime step the acceleration is calculated for each vehicle. Theparameters of these models can be changed to emulate moreaggressive and more considerate drivers, as required.

When modelling vehicular networks over large areas(such as metropolitan areas) the flow of vehicles on a singleroad behaves as an incompressible fluid according to Q =ρV , where ρ is the average density of vehicles (cars/km)and V is the average velocity on the road (km/h) [11].At microscopic levels of simulation (across any field size)each vehicle is treated as an individual entity, which greatlyincreases the computational requirements of the system butprovides a more realistic and component-based approach tomodelling.

The Intelligent Driver Model (IDM) in the simulatorfollows the MOBIL model [10] developed by Treiber. MOBILoperates as a car-following model such that the accelerationand braking are defined by the distance from the car in front.The function of such an acceleration dv/dt is shown in

dv

dt= a

[1−

(v

v0

)δ−(s∗

s

)2]

, (1)

where

s∗ = s0 +(vT +

vΔv

2√ab

)(2)

for acceleration on an open road a, velocity v, desiredvelocity v0, distance s to front vehicle, desired dynamicdistance to front vehicle s∗, velocity difference Δv, a safetime delay between vehicles T , a comfortable braking valueb, a minimum distance between vehicles s0 and finally anexponent δ which is adjusted in order to mimic real vehicularflow patterns.

Lane changing algorithms add a necessary level ofcomplexity to the IDM. In order to decide whether to changelane or not, the current acceleration must be calculated forthe current lane and the acceleration in the new lane (withregard to the car behind and in front in the original lane).If the acceleration in the new lane is greater than that inthe current lane, there is an advantage to be gained bychanging lane. Many models, including those in the originalsimulator, include a bias in the model for particular lanes,which simulates the real scenario of slow lanes and fast lanes.

3.2. MAC Layer Protocol. In simulating wireless fidelitynetworks, the majority of simulations use the IEEE 802.11protocols [12], as these offer the best simulation of theMAC layer functionality. The IEEE 802.11 MAC layer usesa distributed coordination function (DCF), which has beensimulated in [13], to ensure efficient communication on the


medium, and implements controls to reduce collisions. Morerecently the IEEE 802.11p standard has been tested in [2]specifically for intervehicular communications. This allowsthe foundation of underlying strengths in the 802.11 suite tobe enhanced for vehicular networks.

The MAC layer in the simulator operates using anadapted version of IEEE 802.11 which removes the inter-frame spacing (IFS) model, enabling equal priority for allnetwork traffic. Due to implementation, and the need forsimplicity in the model, our implementation of 802.11 doesnot suspend the backoff timer when the medium is busyduring that frame, as 802.11 does.

The network backoff when the medium is busy Xoperates as shown in [11]

X ∈ 2n × [Bmin,Bmax], (3)

where n is the number of times it has previously had toback off in succession. Bmin and Bmax are the minimum andmaximum possible back off time, respectively. Bmin is oftenset at 0. The medium is defined as busy if any car within thetransmitters interference range, Ri, is currently broadcasting.

Every car within the transmission range, represented byRc, (which is usually twice as small as the interference range)will receive the message with probability λ.

3.3. Radiowave Propagation. In the modified simulator thereception of messages is performed by a basic algorithmcontrolled by the simulation engine. In advanced networksimulators the realistic reception of messages depends onthe signal strength at the receiver. The basic propagationmodel is the Friis free-space calculation, which extends theideal free-space propagation formula (Pr ∝ 1/d2) [14]where Pr is received power and d is the distance fromtransmitter) to incorporate the antenna gain(both transmitand receive). The Friis model, however, will only hold truewith a clear line-of-sight (LOS) between transmitter andreceiver, and assumes no level of multipath or shadowing ofthe radiowave (which becomes very apparent in urban andhighway scenarios [15]). Friis operates as shown in

Pr(d) = PtGtGrλ2

(4π)2d2L, (4)

where Pr is the received power at distance d with respect tothe antenna gain and height and the system loss L.

One method of altering the propagation of messagesthrough the simulated network, is to change the transmittedpower and therefore the transmission range. By transmittinginformation further the message has a greater probabilityof retransmission in sparse networks and also a faster dis-semination through the system [16]. The main disadvantageof having a large transmission range is that informationpropagates further, eventually reaching a point where theinformation usefulness is low, and so takes up bandwidthand time for almost redundant data. In a system with limitedtime and bandwidth this can cause localised problems, wheremore pertinent (i.e., geographically closer) information islost to redundant data.

4. Algorithms

This section examines the algorithms used in the simulation.These algorithms form the basis for the work we presenthere and have been designed specifically for vehicle-to-vehicle and VANET scenarios. We have added to the originalsimulator, by Treiber et al., including telecommunicationsfunctionality and adapting both the IDM and lane-changingmodel, based on reactions to information received from anobstacle or danger. From this, we are able to show thatmore efficient information dissemination through a networkcan increase vehicular flow and also reduce stop-and-goformations (presented in our results in Section 6).

4.1. Epidemic Message Passing Algorithm. The propagation ofmessages through a system requires an efficient delivery algo-rithm. In our simulation we use a probabilistic informationdissemination protocol, which is fully defined in Nekoveeand Bogason’s work [11]. We assume that all vehicles knowtheir location (via GPS technology) and that each messagecontains information about location and time of creation.To ensure that propagation does not extend to unnecessarydistances from the source each message has a Time-To-Live(TTL) setting.

The algorithm allows for a reasonable amount of retrans-mission and dissemination through the network and bal-ances the relevance of the information with the distance fromthe source. To this end, information can disseminate quicklyand efficiently and also reduce information spreading tovehicles which do not require the information (as discussedpreviously, this can cause more pertinent information to belost).

The probability for rebroadcasting, P, as described byNekovee [17] is obtained from

P =

⎧⎪⎪⎪⎪⎨⎪⎪⎪⎪⎩

1 if Nf or Nb = 0,

1− exp

⎛⎝−A

∣∣∣Nf −Nb

∣∣∣Nf + Nb

⎞⎠ otherwise,

(5)

whereNf andNb are the number of times the car has receivedthat particular message from front and back, respectively,and A is a protocol parameter which controls redundanttransmissions. In the case of directional message propaga-tion, (5) is modified such that if a message is propagating ineither direction it is only kept alive by nodes near the head ortail of the group. In this case the rebroadcasting probability,P, is computed from

P =

⎧⎪⎪⎪⎨⎪⎪⎪⎩

1 if Nk = 0,

1− exp

(−A Nk

Nk + N−→k

)otherwise,

(6)

where Nk is the count of messages received from the directionof message propagation (e.g., if k is forward, Nk is the countof messages received from vehicles in front), and N−→

kis the

count of messages received from the opposite direction.


4.2. Variable Speed Limit. The initial algorithm we introducereduces the value of the desired velocity (v0 in (5)) by a fixedamount when the vehicle has received the warning message.This achieves an overall slowdown in the network which canreduce the time delay between free-flow and gridlock (wherev = 0) at the obstacle. This particular algorithm change has atransient effect on the network, such that the system will stillbecome congested over time. The idea for this came from theLondon Orbital (M25) variable speed limit which operateson parts of the motorway.

The value we reduce v0 by is of great importance. Initialtests showed that reducing the desired velocity too much(i.e., a reduction of over 10 m/s) when infected, the effecton the network was to cause congestion further back in thesystem, such that gridlock (i.e., all vehicles in the field ofsimulation are static) occurs much sooner. By reducing thevalue of v0 by 2.7 m/s the network slowed well, and the timedelay between free-flow and gridlock was increased withoutcausing congestion further back in the network.

An important algorithmic change is the return to thenormal value of v0 once the obstacle has been passedgeographically, otherwise the recovery from the obstacle willtake a greater amount of time. We did test a proportionalchange in the desired velocity as the obstacle was approached,but this provided little observable effect at great distances anda highly-negative effect closer to the obstacle, as cars werecongesting more smoothly but to a greater extent.

The results of this algorithm change were both transientand often detrimental to the overall velocity of vehicles in thesystem, so the change was dropped from the final algorithmdescribed in Section 4.3. The reason for this negative effect isthought to be related to the size of field we are simulating.In future simulations we will explore a much larger field,where we expect the effect of the algorithm to become morepronounced as the distance from the obstacle increases.

4.3. Lane Changing Algorithm. The original lane changemodel, shown in (7), operates by determining an accelerationadvantage (toward a goal velocity) to be gained by changinglane and then testing if a gain threshold is reached bythe advantage. Early incarnations of our changes to thealgorithm worked to forcibly increase the advantage if amessage had been received and the vehicle was in thelane with the obstacle. This has some positive effects, butcan cause problems when the message propagates a largedistance back through the system. In the case of the messagepropagating beyond the reasonable extent of the need tochange lane, this approach causes unnecessary congestion inthe adjacent lane to the obstacle resulting in total congestionin a short time.

At the start of the simulation several static variablesare applied to the model. A changing threshold is appliedthat indicates the increased acceleration the lane change willyield; this is set by default at 0.3 ms−2 for cars in the field.The other value is the ‘politeness factor’ which reduces theoverall calculated advantage and which simulates the caredrivers take when changing lane (i.e., the model may say itis advantageous to change lane, but the driver may be toopolite or hesitant to do so).

The basic algorithm operates by calculating a value ofadvantage (A), the disadvantage this causes to other (B) andthen calculates whether a function of these values reaches achanging threshold. If the threshold is reached the vehiclechanges lane:

A = an − ao + B, (7)

where an is the acceleration in the new lane and ao is theacceleration in the old lane. B refers to a weighting to keepthe vehicle in the slow lane, as operates in reality:

B = ab(o) − ab(n), (8)

where ab(o) is the acceleration of the car behind in the old laneif the vehicles changes lane and ab(n) is the acceleration of thecar behind in the new lane if I change lane. These values arethen entered into the following equation to return true orfalse to changing lane:

(A− p

)∗ B > T , (9)

where p is the politeness factor and T is the changingthreshold. The form of (9) is multiplicative so the valuesof A and B have a significant impact on each other. In thefollowing equations, these values are only calculated if thevehicle has been infected with a message. If a vehicle isignorant, it will continue to use the algorithm in (9). Theinitial change we made was to add a value to A in (10) assuch

((A + V)− p

)∗ B > T. (10)

This is very much a brute force approach and as such doesnot truly represent real driving in a highway, where the valueof V would increase as the vehicle approaches the obstacleand drop to zero after the obstacle has been passed. Thisproportional addition to A is shown in

τ = Po(Pm − Po)

,

((A + τ)− p

)∗ B > T.

(11)

In this adaption of the original algorithm, the value τ iscalculated as a function of the location of the obstacle andthe vehicle’s distance to it. This value is capped at a maximum(currently 20) to prevent unrealistic behaviour (i.e., cuttingin with zero safety headway), which means the effect isnoticeable but quite subtle, when compared to the brute forcemethod in (10). This means that as the vehicle approachesthe obstacle the incentive to change lane becomes greater,reducing the appearance of congestion at the obstacle in thesame lane.

5. Simulation Scenario

In order to deliver usable and realistic results, the simulatedsystem must be an accurate representation of a real highway.The models for vehicular flow and network transmissionhave been introduced in previous sections, and the metrics


Time 0:49#S : 29#R : 136#I : 24

Main inflowTruck percentage

Imposed speed limitTransmission range

Time warp factor

2350 vehicles/h

120

0%120 km/h100 m5 times

02550(%

) 75100

Message spreading

Time

15 s

IgonrantInfectedActiveBacking off

SleepingStifler

Figure 1: A graphic showing the simulator, with the obstacle in the lower quadrant of the curve to the left.

gathered with the results obtained are shown in Section 6.The system we used for our study is shown in Figure 1. Theroad length is approx. 1400 m and has two lanes with vehiclesmoving in the same direction and an obstacle at 720 m, inthe left-hand lane. Vehicles originate in the top-right of theroad and travel around to the departure point in the bottom-right. The flow of vehicles entering the road can be variedfrom 0 to 4000 vehicles per hour, split across the two lanes.The desired speed (v0 in (1)) can be varied between 20 and140 km/h, which will directly affect the vehicular density asgreater velocity will require greater separation, according tothe IDM.

The transmission of data between vehicles is controlledby them operating in five distinct states: ignorant, wherethe vehicle has received no data and is not transmitting,informed, when the vehicle receives data from the network,transmitting, to resend received data through the network,backing off, when the vehicle wishes to transmit but themedium is busy, sleeping, when the vehicle has forwardeddata but is waiting to make sure all neighbours have receivedthat data and stifling, when the vehicle has performed all thereception/transmission required and therefore is no longeractive. The movement through these states, from ignorantto stifler, can be seen as a loop. The propagation methodsdescribed in Section 6.3 control how the vehicle progressesthrough the states in different ways. The transmission rangeof the vehicles is also set globally for the simulation, between0 and 200 m (it is possible to extend the transmission range ofvehicular networks beyond 200 m but only by using very hightransmission power and assuming little interference). Thevehicle that is causing the obstacle (surrounded by severalcar-lengths of safety barriers) constantly transmits withoutmoving to the sleeping or stifler states, so that any vehiclecoming within range will receive the data, and then begin theprocess of receive-and-forward.

In the next section we present the findings of our work.

6. Simulation Studies

In this section the collection of metrics is discussed withdiagrams and presentation of the results. All the simulations

were run with varying settings, so that an appreciation ofall the situations that may occur (that we can control inthe simulator) can be established. It is important to note, aspreviously mentioned, that in the initial studies we assumedthat all vehicles in the simulation are equipped with thetechnology for message propagation. In reality this marketpenetration will take many years to achieve, but many carmanufacturers are working on supplying this technologysoon [18].

In later studies, as shown in Section 6.5, we have exploreddifferent rates of equipped vehicles in the simulation. Thisrequires that all parameters remain static whilst testingthe equipped vehicle rates, such that we can rule out anyinfluence from transmission range, vehicular density andvehicle velocity. For each run at different parameters wetested equipped vehicle rates from 0% to 100% in incrementsof 10% and performed 20 runs to average out the data andremove any transient or artifactual data.

6.1. Velocity Experiments. These experiments test the effec-tiveness of the algorithm as vehicle velocity changes, to seeif the algorithms are suited to an urban (slow) or highway(fast) environment. When ignorant (no messages have beenreceived), the cars will still attempt to change lane to avoidthe obstacle, but only as part of the original lane-changingalgorithm, and so congestion builds up in a short amountof time, for most simulations. Below a certain network loadthe road will never become congested, so the vehicular loadof the experiment was varied for each experiment. Thevehicular load on the road was also set low enough so thatthe algorithm could affect the flow of cars as, at high loads,this would not be possible. To this end there is, in any system,a critical value of vehicular load after which no action canprevent or reduce congestion.

Some early simulations with low vehicular loads showedthat it is sometimes more efficient to be ignorant of theobstacle, and this must be taken into account, as in thiscase the best course of action is to drive normally, using thenormal algorithm.

Figures 2 and 3 show results from a sample of theexperiments we ran to test this theory. By varying all


00 100 200

InfectedIgnorant

300 400

Time

500 600 700 800 900 1000

20

40

Nu

mbe

rof

cars

exit

ing

60

80

100

120

140

180

160

Low velocity limit (50 km/h)-load 4400 cars/hour

Figure 2: Comparison of exit aggregate for infected and ignorantsimulations at urban velocities (below 50 km/h).

00 100 200

InfectedIgnorant

300 400

Time

500 600 700 800 900 1000

50

Nu

mbe

rof

cars

exit

ing

100

150

200

250

300High velocity limit (120 km/h)-load 4400 cars/hour

Figure 3: Comparison of exit aggregate for infected and ignorantsimulations for motorway velocities.

the parameters available we found that certain velocities,transmission ranges, and vehicular load had different levelsof effectiveness to the overall congestion in the system. In themain the results showed that our algorithm always produceda positive effect. In order to produce valid data we ran severalsimulations with the same parameters and then averaged thisdata, to show results for a typical case.

Figures 2 and 3 show the number of cars exiting the fieldin a simulation run as an aggregate over time. Both show anadvantage for infected cars using the advanced lane-changingalgorithm, but the advantage is greater at higher velocities,where the vehicles have more distance between them for thesame vehicular load, meaning they can more easily changelane.

We found, and Figures 2 and 3 corroborate this, thatthe advanced algorithm increased the time before congestion

00 100

Communication onCommunication off

Accident position

200 300

Time

400 500 600 700

100

Posi

tion

ofla

ne

chan

ge

200

300

400

500

600

700

800Position of lane change over time

Figure 4: Chart showing location on the field of lane changes withand without communication.

began to build up and then once congested, the infectedcars still moved through the system more efficiently. Byrunning the simulation for 15 minutes we can monitor thedevelopment of congestion in the system and how the flow isaffected by the changed algorithm.

The results show interesting behaviour, beyond thereduction of congestion in the system. By analysing thefirst 5–7 minutes of simulation time, we can see that thedevelopment of congestion is also slower once vehicles dostart to slow down. This is because of the algorithm movingvehicles into the opposite lane to the obstacle, reducing theload on the lane with the obstacle and therefore reducing thenumber of stopped vehicles behind the obstacle which, whenchanging lane, cause a dramatic slowdown in the new lane.This reduction in stop-and-go vehicular formation is alsoseen elsewhere in the field when the cars are infected withthe warning message and switch to our adapted algorithm.

6.2. Position of Lane Change. A factor that affects the build-up of congestion in the system is related to the locationof the lane change. The following results show where thelane change occurs with no communication and then usingour enhanced lane change algorithm with communicationactive. The simulation settings were set at 2200 cars/hourload, speed limit of 120 km/h and a transmission range of100 m.

During the simulation the message propagates backwardstowards position 0 and, with the advanced algorithm, thelocation of the lane-change also reduces. When the systemstarts to slow and vehicular density increases, the lane-changemoves right back, causing a slower build-up of congestionand a greater amount of free vehicular flow, as shown byFigure 4. This does place greater load on the opposite laneto the obstacle, but the reduction of stop-and-go behaviournegates this. In Figure 4 the rapid reduction in position ofmost lane changes between 395–405 seconds and again at


00 100

Propagation off

MixedEdge detectionDistance detection

Simple flooding

200 300

Time (s)

400 500 600 700 800

10Flow

profi

le(c

ars

out/

cars

in)

(%)

20

30

40

50

60

70

80

90Flow profile comparing propagation methods

Figure 5: Comparison of message propagation methodologies.

475–500 seconds represents a period when the propagationof the message is continuous, and the periods of little change(of lane change position) are due to reduced propagation ofthe warning message. The initial peak of lane change positionbetween 0 and 35 seconds represents the initialisation of thesystem, that cars can change lane very close to the obstacledue to the road being less loaded. The data to produce thisresult came from a single typical simulation, with parametersset as per the previous paragraph.

6.3. Transmission Method Experiments. In this experimentthe available transmission methods are tested with the newalgorithm, to see how they affect the overall congestion inthe system. Figure 5 shows a simulation run until congestionis present at the origin of the field (i.e., position = 0).The simulation is stopped when the congestion reaches theorigin as after this point the algorithm is not affectingthe vehicular flow. The values represent the proportionof cars leaving the field in relation to the cars entering(Cars Exiting/Cars Arriving). This indicates how vehicles areflowing through the system, where an increase of the gradientrepresents free flow and a decrease represents congestion.

The models shown in Figure 5 are simple flooding, wherethe message is rebroadcast just once, edge detection whichis explained in Section 3.1, and distance detection, whichis a different probabilistic method and a mixture of edgeand distance detection. These models are all running inthe simulation, but the mixed (edge detection and distancedetection) offers the best simulation of a real epidemicprotocol.

As can be seen in Figure 5 the edge detection methodalone offers little improvement over no propagation, andthe simple flooding and distance detection methods offer agood initial advantage (0–200 seconds) but then suffer veryfast congestion build-up. The mixture of edge and distancedetection algorithm, with our changes to the lane changemodel offers excellent results keeping near free flow until

0 200400 600

TimePosition800

10001200

1400 0100

200300

400500

600

0

Ave

rage

velo

city

102030405060

Average velocity map with no transmission

Figure 6: Average velocities across the simulated highway withouttransmission.

0 200400 600

TimePosition800

10001200

1400 0100

200300

400500

600

0

Ave

rage

velo

city

102030405060

Average velocity map with transmission active

Figure 7: Average velocities across the simulated highway withactive transmission

ca. 420 seconds, when the network then slows and starts tocongest, but this takes longer (ca. 250 seconds from the firstslowdown) than the other algorithms.

The addition of this propagation method and thechanged algorithm prevent several congestion-causing situ-ations to occur. The main situation avoided is when vehiclesare unable to change lane to avoid the obstacle and begin toslow down, but then do change lane causing the cars behindto slow. This cause has been seen to initiate the build-upof congestion by earlier warning of the obstacle so that thecars can change lane at a high velocity. Another behaviourof the system is that when the congestion is initiated, thecars will fill up behind the obstacle, unable to change lane.With the adapted algorithm the extra incentive to changelane means that the opposite lane fills first and so vehiclescan still move, increasing the time before the whole systembecomes congested.

6.4. System Velocity. The average velocity through a systemis of great importance. If a higher average velocity can beachieved the number of vehicles passing through the area ofcongestion will be higher than if there is much slowing ofvehicles. Figures 6 and 7 show the average velocity calculatedfor intervals of 10 metres on the x axis and an interval of 30seconds across the y axis. Each point represents the averagevelocity at that time/position interval. In both figures there


00 50 100

Time

150 200 250 300 350

200

400

Post

itio

n

600

800

1000

1400

1200

40% equipped vehicles

(a)

00 50 100

Time

150 200 250 300 350

200

400

Post

itio

n

600

800

1000

1400

1200


(b)

00 50 100

Time

150 200 250 300 350

200

400

Post

itio

n

600

800

1000

1400

1200


(c)

00 50 100

Time

150 200 250 300 350

200

400

Post

itio

n

600

800

1000

1400

1200


(d)

Figure 8: Vehicle trajectories through the system for (a) 40% equipped vehicles, where we see little change from 0% equipped vehicles, (b)60% where some improvement is shown but congestion still builds up, (c) 70% where the improvement is noticeable and (d) 80% wherethe trajectory is only slighty affected by the obstacle. Each coloured trajectory represents a single vehicle as it enters and leaves the simulatedhighway.

is a noticeable slowdown as the vehicles pass the obstacle.This can be accounted for by the IDM attempting to retaina minimum safe distance between vehicles.

Figure 6 shows that after an initial even velocity throughthe field, congestion begins to build up at the position ofthe obstacle between 100 and 200 seconds, which causes aslowdown further back to position 0. By time of 330 secondsthe congestion has reached position 0, and the averagevelocity falls from 20–25 ms−1 to 0–5 ms−1.

As can be seen in Figure 7, there is a uniform averagevelocity before and after the obstacle during the whole periodof the simulation (10 minutes). This reinforces the resultsfrom the other simulations and proves that there is a betterflow of vehicles through the network, as well as a reductionin the build-up of congestion, when effective transmission ofthe road condition occurs.

We note that for both figures there is a spike in velocities(between time 0–10 and position 1200–1400). This is theperiod when the first cars are leaving the field. As they have

no cars in front they can accelerate up to the full speedlimit unhindered. To remove this artefact future simulationswill have a “warm-up” period, with low vehicular load, thatinitialises the field.

6.5. Varying Penetration Rate of Radio-Equipped Vehicles.In all of the simulation studies and results presented inSection 5, we have assumed that 100% of the vehicles inthe simulated field are radio equipped and therefore ableto receive and retransmit (where necessary) the data aboutthe upcoming incident or congestion. This is a somewhatidealised view of how deep uptake of radio-equipped vehicleswill be. More likely, as vehicle manufacturers and radioequipment providers begin releasing the products to operateand support a VANET, the ratio of equipped vehicles willincrease. Due to the critical safety-of-life applications ofVANET, it is important to study the efficiency and resilienceof simulated applications, to study the point at which theybecome usable and reliable. This simulation study involved


the same system as in previous experiments, but here theproportion of equipped vehicles is increased from 10% to90%. The plots in Figure 8 show the flow trajectories ofvehicles in the system.

We ran each simulation several times, to ensure noartifactual or erroneous data was included in our results,but the results shown here represent only one simulation. Ascan be seen at low penetration rates (10%–40%) the vehiclesreach the obstruction and begin to congest, but between 60%and 70% there is a noticeable improvement, or smoothing, ofthe vehicles velocity through the field of study. The figuresshow that above 70%, where there is still some conges-tion build-up around the obstruction, the propagation ofinformation to vehicles further away, and the change to thelane-changing algorithm this triggers, allows for more steadydriving through the field of study, and less congestion.

7. Conclusions

This paper has presented a simulation of a specific roadcondition, that of an accident blocking a lane on a dualcarriageway. The simulation uses the coupled approach tomobility modelling and network simulation in a singleprocess. The models and algorithms that represent thevehicular flow and network traffic have been implementedaccording to well-known and standard models. We haveshown that adapting the algorithms when information aboutthe accident is received via wireless transmission, we canreduce the build-up of congestion and increase the flow ofvehicles passing the obstacle, both in terms of quantity of carsand average velocity.

The various algorithms we tested achieved overallimprovement in the majority of cases. In some simulationsthe improvement was not only in the prevention of con-gestion overall, but by also keeping a consistent averagevelocity through the network, which helps to reduce theeffects of stop-and-go vehicular flow and smooth possiblecongestion “waves” that emanate from the source backwards.The driver model and lane-changing algorithms come fromwell validated sources, and so the adaptions we have madeare highly realistic and can show the effect of even simplechanges (as in the brute force addition of a value to A).In order to fully test our algorithms we ran numerous testswith a wide variety of parameters, to test for any transient orartifactual effects. From these repetitive runs we establishedthat most effects were long-lasting and that where those weretransient, this was only due to a vehicular load on the roadwhere congestion could no be prevented (as corroborated bycontrol tests).

In the common situation where the obstacle is temporary(i.e., a vehicle malfunction), any reduction in congestionbuild-up allows for the obstacle to be removed, before thevelocity of all vehicles behind the obstacle drops to zero. Inmore complex roadways the advanced warning could alsolead to a change of route, so drivers can avoid the sectionof road where the obstacle exists.

The real-world application of this algorithm is limitedwith current technology, as the recommendation for lanechange can be ignored by a human driver. The current possi-

bility for this algorithm is for an advanced driver informationsystem, where the details of the obstruction are suggested anddisplayed in-car. The modelling of penetration rates that weperformed (Section 6.5) also studies the obedience of humandrivers. In the future, when automated driving systems are inplace and more relied on, this algorithm would form a part ofthe plethora of functions and processes used to drive a vehicleunder computer control.

Buscher et al. [19] discuss the challenges of systemssuch as ours, from a social perspective. The authors workmotivates the need for accurate human-behaviour modellingin vehicular network simulation and what impact the socialaspects of mobility provide. The simulation system wedesigned to perform our studies uses accurate mobilitymodelling, but for a real implementation the social aspects ofthe driver and the context in which they are flowing becomesessential.

In the size of field simulated here we are between themicroscopic and macroscopic scale of simulation, whichis achieved seamlessly by the use of a coupled model ofsimulation. The tool we performed the simulations withwas lightweight and so we could easily implement newalgorithms and protocols. In order to run more complexnetworks we would require a more complex simulationengine for vehicle flow. With this increase in complexity thefield size will increase and therefore a more powerful networksimulator is required. The proof of concept that this paperprovides will lead to further work in this field, including theuse of parallel and distributed computational resource.

Acknowledgments

M. Nekovee acknowledges the Royal Society for supportinghis work through an Industry Fellowship. T. Hewer acknowl-edges the support of British Telecom as part of a CASEIndustrial PhD and funding from the UK Engineering andPhysical Sciences Research Council.

References

[1] M. Nekovee, “Modeling the spread of worm epidemics invehicular ad hoc networks,” in Proceedings of the 63rd IEEEVehicular Technology Conference (VTC ’06), vol. 2, pp. 841–845, May 2006.

[2] J. Yin, T. Elbatt, G. Yeung et al., “Performance evaluation ofsafety applications over DSRC vehicular ad hoc networks,”in Proceedings of the 1st ACM International Workshop onVehicular Ad Hoc Networks (VANET ’04), pp. 1–9, October2004.

[3] S. Eichler, B. Ostermaier, C. Schroth, and T. Kosch, “Sim-ulation of car-to-car messaging: analyzing the impact onroad traffic,” in Proceedings of the 13th IEEE InternationalSymposium on Modeling, Analysis and Simulation of Computerand Telecommunications Systems (MASCOTS ’05), pp. 507–510, September 2005.

[4] M. Torrent-Moreno, Inter vehicle communications, acheivingsafety in a distributed wireless environment: challenges, systemsand protocols, Ph.D. thesis, Universitat Karlsruhe, 2007.

[5] J. Anda, J. LeBrun, D. Ghosal, C.-N. Chuah, and M. Zhang,“VGrid: vehicular ad hoc networking and computing grid for


intelligent traffic control,” Tech. Rep., University of Californiaat Davis, 2005.

[6] K. Maeda, T. Umedu, H. Yamaguchi, K. Yasumoto, andT. Higashino, “MobiREAL: scenario generation and toolsetfor MANET simulation with realistic node mobility,” inProceedings of the 7th International Conference on Mobile DataManagement (MDM ’06), p. 55, May 2006.

[7] O. Wellnitz, S. Lahde, E. Eden, and L. Wolf, “A scenario editorfor mobile ad hoc networks,” Tech. Rep., IBR, TechnischeUniversit at Braunschweig, 2006.

[8] A. Chen, B. Khorashadi, C.-N. Chuah, D. Ghosal, and M.Zhang, “Smoothing vehicular traffic flow using vehicular-based ad hoc networking & computing grid (VGrid),” inProceedings of the IEEE Intelligent Transportation SystemsConference (ITSC ’06), pp. 349–354, September 2006.

[9] M. Treiber, A. Hennecke, and D. Helbing, “Congested trafficstates in empirical observations and microscopic simulations,”Physical Review E, vol. 62, no. 2, pp. 1805–1824, 2000.

[10] A. Kesting, M. Treiber, and D. Helbing, “General lane-changing model MOBIL for car-following models,” Trans-portation Research Record, vol. 1999, pp. 86–94, 2007.

[11] M. Nekovee and B. B. Bogason, “Reliable and efficient infor-mation dissemination in intermittently connected vehicularadhoc networks,” in Proceedings of the 65th IEEE VehicularTechnology Conference (VTC ’07), pp. 2486–2490, April 2007.

[12] IEEE, “IEEE standard for information technology-telecommunications and information exchange betweensystems-local and metropolitan area networks-specificrequirements—part 11: wireless LAN medium access control(MAC) and physical layer (PHY) specifications,” IEEEStd 802.11-2007, (Revision of IEEE Std 802.11-1999), pp.C1–1184, June 2007.

[13] J. Weinmiller, H. Woesner, and A. Wolisz, “Analyzing andimproving the IEEE 802.11-MAC protocol for wireless LANs,”in Proceedings of the 4th IEEE International Workshop onModeling, Analysis, and Simulation of Computer and Telecom-munication Systems (MASCOTS ’96), pp. 200–206, 1996.

[14] R. Struzak, “Radio-wave propagation basics,” Tech. Rep.,ICTP-ITU-URSI School on Wireless Networking for Develop-ment, 2006.

[15] K. Laasonen, “Radio propagation modeling,” Tech. Rep.,University of Helsinki, 2003.

[16] M. Torrent-Moreno, “Inter-vehicle communications: assess-ing information dissemination under safety constraints,” inProceedings of the 4th Annual Conference on Wireless onDemand Network Systems and Services (WONS ’07), pp. 59–64, January 2007.

[17] M. Nekovee, “Epidemic algorithms for reliable and efficientinformation dissemination in vehicular ad hoc networks,” IETIntelligent Transport Systems, vol. 3, no. 2, pp. 104–110, 2009.

[18] D. Waters, “Connected cars ‘promise safer roads’,” BBC News,2007.

[19] M. Buscher, P. Coulton, C. Efstratiou, et al., “Intelligent mobil-ity systems: some socio-technical challenges and opportuni-ties,” in Communications Infrastructure: Systems and Applica-tions in Europe, R. Mehmood, E. Cerqueira, R. Piesiewicz, andI. Chlamtac, Eds., pp. 140–152, 2009.


Research Article

Reliable Delay Constrained Multihop Broadcasting in VANETs

Martin Koubek, Susan Rea, and Dirk Pesch

NIMBUS Centre for Embedded Systems Research, Cork Institute of Technology, Cork, Ireland

Correspondence should be addressed to Susan Rea, [email protected]

Received 26 November 2009; Accepted 5 September 2010


Copyright © 2010 Martin Koubek et al. This is an open access article distributed under the Creative Commons AttributionLicense, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properlycited.

Vehicular communication is regarded as a major innovative feature for in-car technology. While improving road safety isunanimously considered the major driving factor for the deployment of Intelligent Vehicle Safety Systems, the challenges relating toreliable multi-hop broadcasting are exigent in vehicular networking. In fact, safety applications must rely on very accurate and up-to-date information about the surrounding environment, which in turn requires the use of accurate positioning systems and smartcommunication protocols for exchanging information. Communications protocols for VANETs must guarantee fast and reliabledelivery of information to all vehicles in the neighbourhood, where the wireless communication medium is shared and highlyunreliable with limited bandwidth. In this paper, we focus on mechanisms that improve the reliability of broadcasting protocols,where the emphasis is on satisfying the delay requirements for safety applications. We present the Pseudoacknowledgments(PACKs) scheme and compare this with existing methods over varying vehicle densities in an urban scenario using the networksimulator OPNET.

1. Introduction

The US Federal Communications Commission (FCC) andlater the European Telecommunications Standards Institute(ETSI) approved a frequency band reservation in the 5.9 GHz(in Europe 5 GHz) band for wireless communicationsbetween vehicles (V2V) and roadside (V2R) infrastructures.At present, the IEEE group is completing the final draftsof the IEEE 802.11p and IEEE P1609 “Standard for Wire-less Access in Vehicular Environments (WAVEs)” [1]. TheEuropean Commission through programmes like the i2010Intelligent Car Initiative [2], which is a followup of eEurope2005 [3] is driving the rollout of intelligent vehicle systemsin both European and international markets, by supportingICT research and development in the area of transport.Under i2010, eSafety is a collaborative initiative involving theEuropean Commission, industry, and other stakeholders tohasten the development, deployment, and use of IntelligentVehicle Safety Systems (IVSSs) as a means of increasing roadsafety and reducing the number of road traffic accidentswithin Europe.

Integrating a network interface, GPS receiver, sensors,and on-board computers presents an opportunity to build

a powerful car-safety system, capable of gathering, process-ing, and distributing information. By collecting accurateand up-to-date information concerning the status of thesurrounding environment, a driver assistance system canquickly detect potentially dangerous situations and notifythe driver regarding this impending peril. Notifying otherdrivers can be achieved via vehicle-2-vehicle (V2V) commu-nications typically relying on broadcasting as the underlyingdissemination technique. However, broadcasting is a veryexpensive dissemination technique that needlessly consumeschannel communication capacity with increased collisionsand packets losses [4]. A broadcasting protocol for VANETsmust guarantee fast and reliable delivery of informationto all vehicles in the neighbourhood, where the wirelesscommunications medium is shared, very unreliable, andwith limited bandwidth. It must guarantee high deliveryrates for priority messages with emergency payload datain all situations from small vehicle densities (rural areas)to crowded roads in cities during peak times with thecommunication network may be well saturated.

Broadcasting protocols (e.g., [5–9]), that have beenproposed for VANETs have a common factor in that theycannot guarantee high reliability for safety-related data


dissemination with [5] concluding that the probabilityof successful reception of the data decreases with grow-ing distance from the sender. These factors have seriousconsequences for safety-related data dissemination wheredangerous situations can be aggravated through unsuccessfulbroadcast communications.

In this paper, we propose a scheme called Pseudoac-knowledgments (PACK) that interprets successful multi-hop broadcast transmission through overhearing successiverebroadcasts by its neighbours. As the broadcast packettraverses the network, each hop creates dynamic time slotsin order to rebroadcast. Intermediate hops that receive thebroadcast wait until the dynamic slot time expires and thenrebroadcasts thereby acknowledging a link between itself anda previous hop. If the previous hop does not overhear therebroadcast, it repeats the rebroadcasting. The dynamic slotsare created locally at individual nodes and do not require aglobal clock.

The advantage of the PACK method is that it does notneed any extra hardware and can be implemented on top ofany broadcasting protocol, however, our simulation resultshave demonstrated that most gains in efficiency are achievedwith location-based p-persistent CSMA/CA broadcastingprotocols. The PACK schemes rapidly increase receptionprobability of broadcasting protocols with minimal addi-tional overhead in terms of latency and retransmissions. Inthis paper, we compare the efficiency of the PACK methodwith existing schemes for reliable multihop broadcasting thatincrease the reception probability. The network simulatortool OPNET [10] is used to develop an accurate urbanscenario based on the VANET specific WAVE communica-tions protocol with realistic vehicle mobility patterns, radiopropagation model using 802.11p.

2. Related Work

One of the primary concerns for broadcast protocols lies inthe unreliable packet delivery. Protocols such as ALOHA andCSMA are some of the earliest works that focus on mitigatingpacket collisions in uncoordinated networks. Following onfrom this CSMA with Collision Avoidance was developedwhich is the basis for the IEEE 802.11 suite of communi-cations protocols of which IEEE 802.11p for V2V commu-nications is part of it. An RTS/CTS handshake exchangemechanism has been developed for unicast transmissionsto increase reliability, however, broadcast transmissions stillhave to rely only on pure CSMA/CA protocol withoutRTS/CTS. A common concern for broadcasting algorithmsin VANETs is their inability to achieve a packet reception rateclose to 100% [5].

2.1. Multihop Broadcasting Schemes. For multihop broad-casting protocols, several works have proposed acknowl-edging techniques to increase reliability in MARQ [11],BACK [12], and BSMA [13] schemes. These methods arebased on reserving time slots where a sender allocatesvirtual time slots for all its neighbours and transmits thebroadcast data. All its neighbours transmit ACKs in their

virtual slot. The reserving of virtual time slots for individualACK transmissions is problematic in denser networks asit leads to a dramatic increase in latency, a fundamentalconcern for the dissemination of safety related data. Theauthors in [6] proposed a broadcasting protocol called UMBthat uses a handshake like RTS, CTS and ACK for one-directional broadcasting, however, this protocol requiresthe positioning of intersection repeaters that acknowledgethe broadcast along the physical roadways. Other multihopbroadcast protocols presented in [7] include V-TRADE andHV-TRADE. A node wishing to transmit or retransmita broadcast transmits a position request at first to allneighbours and waits until all neighbours reply. After allreplies have been received, the node transmits the broadcastwith a list of selected nodes that act as forwarders similarto OLSR [14]. This was one of the earlier works to addressbroadcasting in VANET, and the overhead incurred with theposition request and reply packets at each hop can contributeto network congestion in denser networks and also increasedelay.

From best of our knowledge, there is no method toincrease broadcast reliability in multihop broadcast protocolsfor VANET networks that do not suffer from dramaticallyrising latency and/or increased load on the physical mediumthrough numerous redundant transmissions. As a precursorto presenting the proposed PACK method, we discuss themechanisms previously developed to increase reliability for1-hop broadcasting.

2.2. 1-Hop Broadcasting Schemes. In [15], the authors haveidentified protocols that increase the reliability of 1-hopbroadcasting schemes and have grouped the schemes basedon their channel access methods.

(i) The first group is based on CSMA/CA where pro-tocols (e.g., [11–13]) use a handshake mechanismcomprising of short packets similar to RTS/CTS/ACKpackets.

(ii) The second group of protocols relies on reservingtime slots in the physical medium. For the RR-ALOHA [16], vehicles must continuously exchange2-hop information to reserve free time slots withoutany central coordination units. The RR-ALOHAwas proposed within the European research projectCarTalk2000 [17].

(iii) The third group relies on the repetition of broad-casting transmissions. The SFR [18, 19] protocolrandomly repeats broadcasted transmissions. Theauthors in [15] propose the OOC code that dynam-ically affects the number of repetition. The OOCmethod performed better against SFR [15, 20], butfor fast moving vehicles the OOC protocol hasdifficulties with codeword synchronisation.

(iv) Another group of protocols not discussed in [15]investigated changing the transmission power ofbroadcasting messages to control the wireless band-width [21, 22]. The Adaptive Transmission Power


(ATP) protocol [21] changes the transmit powerdepending on the number of 1-hop neighbours.

In recent years, several 1-hop broadcasting schemeshave been developed for VANETs whereas not many effortswere invested in improving existing multihop broadcastingschemes described in Section 2.1. For safety-related datadissemination, there will be a prerequisite to disseminationdata beyond a single hop with high reliability for datadelivery over several hops with minimal delay and low datacollisions. In this paper, we propose a multihop scheme that

(i) improves the reliability of multihop broadcast proto-cols,

(ii) with a marginal increase in latency and link load.

The proposed approach is based on creating flexible timeslots at the transmitter and the pseudoacknowledging oftransmissions by rebroadcasting nodes through overhearing.We choose three 1-hop reception schemes namely RR-ALOHA, SFR, and ATP that we have extended for use withmultihop broadcasting and compared their performancewith our proposed scheme we refer to as Pseudoacknowl-edgments (PACKs). We tested the schemes under differentvehicle densities where we emulated local (accidents) andglobal (raining) events. An urban environment has beenselected for experimental evaluation as opposed to a ruralor motorway scenario as this environment will be denselypopulated with slower moving vehicles that force the use ofmultihop broadcast protocols as transmission distances areseverely attenuated with obstacles present in the environ-ment such as buildings, traffic lights, and restricted roadwaysthat cause a build up and congestion in traffic flows.

3. Multihop Broadcast Protocol

The methods for increasing multihop broadcast protocolreliability have been overlaid on the same underlying basebroadcast protocol namely the low latency Slotted RestrictedMobility-Based (SRMB) protocol as opposed to using flood-ing. The SRMB protocol is an extension of the RestrictedMobility-Based (RMB) [23] broadcasting protocol withSRMB minimising data collisions on forwarding broadcastsby using a dynamic slot wait time generated in the upperMAC layer in the order of milliseconds. PACK can be usedwith any broadcasting protocol, but dynamic slot wait times(SRMB) have been shown to reduce collisions by modifyingthe channel access times. Protocols, which include alreadysome form of slot wait times, for example, in case of AODV[24] and OLSR [14] random wait time in the range of 0 to100 ms, do not need necessary integrate SRMB extensions touse PACK.

The RMB, SRMB, and PACK algorithms are describednext, prior to the presentation of the experimental evalua-tions.

3.1. Restricted Mobility-Based (RMB) Protocol. We havepreviously presented the RMB algorithm in [23]. RMB isa flat (nonclustered), uncentralised, p-persistent CSMA/CA

S1S2

S3

S4

Figure 1: Directional sectors are defined about the transmittingnode with a radius defined by the theoretical transmission distancewith each sector having a 90◦ spread.

broadcasting protocol that reduces redundant broadcasttransmissions using 1-hop location knowledge obtainedfrom beacons. RMB was compared with the DV-CASTprotocol [25], with RMB having fewer transmissions, lowerend-to-end delay, and a high delivery ratio.

The basic principle of this algorithm is that beforebroadcasting (rebroadcasting) a transmitter Mi determines asmall set of its neighbours MPRi

1···N (Multipoint Relay set asused in OLSR [14]) with each node lying in a geographicallydifferent sector (maximum N ≤ 4 sectors)

as shown in Figure 1. The transmitter records theshortened MAC addresses of the MPRi

1···N nodes in thepacket header and broadcasts. A node Mj that receives thepacket and has its MAC address recoded in the packet headerassigns a Backoff time slot “0” for rebroadcasting in the MACbuffer. A node Mk, which receives the packet and finds thatits MAC address does not match any address recorded inthe packet header, assigns its Backoff time slot depending onits position, speed, and motion vector compared against thetransmitter in range from 1 to the maximum value of theparticular Contention Window (CW). The maximum size ofCW depends on the type of traffic (voice, video, and data)and ranges from 3 to 15 [1, 26]. A Backoff time slot of “1”refers to nodes that are sufficiently far from the transmitterand have similar speed and motion vector as the transmitter.Larger Backoff time values indicate that nodes have differentmotion vectors and speeds compared to the transmitter [23,Section 3].

To avoid redundant transmissions during broadcasting,each node Mi (MPR, non-MPR) assesses whether all of itsneighbours have received the broadcast packet. This is basedon the knowledge of the position of the transmitter and allneighbours and the knowledge of transmission distance. Ifall neighbours are assessed by the Mi to have received thebroadcast and the Mi has the same broadcast to transmit,then Mi discards the packet and does not rebroadcast.

The RMB scheme ensures that during broadcasting ifa collision occurs at an MPR node, some other non-MPRnode with the second highest priority substitutes as the MPRand rebroadcasts. A strong advantage of this scheme lies in


t1t2 t2

t3 t3

Collision

Figure 2: RMB.

the fast broadcasting process where nodes wait for retrans-mitting generally less than a millisecond. A disadvantage isthat the contention window size of a traffic class may notbe sufficiently large enough to transmit without collisions atnon-MPRs, for example, considering the “Voice” traffic class,there are only 3 Backoff time slots. This implies that non-MPR nodes can, with a high probability, be assigned the sametime slot which leads to collisions, thus, effectively stoppingthe broadcast.

3.2. SRMB and PACK. The RMB algorithm suffers from thehidden terminal problem as illustrated in Figure 2.

A source node Mi broadcasts at time t1 with MPR setMPRi

j,k. All neighbours of Mi that received the broadcastand because MPR nodes (Mj and Mk) set the Backoff timeto “0” and are not within transmission range of each other(Hidden Terminal problem), Mj and Mk rebroadcast attime t2. Because Mj and Mk transmit in the same time(or within a short proximity), a collision occurs aroundMi in t2. The collision generally does not have an effecton surrounding nodes of Mi because these nodes alreadyreceived the broadcast in time t1. But Mi does not overhear(receive) the broadcasts sent by Mj and Mk correctly andso Mi does not know if its own broadcast transmission wassuccessfully received at its neighbours.

To minimise collisions at the source node (and likewiseat intermediate nodes that act as forwarders), we developedthe Slotted Restricted Mobility-Based (SRMB) algorithm andthe Pseudoacknowledgments (PACK) scheme.

3.2.1. Slotted Restricted Mobility-Based (SRMB) Algorithm.The main contribution of the SRMB extension is thatrebroadcasting is carefully scheduled (spread in time) usingdynamic slot wait times (Figure 3). Each node that receivesa broadcast packet assigns a dynamic wait time slot forrebroadcasting to ensure that nodes have sufficient time forrebroadcasting. The wait time slot is derived from the max-imum transmission time TL MAC (1) including processing atlower MAC layer and the time needed for transmission

TL MAC(ac) = LDATA

RDATA+D

c+ SIFS

+ ·TBoSlot · (AIFSN + CW[ac]).

(1)

(The equation is valid only for lightly loaded networks, Inbusier networks, if a transmission is heard while a node is inBackoff, then the new Backoff time is set and transmissiondelay (1) is increased.)

Table 1: Parameters in different traffic categories.

Access Category AIFSN CWmax

CW[background∼WSA] 7 15

CW[voice∼WSM] 2 3

(i) LDATA is the size of data transmitted over the physicalmedium in bits. It contains the data payload, WAVE,and MAC headers.

(ii) RDATA is the data rate in bits per second.

(iii) D is the theoretical distance within which packetscan be successfully received. This depends on theenvironment radio propagation characteristics. Inour simulation, we set the transmission distance to200 m, which has been determined from empiricaldata measurements and is described in Section 4.

(iv) c is the speed of light set to 3× 108 m/s.

(v) SIFS is the short interframe space with a length of16 μs.

(vi) AIFSN specifies the number of “slot” periods withinthe AIFS (Arbitration Interframe Space) value usedby an access category during contention (Table 1).

(vii) AIFS is the lag time between the medium becomingidle and the time when the access category starts orresumes a random Backoff period.

(viii) CW is a number of slots in particular ContentionWindow (Table 1).

(ix) ACs are the Access Categories used by 802.11e andWAVE MAC to manage different traffic classes (voice,video, and data).

(x) TBoSlot is the duration of a slot, this is set to 9 μs.

The SRMB algorithm extends the RMB principle andworks as follows.

A station Mj receives the packet and encapsulates thelist of MPRi

1···N addresses from the incoming packet. If anyaddresses of MPRi

1···N match the Mj address, then before aretransmission Mj adds a delay in length of wait time slotTslot as follows:

Tslot(J) = (J − 1) ·m ·max(TL MAC). (2)

(i) J is J ∈ (1 ≤ N) and is the order of the node Mj inthe list of MPRi

1···N .

(ii) m is a multiplier added to avoid collisions whenthe networks become busy and (1) expires. Thevalue is set to 1.5, which has been determined fromsimulation investigation.

Else if Mj address does not match any of the addressesin MPRi

1···N ,then before a transmission node Mj adds a timedelay according to (2), where

(i) J = N + S;


t1t2

t3 t3

t4

Figure 3: SRMB.

(ii) N is the maximum number of nodes in MPRi1···N ;

(iii) S is the order of the sector where Mj is positioned(Figure 1). A sector is defined about the transmittingnode with a radius defined by the theoretical trans-mission distance with each sector having a 90◦ degreespread [23].

Time slots are chosen based on an MPR node priority,and MPR nodes transmit one by one leaving sufficienttime to avoid collisions at a source node and also to avoidcollisions between other non-MPR nodes in different sectors.

3.2.2. Pseudoacknowledgments (PACK). The principle ofSRMB is that nodes Mj···n broadcast one by one withoutcollisions at the source node or previous forwarding hopMi. As previously described, a broadcasting node definesgeographical sectors and selects its MPR set MPRi

1···N andbroadcasts. Selected neighbours of Mi that receive the broad-cast say Mj and Mk then rebroadcast. The rebroadcastingby Mj and Mk is also received (overheard) at Mi (Figure 3)assuming no collisions. Collisions are mitigated due to thespreading of the retransmissions over dynamic wait timeslots, and so each rebroadcast node should transmit in turnand be overheard by Mi. This overhearing is interpreted bythe PACK method as a form of pseudoacknowledgementfor the individual sectors. If an unacknowledged sector(s)remains after some predefined time (as per (3)), then thenode Mi repeats the broadcast with a new list of MPRi

1···Mthat contains only the missing sector(s). The algorithm isrepeated until all sectors are acknowledged or a maximumnumber of repetitions are reached for the broadcast. Thebroadcast repetition interval Trep is calculated according tothe following equation:

Trep = 2 ·N ·max(TL MAC)

+ Rand (N ·max(TL MAC)).(3)

(i) N is the maximum number of nodes in MPRi1···N .For

other broadcasting protocols other than SRMB, Nrepresents the number of nodes that can possiblyretransmit.

(ii) Rand is a random value uniformly distributed in therange 0 to (N · max(TL MAC)) to further randomiserepetitions over a short time interval to avoid colli-sions.

t1 t1

t2

t3Collisiont4

t5

t6

Figure 4: SRMB+PACK.

The PACK scheme partly solves the Hidden TerminalProblem by using repetitions Figure 4. The maximumrepetitions are set to 3 by default.

The fundamental difference between SRMB+PACK andslotted protocols such as RR-ALOHA is that SRMB+PACKuses access CSMA to the physical medium. Only specificnodes act as forwarders for the broadcast and in turn createvirtual time slots during the broadcasting process at theupper MAC layer to further randomise the channel accesstime to decrease packet collisions. Nodes set the start of therepetition slots based on the time the packet is received soglobal synchronisation is not required and the slot size isdetermined using (3). After this wait time slot expires, thebroadcast packet is passed from the upper MAC layer to thelower MAC layer for transmission according to the particularMAC standard (e.g., [26]).

The converse is true for slotted protocols such asRR-ALOHA, where TDMA is used to access the physicalmedium. All nodes must rely on a global clock for syn-chronization, and each node has its own reserved time slotto transmit with a fixed length, which makes this schemeunsuitable for variable length packets or event/bursty traffic.

3.3. Reliable Broadcast Schemes under Test. In this paper,we compared the proposed multihop SRMB+PACK schemewith 3 other reliable broadcast methods. These mechanismsalso used SRMB as the underlying broadcasting mechanismwith WAVE [1] as the communications protocol.

3.3.1. Synchronous Fixed Retransmission (SFR). SFR hasbeen presented in [18, 19] and is based on repetitivelybroadcasting the same message by a sender. The numberof rebroadcasts is not constant and is randomly chosenaccording to the following principle.

Messages are assumed to have a specific lifetime, and thislife time, is divided into time slots, and from this a randomnumber of these slots are chosen to repetitively transmitthe broadcast. The time slots are synchronized to a globalclock. The authors have proposed other mechanisms buthave shown that SFR achieves the better performance.

In [18], the message lifetime is set to 100 ms, which wesee as unsuitable because it significantly lengthens safetymessaging. Considering this, we decreased the lifetime to10 ms and derived a suitable slot length of 1 ms using (1) and(2), giving sufficient time to perform repetitive broadcasting.


Each sender can randomly choose from 0 to 9 repetitions forbroadcasting and then broadcast in the selected slot.

3.3.2. Adaptive Transmission Power for Beacons (ATPBs). TheAdaptive Transmission Power (ATP) protocol presented in[21] is based on nodes listening to the medium and countingthe collisions that occur in this period. Depending on thisvalue and the number of neighbours, a node decreases orincreases its transmission power appropriately. In [21], thethreshold for the number of neighbours is set to 30, whenthis value is exceeded the transmit power is controlled usingATP.

Irrespective of the packet type, the same power is usedto transmit all messages. Authors [27] highlight that suchan approach leads to dangerously reduced transmissionranges for emergency data and this is counter productive,where emergency data is typically sent on the maximumtransmit power to cover as many nodes as possible. Improvedperformance is achieved using the maximum transmit poweras opposed to broadcasting over multiple hops. As analternative to ATP, we developed the Adaptive TransmissionPower for Beacons (ATPBs), which relies on the same methodof assessing channel, but the transmit power is only modifiedfor the periodic beacons to spare communication capacityfor safety messages that are transmitted with the maximumpossible transmit power.

3.3.3. Reliable Reservation-ALOHA (RR-ALOHA). The RR-ALOHA protocol presented in [16] has been developedwithin the European research project CarTalk2000 [17]. Thisis a slotted technique (TDMA access), where nodes rely onsynchronised time slots for communications, where nodesare assigned a single dedicated slot for transmission. Toprevent nodes from using the same slot, the Reservation-ALOHA (R-ALOHA) [28] protocol uses a central repeaterthat announces used slots. This concept is impracticable foruse within VANETs because the inclusion of static infrastruc-ture would restrict VANET communications to centralisedvehicle-2-Infrastructure communications. To avoid the useof central repeaters, RR-ALOHA [16] was developed andproposes that each node sends beacons containing informa-tion identifying which slot is used for communications withtheir 1-hop neighbours. A node, which receives beacons fromits 1-hop neighbour nodes, indirectly receives informationidentifying the used slots for its 2-hop neighbours. Thisallows nodes to access free slots and to avoid the HiddenTerminal Problem.

4. Simulation Environment

We have developed a VANET simulation environment usingthe network simulation tool OPNET V.12 [10] to evaluatethe performance of the PACK algorithm over the SlottedRestricted Mobility-Based (SRMB) broadcast algorithm andintegrated this with the VANET specific Wave Short MessageProtocol (WSMP) based on a simplified model of theWave communications standard (parameters are shown inTable 2). The Wave model contains one Control Channel

Table 2: Scenario description.

Scenario Urban scenario

Transmit power 18 dBm

Frequency 5.9 GHz

Data rate 6 Mbit/s

Bandwidth of channel 10 MHz

Transmit range Two-Ray Ground modelwith shadowing

Minimum broadcast distance 500 m

Maximum num hops 10

Speed of vehicles 0–50 km/h

Scenario dimensions 2 km × 2.5 km

Density of nodes/km2 10–140

Number of Hazardous locations 3 (accident), 1 (rain)

Repetition interval of safety messages 1 s

Size of beacons 480b

Size of safety messages 368b

Beaconing interval 100 ms

(CCH) and one Service Channel (SCH) interface with totalchannel duration of 100 ms with 50 ms per channel thatswitch periodically at 50 ms intervals.

To approximate real world radio propagation, we imple-mented a realistic radio propagation models in OPNET.The model is based on the Two-Ray Ground model withshadowing, where the parameters are set based on empiricaltesting of 802.11p radio modules [29]. The packet loss ratiois in the region of 40% for distances up to 100 m between thetransmitter and receiver while the losses increase to 90% withdistances of between 100 m and 150 m, and 100% losses areachieved with distances beyond 200 m.

For experimental investigation, we modelled an urbanscenario using the road traffic simulator SUMO [30], wherethe scenario represents a topology of collector roads in a5 km2 area in the Bishopstown district in Cork City, Ireland.The traffic model contained dynamically moving vehicleswith varying speeds that are restricted to a maximum speedof 70 kmph along 2-lane roads with a mixture of signalledintersections, traffic circles, and stop signs. The density ofvehicles ranged from 10 to 140 vehicles per km2, whichrepresented traffic flows at night time to peak time. Two typesof emergency situations were investigated representing safetyof life applications and low-priority hazard/environmentalwarning applications.

The first scenario emulates 3 accidents in 3 roads inlow, medium, and high density road sections. Accidentscan be detected by vehicles within 50 m of the accidentlocation. A vehicle entering this 50 m sensing range detectsand immediately invokes a broadcast relating to this emer-gency. A vehicle that is within this 50 m range when theaccident occurs selects a random wait time over a uniformlydistributed interval of 100 ms (corresponds to the WAVESYNC INTERVAL) before broadcasting. This distributes thegeneration of broadcasts over the complete WAVE frame andrandomises the intervals at which vehicles rebroadcast and


lessen collisions due to broadcast storms. The broadcastingis repeated at 1 s intervals.

The second scenario was designed to focus on thethroughput of the whole network and emulates an environ-mental network wide event, rain detection in this case. Eachscheme was tested with different loads in the network. Allvehicles detect the rain event uniformly distributed in timeover 1 s and repeatedly broadcast every 1 s.

5. Performance Analysis

In the simulated environment, only two types of messagesare transmitted. Beacon messages WSA [1] were transmittedevery 100 ms by each node, and safety messages wereencapsulated in WSM [1] packets and broadcasted withthe Minimum broadcast distance being set to 500 m andthe Maximum hops being set to 10 hops. We collected thesimulation results from 3 seeds with at least 200 runs foreach seed. The metrics recorded from the experiments areoutlined below and shown in Figures 5–10.

Network overview (Figure 5)—shown in this diagram—is the mean number of 1st hop and 2nd hop neighbours thatnodes have in the network. The diagram shows a limitationof ATPB and RR-ALOHA schemes.

Link Load (Figure 6)—this is calculated as the meanratio of the number of nodes that transmit safety broadcastpacket against the number of nodes that receive the packet.The lower this value the better as this indicates that fewertransmissions are needed to disseminate the broadcastpacket.

End-to-End Delay (Figure 7)—this is a measure of themean time delay between the source of a safety message andthe node that receives the broadcast last. This also coversthe time delay created by time slots CCH TS and SCH TSin the Wave protocol. In the case of the SFR scheme, thisis measured as the delay between the source node and thereception of last repetition broadcast.

Delivery Ratio (Figure 8)—this measure is dependenton the density of a network and it is the mean deliveryratio taken as the number of nodes inside an area thatreceive safety broadcast versus the number of nodes in thatarea. The area was defined by a source node as an areainside a circle with the source node at the centre, and theradius is defined by the Minimum broadcast distance. Forthe SFR scheme, this was measured based on the numberof nodes inside the area that received safety broadcast(from any repetition) versus the number of nodes in thearea.

Delivery Ratio versus Distance (Figure 9)—this shows theeffect on the mean delivery ratio against increasing distancefrom the source up to the Minimum broadcast distance.

Throughput (Figure 10)—this measure was collectedover the complete network and refers to global networkevents. All vehicles in a scenario detect a global event (e.g.,raining) using sensors and all vehicles broadcast this event.The purpose of this measurement was to investigate theimpact of the broadcast repetition interval, which was variedfrom 0.01 to 3 packets per second, on the delivery ratio (the

number of nodes that receive the broadcast) in a network thatwas moderately busy, with 60 vehicles/km2.

6. Theoretical and Experimental Results

6.1. Theoretical Results. We compare the proposed PACKscheme with 3 existing schemes, namely SFR, ATPB, andRR-ALOHA. All the schemes were overlaid on the SRMBbroadcasting protocol. According to the WAVE standard [1],time was divided to frames (Sync interval) with a length of100 ms. Each frame contains two slots the Control Channel(CCH TS) and the Service Channel (SCH TS) time slots,each with a length of 50 ms. Each of these slots begins witha Guardian time of 5 ms to allow a unit to switch fromone channel to another. In the Guardian time interval, nomessages can be sent. Beacon messages and safety messageswere sent only in CCH TS after the Guardian time. If asafety message was sent in CCH TS, the beacon message wasomitted to prevent overloading the medium.

For repeated broadcasting of an event (local, global), theinvoking of safety messages was uniformly distributed acrossthe Sync interval with a length of 100 ms. If a safety messagewas invoked during the SCH TS 50 ms interval or Guardiantime 5 ms duration, then it waited until the beginning ofthe CCH TS where it was immediately transmitted. A meantime delay TH MAC for waiting emergency data (WSM) atthe upper MAC layer before being passed to the lower MAClayer to access the CCH TS is calculated as per the followingequation:

TSRMBH MAC =

TSCH+G

Tsync· TSCH+G

2≈ 15 ms. (4)

(i) TSCH+G is the time in length of SCH TS (50 ms) plusGuardian time (5 ms) when emergency data cannotbe sent.

(ii) Tsync is the length of Sync interval 100 ms specifies inWave [1].

The Mean theoretical overall time delay for multihopbroadcasting TSRMB is calculated as per equation (5), whichis derived from (1), (2), and (4) as follows:

TSRMB = TSRMBH MAC + H · (Tslot(J) + TL MAC)

TSRMB ≈ 18 ms.(5)

(It presumes that all transmissions were made in one CCHTS. Otherwise the TH MAC was extended to 55 ms (length ofSCH TS and Guardian time).)

(i) LDATA in (1) is the size in bits of an emergency packet(WSM) with a value of 368 bits.

(ii) H is the mean number of hops and is set to 6. Thenumber was taken from mean number of hops in thesimulations that increased with increasing density.

(iii) It is presumed that Tslot with J ∈ (1 ≤ N) is thedelay applied mainly at the origin of the broadcast,


where broadcasts are sent in different sectors basedon the priority of the MPR nodes. Here, J representsthe average number of MPR nodes per hop, based onsimulation evaluation this was set to J = 1.5.

6.1.1. Pack. In case of using the PACK scheme, the overallmultihop delay TSPACK is slightly increased due to thefollowing repetitions:

TPACK = TSRMB + k · Trep,

TPACK ≈ 22 ms,(6)

where k is the number of repetitions. This value dependson the data traffic on the physical medium, where inless busy network the repetition value was approximatelyone repetition per the complete broadcast and this wentup to approximately 7 for busy networks. For theoreticalestimation, we set k = 2.5, which is compared to mediumbusy network.

The SRMB+PACK scheme increased end-to-end delay ofSRMB protocol by 22% (18 ms compared to 22 ms).

6.1.2. Synchronous Fixed Retransmission (SFR). In case ofusing the SFR scheme, the overall multihop delay TSFR wascalculated as follows:

TSFR = TSRMB + k · TSFR slot,

TSFR ≈ 23 ms.(7)

(i) k is the mean number of broadcast repetitions equallydistributed from 0 to 9 as specified by the SFRscheme.

(ii) TSFR slot is a slot in length of 1 ms specifies by SFRscheme.

The SFR scheme increased end-to-end delay of SRMBprotocol by 28% and by 5% when compared againstSRMB+PACK.

6.1.3. Adaptive Transmission Power of Beacons (ATPBs). Thetheoretical overall time delay of multihop broadcasting waskept the same as in SRMB protocol. From the perspectiveof broadcasting delay, the ATPB and SRMB schemes workon the same principle. ATPB only affects the transmissionpower of the beacons and does not straight impact on thedissemination of emergency (WSM) data.

6.1.4. Reliable Reservation-ALOHA (RR-ALOHA). In RR-ALOHA, the beacon (WSA) contained a list of all time slots,where each entry relates to particular time slot. Each entryin the list had a size of 11 bits and contained informationrelating to the state of the channel (busy or idle) and theshort MAC address of the node transmitting on that timeslot. Because we implemented RR-ALOHA over the WAVEstandard, we had to derive the maximum number of slots

first. The size of beacons LDATA used by RR-ALOHA wascalculated in as follows:

LRR-ALOHADATA = LMAC + LWSA + LRR-ALOHA,

LRR-ALOHA = 11 bits · S,(8)

where LMAC is the size of the MAC header with 272 bits, andLWSA is the size of the WSA beacons with length of 480 bits.From a knowledge of the maximum available time of 45 msin the CCH TS and from maximum transmission delay (1),we determine that the maximum number of time slots S usedby RR-ALOHA is 90 with length of a slot being 0.5 ms. Theoverhead LRR-ALOHA was calculated as 11 bits × 90 time slots,which is 990 bits.

The mean theoretical overall multihop delay TRR-ALOHA

was calculated as follows:

TRR-ALOHAH MAC ≈ 50 ms,

TRR-ALOHA = TRR-ALOHAH MAC ·H + TL MAC,

TRR-ALOHA ≈ 300 ms.

(9)

The delay TRR-ALOHA depends on the number of hops Hand how many retransmit nodes are chosen. Theoretically,with 10 hops (10 hops in the maximum number of allowablehops for a broadcast) the delay can vary from 18 ms (see (5))to 1000 ms (see (9)) depending on the selecton of forwardinghops and their time slot.

RR-ALOHA gives the longest delay, 17 times higher thanSRMB and 14 times higher than SRMB+PACK and SFR.

6.2. Experimental Results. All the results presented arerepresented by mean values for individual data points whichare averaged over approximately 600 values with 3 seeds. Thedata sets in most cases have a skewed distribution, so it ispreferable to use the first and third quartiles (q25, q75) asdescriptive statistics.

6.2.1. Network Overview. Network overview (Figure 5)—thisshows the mean number of 1-hop and 2-hop neighbours thatnodes have in the network. In [21], for the ATP scheme theneighbour threshold is set to 30 nodes, meaning that if anode has more than 30 neighbours, then the node shouldchange its transmission power which would then affect thebroadcasting performance. As can be seen in Figure 5, thenumber of neighbours exceeds 30 between x = 40 andx = 60 (vehicles/km2). The RR-ALOHA protocol uses timeslots, where the number of time slots, was set to 90 using(1) and (8). The number of 1st hop and 2nd hop neighboursexceeds the maximum number of slots, that is, 90 at x = 60(vehicles/km2), beyond this density some nodes will have toshare the same time slot.

The results show a limitation of the ATPB and RR-ALOHA as the number of neighbouring nodes can affectthe broadcast performance. SRMB+PACK and SFR are notrestricted by number of neighbours and can work across allneighbour densities.


0

50

100

150

200

0 50 100 150

Nu

mbe

rof

nei

ghbo

urs

(vehicles/km2)

1st neighbours2nd neighbours

Figure 5: Network overview shows the mean number of 1st hopand 2nd hop neighbours that nodes have in the network.

6.2.2. Link Load. Link Load (Figure 6) showed that allschemes (except SFR because of repetitions) have a rapidlydecreasing link load trend. As the vehicle density increases,the network connectivity goes from sparsely connected towell connected. After SFR, the SRMB protocol performsthe next worst in terms of link load (with a Link LoadRatio mean value of #LL = 0.25, with 1st and 3rdquartiles being q25 = 0.14, q75 = 0.25, taken at thehighest density of vehicles with x = 140 vehicles/km2)in denser networks. The PACK scheme in lower densitynetworks performs marginally poorer (5%, this drop inperformance is attributed to the repetition of broadcastsfor unacknowledged sectors) than SRMB with #LL = 0.77,q25 = 0.5, and q75 = 1, at a vehicle density x = 10/km2

in less busy networks. For higher density networks withx = 140 vehicles/km2,values of #LL = 0.22, q25 =0.18, and q75 = 0.27 are achieved, and this represents animprovement of 12% when compared with SRMB. For moresaturated networks, the pseudoacknowledgements used byPACK to acknowledge sectors reduce the probability of non-MPR nodes rebroadcasting and thus reduce the probabilityof collisions which results in fewer transmissions in thecongested medium. The ATPB and SRMB schemes have asimilar performance as the power control aspect of ATPBonly applies to the beacons. The best performance acrossall densities was achieved by RR-ALOHA as expected witha 40% improvement over SRMB at a vehicle density of x =140/km2 and #LL = 0.15, q25 = 0.12, and q75 = 0.15. Thisperformance is attributed to the fact that RR-ALOHA usesone slot per node transmissions and will always outperformCSMA/CA methods, on which the other schemes are based.The better performance in terms of link load is offset bythe poor end-2-end delay and throughput achieved with RR-ALOHA. The worst performance, that is, the greatest numberof transmissions was attributed to SFR, which significantlydiffers from the other schemes. In the lightest density, SFRreached a value (#LL = 4, q25 = 2, q75 = 6, and densityx = 10) 5 times greater than SRMB. In the heaviest density

0

1

2

3

4

5

6

7

8

0

0.2

0.4

0.6

0.8

1

0 50 100 150

Lin

klo

adra

tio

(vehicles/km2)

SRMB+PACK+ATPB

+RR-ALOHA+SFR

Figure 6: Link Load Ratio is calculated as the mean ratio of thenumber of nodes that transmit a broadcast packet against thenumber of nodes that receive the packet. Second right y axis is forSFR scheme, which significanty differ from the others.

network, SFR flooded the network, which led to rapidlyincreasing unsuccessful transmissions (#LL = 7, q25 = 0.5,q75 = 2, and density x = 140) with values 30 times greaterthan in SRMB.

The link load results show that all schemes (exceptSFR) perform broadcasting with a very low number oftransmissions and decrease the number with increasingdensity of vehicles as network increasing in connectivity dueto a larger number of nodes. At higher densities, SFR floodednetwork because of repetitions and is actually worse thanusing a simple flooding protocol making SFR unsuitable forVANETs.

6.2.3. End-to-End Delay. End-to-End Delay (Figure 7)—Asexpected due to their similar operation, the results for end-to-end delay showed that the SRMB protocol and the ATPBscheme maintain the same relatively constant short timedelay (End-to-End Delay, #EE = 20 ms, q25 = 4, q75 = 37,and density x = 140), which matched the theoretical resultachieved with (5). In comparison, the PACK method hada slightly increased delay across all densities from lighterdensities (#EE = 18 ms, q25 = 0.4, q75 = 33, and densityx = 10) with a deterioration in performance when comparedwith SRMB of 12% and in larger densities (#EE = 33 ms,q25 = 8, q75 = 50, and density x = 140) a deterioration of50% again comparing to SRMB. Using (6) and a repetitionfactor of 2.5 and looking at a medium density network withx = 40/km2,theoretical results matched experimental result(#EE = 22 ms, q25 = 4, q75 = 39, and density x = 40).The SFR scheme gave the 2nd longest delay across a low-density network (#EE = 22 ms, q25 = 5, q75 = 40, anddensity x = 10) to a high-density network (#EE = 47 ms,q25 = 11, q75 = 60, and density x = 140) with a deteriorationin performance ranging from 40% to 240% when comparedagainst SRMB. Using (7), the theoretical end-to-end delay


0

200

400

600

800

0

20

40

60

80

100

En

d-to

-en

dde

lay

(ms)

0 50 100 150

(vehicles/km2)

Figure 7: End-to-End Delay is a measure of the mean time delaybetween the source of a safety message and the node that receivesthe broadcast last. Second right y axis is for RR-ALOHA scheme,which significanty differ from the others. Label is the same as atFigure 6.

does not match the empirical result. Equation (7) is derivedusing the maximum transmission time TL MAC from (1)which does not consider a saturated case (i.e., collisions arenot considered). Equation (1) is valid only for lightly loadednetworks. In more dense networks if a transmission on themedium is detected while a node is in Backoff, a new Backofftime is set and the transmission delay (1) is increased. Withthe SFR protocol, we have an increasing load on the physicalmedium as a consequence of repetitions that saturate thenetwork and lead to collisions. The longest delay is givenby RR-ALOHA across all densities from the lowest (#EE =120 ms, q25 = 40, q75 = 176, and density x = 10) tohighest (#EE = 740 ms, q25 = 580, q75 = 920, and densityx = 140) with a deterioration from 7.5 times to 35 timesthat of SRMB. Using (9), we see that the theoretical resultdepended strongly on the number of hops and the selectionof the next hops based on their time slots. This variation wasdescribed in Section 6.1.4 and matched experimental results.

The results showed that the SRMB, ATPB, PACK, andSFR schemes reach a fraction of driver reaction time (around0.05 s of 0.7 s [31]). On the basis of the results, we show thatthese schemes in terms of end-to-end delay are appropriatefor VANETs. As RR-ALOHA has prohibitively long end-to-end delays across all densities, we conclude that thismethod based on comparison with driver reaction speeds isunsuitable for emergency data dissemination in VANETs.

6.2.4. Delivery Ratio. Delivery Ratio (Figure 7)—Resultsshowed that the SRMB protocol reached relatively constantvalues for Delivery Ratio, #DR = 0.62, q25 = 0.48, q75 = 0.90,and density x = 40 to #DR = 0.61, q25 = 0.43, q75 =0.92, and density x = 140. Similar results were achievedwith ATPB and acknowledged that sparing communicationcapacity by decreasing transmit power of beacons did nothave significant effect on delivery ratio. The PACK methodin low-density network gave values of (#DR = 0.35, q25 =0.18, q75 = 0.45, density x = 10) and in high density

0.2

0.4

0.6

0.8

1

0 50 100 150

(vehicles/km2)

SRMB+PACK+ATPB

+RR-ALOHA+SFR

Del

iver

yra

tio

Figure 8: Delivery Ratio is a measure of the mean delivery ratiotaken as the number of nodes inside an area that receive a broadcastversus the number of nodes in that area.

gave (#DR = 0.83, q25 = 0.77, q75 = 0.96, and densityx = 140). These results reflect improvements of 2% to36% when comparing against SRMB from medium- to high-density networks (#DR = 0.70, q25 = 0.57, q75 = 0.97,and medium density x = 40) with PACK improving overallother methods in medium- to high-density networks. SFRgives the best performance in lower density networks becauseof the repetitions in a sparsely connected network (#DR =0.42, q25 = 0.30, q75 = 0.55, and density x = 10) withimprovements of 32% over SRMB. In medium busy densities(#DR = 0.70, q25 = 0.60, q75 = 0.98, and density x = 40),SFR has a slight deterioration of 4% when compared toPACK and in the highest density (#DR = 0.78, q25 = 0.77,q75 = 0.97, and density x = 140), the decline in performancefalls to 6% when compared to PACK. The RR-ALOHAscheme gave a slightly poorer results when compared toPACK (low density (#DR = 0.33, q25 = 0.18, q75 = 0.45, anddensity x = 10) to high density (#DR = 0.80, q25 = 0.76,q75 = 0.96, and density x = 140) with a 5% decline inperformance.

The results show that PACK, SFR, and RR-ALOHAsignificantly improved delivery ratio across all densities withATPB giving a performance again similar to SRMB. Againthis shows the unsuitability of the ATPB protocols for reliablebroadcasting in VANETs as it only refers to beacon frames.

6.2.5. Delivery Ratio versus Distance. Delivery Ratio versusDistance (Figure 8)—these results were captured at a densityof 60 vehicles/km2 (medium busy network) and showed thatfor all schemes the Delivery Ratio fell of with increasingdistance. Again SRMB and ATPB give similar results. PACK,SFR, and RR-ALOHA improve on SRMB and give verysimilar values up to a distance, x = 250 m from a sender,with an improvement of 18% over SRMB.


0.2

0.4

0.6

0.8

1

0 100 200 300 400 500

Del

iver

yra

tio

Distance (m)

SRMB+PACK+ATPB

+RR-ALOHA+SFR

Figure 9: Delivery Ratio shows the effect on the mean DeliveryRatio against increasing distance.

Further, from the sender at a distance x = 350 m, SFRgives an improvement of 14%, PACK gives 22% and RR-ALOHA 28% over SRMB. On the boundary of the Minimumbroadcast distance (500 m), improvements with SFR were23%, RR-ALOHA 25%, and PACK 28% over SRMB.

The results showed that with increasing distance thedelivery ratio falls off, but again SFR, PACK, and RR-ALOHAimprove broadcasting performance.

6.2.6. Throughput. Throughput (Figure 10)—results showedthat all schemes have a decreased delivery ratio for broadcastswith increasing load in the network. As before, SRMB andATPB perform similarly. SRMB (likewise ATPB) give thelowest throughput ratio #TP = 0.45, q25 = 0.27, q75 = 0.72,and load x = 0.01 packets/1 s under low load and thisdecreases #TP = 0.32, q25 = 0.04, and q75 = 0.53, withincreasing load x = 3 packets/1 s (high load). In low load,x = 0.01 packets/1 s RR-ALOHA has #TP = 0.49, whichimproves on SRMB by 9% and maintained a trend similar toPACK until reaching a moderately loaded network x = 0.3packets/1 s, where #TR = 0.25, q25 = 0.02, q75 = 0.43,and load x = 3 packets/1 s fell to 22% below SRMB. Thisperformance deterioration is attributed to the fact that RR-ALOHA uses TDMA access and will always perform worsethan CSMA/CA access, on which the other schemes arebased. SFR reached the highest values with #TR = 0.58, q25 =0.30, and q75 = 0.87 under low load x = 0.01 packets/1 s anddeteriorated #TR = 0.42, q25 = 0.30, and q75 = 0.73 undergreater load x = 3 packets/1 s due to saturating the networkwith repeat broadcasts but still improved on SRMB by 31%.The PACK scheme reached values similar to RR-ALOHAunder low load. With increasing load, the PACK improveson these with #TR = 0.45, q25 = 0.22, and q75 = 0.68, highload x = 3 packets/1 s giving an improvement of 40% overSRMB, 80% over RR-ALOHA, and 7% over SFR.

The results showed that with increasing load thethroughput ratio dropped across all schemes. TDMA-based

0

0.2

0.4

0.6

0.8

1

10.01 0.1 10

Th

rou

ghpu

t

Invoked broadcasts per second per vehicle in 60 vehicles/km2

SRMB+PACK+ATPB

+RR-ALOHA+SFR

Figure 10: Throughput measure an impact of the broadcastrepetition interval on the delivery ratio.

channel access in RR-ALOHA shows reduced throughputagainst CSMA/CA-based access in highly loaded networks.SFR scheme showed decreased throughput due to overload-ing of network with many repetitions but still maintainsa higher performance then SRMB. The PACK scheme inmoderate to highly loaded networks gives a better perfor-mance than the SFR, RR-ALOHA, and SRMB (ATPB). Theseadvantages of the PACK algorithm can be further highlightedwhen safety application is required to report with a higherrate (<1s) as it can maintain high delivery ratio for moderate-to high-density networks (e.g., vehicle platooning).

7. Discussion, Conclusions, and Future Work

In this paper, we have concentrated on techniques thatincrease reliability of a multihop broadcast protocol. We haveproposed the Pseudoacknowledgments (PACK) scheme thatimproves reliability in multihop broadcasting protocols byrepeating broadcast transmissions on unsuccessful links. Thescheme was compared with existing mechanisms over anurban scenario using the network simulator tool OPNET[10] with an empirical-based propagation model [29], realis-tic mobility patterns using the road traffic simulator SUMO[30], and the Wave [1] standard. All schemes were overlaidon the low-latency p-persistent CSMA/CA broadcastingprotocol called the Slotted Restricted Mobility-Based (SRMB)protocol.

The scenarios were designed to test safety-related datadissemination and measured relevant broadcasting statistics:link load, end-to-end delay, delivery ratio, and throughput.From our results, we draw the following conclusions andidentify tasks for future work.

(i) Changing transmission power of beacons does nothave a significant effect on performance of broad-casting, as has been demonstrated by examining theperformance of ATPB.


(ii) Repeating broadcasts leads to increased delivery ratio,but it also increases the number of transmissions inthe network. This can lead to flooding the networkwith repetitions and can decrease the delivery ratioin denser networks. The throughput results show thatSFR scheme can easily saturate the network underhigher loads which leads to a rapidly decreasing deliv-ery ratio. The redundancy incurred as a consequenceof repetitions which can lead to flooding makes thisscheme unsuitable for VANETs.

(iii) Using small time slots for broadcasts leads to ahigh delivery ratio due to minimum collisions inparticular slots but this increases end-to-end delayas broadcasts are delayed at successive rebroadcastnodes. In the case of RR-ALOHA, the delay reacheslarge values that cannot be tolerated for safety-relateddata dissemination. Another disadvantage of slottingis that it decreases throughput in densely loadednetwork which corresponds with the throughputperformance when comparing CSMA/CA access withALOHA access. When the number of nodes exceedsthe maximum number of available slots, nodes mustshare slots. However, this limitation does not have asignificant effect on the broadcasting performance inour simulations.

(iv) In Summary, PACK in moderate-to high-densitynetworks achieves an increase in delivery ratio ofapproximately 5% over SFR and 3% over RR-ALOHAwhile this may appear to be a minor incrementRR-ALOHA is unsuitable for VANETs due to itsexcessive end-to-end delay, for example, 741 ms forRRALOHA against 32 ms for PACK in high-densitynetworks. While SFR has a tolerable end-to-enddelay for VANETs, it does so at the expense ofsaturating the network with repeat broadcasts andthis is highlighted by the link load metric whichat a vehicle density of 140/km2 is 35 times higherthan of PACK. While PACK and SFR give similardelivery ratio metrics, SFR does so at the expense ofexcessive bandwidth usage in comparison to PACK.Furthermore, due to less saturating of the networkPACK achieved the highest throughput in moder-ate to high density networks. This makes PACKsuitable for applications that need to frequentlyreport (e.g., vehicle platooning, crashed vehicledetection).

(v) From the experimental results presented in thispaper, we can conclude that the PACK mechanismincreases the reliability of multihop broadcasting andis suitable for safety-related data dissemination. Interms of end-to-end delay and bandwidth savings,PACK outperforms RR-ALOHA and SFR, respec-tively, making PACK a more reliable protocol forsafety data dissemination in VANETs.

Although SRMB+PACK protocol has been primarilytested on a specific case of safety application (accidentreporting), the SRMB+PACK mechanism can be used as

a multihop data dissemination mechanism for a rangeof applications that require high packet delivery and lowlatency in very dynamic ad hoc networks. The SRMB+PACKprotocol could be used in route discovery for reactive routingprotocols in VANETs. From the route discovery, perspectiveroutes would be built based on latency, bandwidth con-sumption, and mobility of nodes in the source-destinationpath. Nodes with similar mobility behaviour (speed, motionvector) would be selected as intermediate hops as thissupports the generation of stable routes and reduces routemaintenance overhead.

References

[1] “Intelligent Transportation Systems Standards Fact Sheet,”IEEE 1609—Family of Standards for Wireless Access inVehicular Environments (WAVE), ITS Standards Program.

[2] http://europa.eu/legislation summaries/transport.[3] http://ec.europa.eu/information society/eeurope/2005/index

en.htm.[4] W. Chen, R. K. Guha, T. J. Kwon, J. Lee, and I. Y. Hsu, “A survey

and challenges in routing and data dissemination in vehicularad-hoc networks,” in Proceedings of the IEEE InternationalConference on Vehicular Electronics and Safety, pp. 328–333,Columbus, Ohio, USA, September 2008.

[5] M. Torrent-Moreno, D. Jiang, and H. Hartenstein, “Broadcastreception rates and effects of priority access in 802.11-based vehicular ad-hoc networks,” in Proceedings of the 1stACM International Workshop on Vehicular Ad Hoc Networks(VANET ’04), pp. 10–18, Philadelphia, Pa, USA, 2004.

[6] G. Korkmaz, E. Ekici, F. Ozguner, and U. Ozguner, “UrbanmultiHop broadcast protocol,” in Proceedings of the 1st ACMInternational Workshop on Vehicular Ad Hoc Networks, pp. 76–85, Philadelphia, Pa, USA, 2004.

[7] M. Sun, W. Feng, T. Lai, K. Yamada, H. Okada, and K.Fujimura, “GPS-based message broadcasting for inter-vehiclecommunication ,” in Proceedings of the International Confer-ence on Parallel Processing, pp. 279–286, Toronto, Canada,2000.

[8] M. Heissenbuttel, T. Braun, M. Walchli, and T. Bernoulli,“Optimized stateless broadcasting in wireless multi-hop net-works,” in Proceedings of the 25th IEEE International Confer-ence on Computer Communications (INFOCOM ’06), pp. 1–12,Barcelona, Spain, April 2006.

[9] W. Si and C. Li, “RMAC: a reliable multicast MAC protocol forwireless ad hoc networks,” in Proceedings of the InternationalConference on Parallel Processing (ICPP ’04), pp. 494–501,August 2004.

[10] http://www.opnet.com.[11] J. Peng, “Slim ARQ for reliable broadcasting in wireless

LANs,” in Proceedings of the IEEE Wireless Communicationsand Networking Conference (WCNC ’08), pp. 2164–2169, 2008.

[12] S.-T. Sheu, Y. Tsai, and J. Chen, “A highly reliable broadcastscheme for IEEE 802.11 multi-hop ad hoc networks,” inProceedings of the IEEE International Conference on Commu-nications (ICC ’02), vol. 1, pp. 610–615, New York, NY, USA,2002.

[13] K. Tang and M. Gerla, “Random access MAC for efficientbroadcast support in ad hoc networks,” in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC ’00), pp. 454–459, Los Angeles, Calif, USA, 2000.

[14] http://www.ietf.org/rfc/rfc3626.txt.


[15] F. Hassanzadeh, Reliable broadcast of safety messages in vehic-ular ad hoc networks, M.S. thesis, Graduate Department ofElectrical and Computer Engineering University of Toronto,Toronto, Canada, 2008.

[16] F. Borgonovo, A. Capone, M. Cesana, and L. Fratta, “RR-ALOHA, a Reliable R-ALOHA broadcast channel for ad-hocinter-vehicle communication networks,” 2002, ftp://ftp.elet.polimi.it/users/Flaminio.Borgonovo/PUBBC/2002medhoc.pdf.

[17] http://www.cartalk2000.net.[18] Q. Xu, T. Mak, J. Ko, and R. Sengupta, “Medium access control

protocol design for vehicle—vehicle safety messages,” IEEETransactions on Vehicular Technology, vol. 56, no. 2, pp. 499–518, 2007.

[19] Q. Xu, R. Senguptay, T. Mak, and J. Ko, “Vehicle-to-vehiclesafety messaging in DSRC,” in Proceedings of the 1st ACMInternational Workshop on Vehicular Ad Hoc Networks, pp. 19–28, 2004.

[20] F. Farnoud and S. Valaee, “Repetition-based broadcast invehicular ad hoc networks in Rician channel with capture,” inProceedings of the 27th Conference on Computer Communica-tions (INFOCOM ’08), pp. 1–6, 2008.

[21] L. Yang, J. Guo, and Y. Wu, “Channel adaptive one hopbroadcasting for VANETs,” in Proceedings of the 11th IEEEConference on Intelligent Transportation Systems (ITSC ’08),pp. 369–374, Beijing, China, 2008.

[22] M. Torrent-Moreno, P. Santi, and H. Hartenstein, “Distributedfair transmit power adjustment for vehicular ad hoc net-works,” in Proceedings of the 3rd Annual IEEE CommunicationsSociety on Sensor and Ad Hoc Communications and Networks(Secon ’06), vol. 2, pp. 479–488, Reston, Va, USA, 2006.

[23] M. Koubek, S. Rea, and D. Pesch, “Effective emergencymessaging in WAVE based VANETs,” in Proceedings of the1st International Conference on Wireless Access in VehicularEnvironments (WAVE ’08), December 2008.

[24] http://www.ietf.org/rfc/rfc3561.txt.[25] O. Tonguz, N. Wisitpongphan, F. Bai, P. Mudalige, and V.

Sadekar, “Broadcasting in VANET,” in Proceedings of theMobile Networking for Vehicular Environments (MOVE ’07),pp. 7–12, May 2007.

[26] “802.11e-2005, IEEE Standard for Information technology—telecommunications and information exchange betweensystems—local and metropolitan area networks—specificrequirements—part 11: Wireless LAN Medium Access Control(MAC) and Physical Layer (PHY) specifications Amendment8: Medium Access Control (MAC) Quality of Service Enhance-ments,” 2005.

[27] J. Mittag, F. Thomas, J. Harri, and H. Hartenstein, “Acomparison of single- and multi-hop beaconing in VANETs,”in Proceedings of the 6th ACM International Workshop onVehicular Inter-Networking (VANET ’09), pp. 69–78, 2009.

[28] W. Crowther, R. Rettberg, D. Waldem, S. Ornstein, and F.Heart, “A system for broadcast communications: reservationALOHA,” in Proceedings of the 6th Hawaii InternationalConference on System Sciences, January 1973.

[29] O. Brickley, M. Koubek, S. Rea, and D. Pesch, “A networkcentric simulation environment for CALM-based cooperativevehicular systems,” in Proceedings of the 3rd InternationalConference on Simulation Tools and Techniques (SIMUTools’10), 2010.

[30] http://sumo.sourceforge.net.[31] M. Green, ““How long does it take to stop?” Methodological

analysis of driver perception-brake times,” TransportationHuman Factors, vol. 2, no. 3, pp. 195–216, 2000.


Research Article

Traffic Flow Condition Classification for Short Sections UsingSingle Microwave Sensor

Muhammed G. Cinsdikici1 and Kemal Memis2

1 International Computer Institute, Ege University, 35100 Bornova, Turkey2 Traffic Systems Department, Aselsan Mil.Electrnoics Company, ASELSAN Inc., 10016 sok.No. 16, A.O.S.B, Cigli, 35620 Izmir, Turkey

Correspondence should be addressed to Muhammed G. Cinsdikici, [email protected]

Received 19 October 2009; Accepted 2 September 2010


Copyright © 2010 M. G. Cinsdikici and K. Memis. This is an open access article distributed under the Creative CommonsAttribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work isproperly cited.

Daily observed traffic flow can show different characteristics varying with the times of the day. They are caused by traffic incidentssuch as accidents, disabled cars, construction activities and other unusual events. Three different major traffic conditions can beoccurred: “Flow,” “Dense” and “Congested”. Objective of this research is to identify the current traffic condition by examiningthe traffic measurement parameters. The earlier researches have dealt only with speed and volume by ignoring occupancy. Inour study, the occupancy is another important parameter of classification. The previous works have used multiple sensors toclassify traffic condition whereas our work uses only single microwave sensor. We have extended Multiple Linear Regressionclassification with our new approach of Estimating with Error Prediction. We present novel algorithms of Multiclassification withOne-Against-All Method and Multiclassification with Binary Comparison for multiple SVM architecture. Finaly, a non-linearmodel of backpropagation neural network is introduced for classification. This combination has not been reported on previousstudies. Training data are obtained from the Corsim based microscopic traffic simulator TSIS 5.1. All performances are comparedusing this data set. Our methods are currently installed and running at traffic management center of 2.Ring Road in Istanbul.

1. Introduction

Traffic flow characteristic shows dynamical change at dif-ferent time periods of the day. Many traffic incidents suchas accidents, disabled cars, construction activities, highdemands on traffic, and other unusual events cause thischange. For dealing with these unstable traffic problems,traffic conditions should be clarified. Mainly there exist threedifferent major traffic conditions: “Flow Traffic,” “DenseTraffic,” and “Congested Traffic”. Clarification of them req-uires careful/detailed examination of the flow parameters ofspeed, volume, and occupancy (SVO).

Accurate interpretation of SVO supports traffic manage-ment centers to make proper decision on directing the trafficto the less intensive roads. Hence, the response time forintervention in an incident will be reduced.

Measurement methods for obtaining SVO have changedfor the last 60 years of the span (i.e., especially the last 40years with the rapid rise in the number of freeways). Indeed,

they are still changing [1]. Some of them are

(i) measurement at a point,

(ii) measurement over a short section (about 10 meters),

(iii) measurement over a length of road (at least 0.5 km),

(iv) measurement with mobile observer in the trafficstream,

(v) measurement with multiple simultaneous mobilevehicles, as part of ITS (Intelligent TransportationSystems).

Although all measurement methods above produce speedand volume of SVO, the only method producing occupancy(i.e., the percentage of unit time that the detection zone ofthe instrument is occupied by vehicles) is the measurementover a short section [1].

Detectors used for short section measurement can bebased on inductive loop (IL), microwave, radar, photocell,


ultrasonic, and analog/digital camera technologies [1]. Sincethe quick and effortless installation is possible only on ILtechnology, we have used Radio Transmissions MicrowaveSensor-based (RTMS) IL detector. We have collected SVOparameter values at different times of day by using RTMS-IL.

The collected SVO data is analyzed through our threedistinct novel approaches to classify traffic flow as “Flow”,“Dense”, or “Congested”. These approaches are estimatingerror prediction for multiple linear regression analysis, twoimproved variants of support vector machines (SVM), andbackpropagation neural network. First two methods arelinear classifiers, whereas NN is non-linear.

In Section 2, the contributions of our paper will be givenin the related works. Then, background is presented. In theSection 4, all details about the proposed study are going tobe given. Section 5 summarizes experimental results. Finally,conclusion is given.

2. Related Works

Flow theory has been tried to analyze traffic through thespeed, volume, and the vehicular concentration parameters.

Temporal vehicular concentration named as occupancycan be measured only over a short section (i.e., shorter thanthe minimum vehicle length). So, this parameter becomesunmeaning for long section measurements.

Density as an alternative vehicular concentration hasbeen a part of traffic measurement since 1930’s. It depictsthe number of vehicles over a long section (i.e., one mile orkilometer) in contrast to occupancy.

Although vehicular concentration encompasses bothdensity and occupancy parameters, indeed, it would be fairto say that the majority opinion is in favor of using densityduring the evolution of traffic flow theory.

However, a minority view has intended to use occupancyin theoretical works. Although there are well-defined factsput forward by the majority for the continued use of density,the minority also propounds major reasons for making moreuse of occupancy. The most crucial reason among themcan be given as density (i.e., vehicles per length of road)ignores the effects of vehicle length and traffic composition.Occupancy, on the other hand, is directly affected by both ofthese variables and therefore gives a more reliable indicatorof the amount of a road being used by vehicles.

First mathematical model-based on speed, volume, anddensity variables had been developed by Greenshild inmid 1930s [2] using the aerial photographs. In his work,relationship between speed and density is introduced relyingon simple linear regression approach.

After World War II, with the tremendous increase inuse of automobiles and the expansion of the highwaysystem, there was also a surge in the study of trafficcharacteristics and the development of traffic flow theories.In 1950’s, theoretical developments based on a varietyof approaches, such as car-following, traffic wave theory(hydrodynamic analogy), and queuing theory has emerged.Some of the seminal works of that period include the worksby Reuschel (1950) [3–5], Wardrop (1952) [6], Pipes (1953)[7], Lighthill and Whitham (1955) [8], Newell (1955) [9],

Webster (1957) [10], Edie and Foote (1958) [11], Chan-dler et al. (1958) [12], and other papers by Herman et al.

Reuschel and Pipes offered a microscopic traffic modelthat identifies the linear dependency between the speed ofa vehicle and the distance between the vehicles in a singlelane. The models described by Reuschel and Pipes werereasonable in concept, but no experimental verification oftheir conclusions was pursued for many years [13].

Wardrop’s theory was based on two major principles.The first one was stating that travel times between the sameorigin and destination pairs for any used routes are lessthan or equal to the travel times for all unused routes. Thisis referenced as Dynamic User Equilibrium (DUE) in theliterature and used for large-scale networks. Diverting trafficwith DUE is inefficient and difficult-to-implement systemoptimum (i.e., min. average travel time) [14]. The secondone aims at minimizing both the total travel times andaverage travel times for all assigned routes for all drivers onthe whole network. However, individual choice of drivers wasby no means guaranteed to satisfy this principle and in mostcases did not [13].

A few years later, Lighthill and Whitham (L-W) togetherset out the first comprehensive theory of kinematic waves.In L-W traffic model, there exists correlation betweentraffic flow (cars/hour) and traffic density (cars/mile) [8].Propagation of shock waves, generated by traffic transitionsfrom one steady state to another, was also determined byL-W model. Nevertheless, this model was viable only fordescribing the density changes. Unfortunately, the modelproduces larger densities exceeding the possible maximumvehicle density.

Greenberg improved the existing mathematical modelsby adding the nonlinearity into the model structure [15]. Hetreats the traffic stream as a continuous fluid and derives therelations between speed, density, and flow by fluid dynamics.

The distinction of free (i.e., non-congested) and con-gested flow on the speed-density model was carried out byEddie [16] and Underwood [17] and was investigated in animportant empirical test by Drake et al. [18].

Athol [19] suggested using the occupancy rather thanthe density for the flow-concentration works, however hissuggestion became popular one decade later.

Speed, flow, and occupancy (SFO) relationships havebeen studied at the free flow level by Hurdle and Datta[20], Persaud and Hurdle [21], F. L. Hall and L. M. Hall[22], Banks [23], Smith et al. [24], Hall et al. [25], andWemple et al. [26]. Also, SFO have been studied at congestedflow level and by Hall and Montgomery [27], Zhou and Hall[28], and Banks [29, 30].

SVMs are usually employed for incident detection algo-rithms identifying the anomalies of traffic flow model usingsingle or two sensors [31]. The process for determiningthe presence of an incident is twofold. The first is adetermination of congestion (exits/not-exists), one of thestates we also want to identify. Next is the binary analysisof type of congestion (incident occurred/not-occurred). Pastresearches and applications usually tend to use the trafficdescriptors of two traffic sensors as their input parametersfor incident detection with SVM methods.


In ITS, neural networks can be found on the areas likevehicle detection, road detection, and single loop vehicle typeclassification [32, 33]. Rarely some work related with trafficflow control can be found and they are also not well defined[34].

All retrospective researches so far show the binaryrelationships between the traffic parameters such as flow rateand occupancy (FR-O), flow rate and speed (FR-S), speedand occupancy (S-O) at congested or flow levels. Trafficstatus is assessed according to these binary relationships orby inspecting only one parameter (like speed) rather thanexamining all of them.

The first contribution of this paper to the traffic studiesis focusing on another major traffic flow level ignored inearlier researches called dense flow level. All the previousworks studies congested and flow traffic levels.

The second contribution of the paper is that no one inthe earlier works is based on “single sensor” (i.e., all of themused double sensor placed within some distance) with usingSVO triplets (i.e., triple relations of traffic parameters).

The third one is our new approach of “estimation witherror prediction” for multiple-linear regression.

The fourth and most important one is our novelalgorithms of “Multiclassification with One-Against-All” and“Multiclassification with Binary Comparison” as variants forsupport vector machine (SVM) classifier for the traffic flowmodel.

The final contribution is that none of the earlier worksuses neural network model to classify the traffic control byshort section with single ILD.

3. Background

One of the short section detectors, RTMS, is capable ofproducing some kind of traffic parameters periodically foreach lane on the freeway. These parameters are volume,speed, and occupancy.

(i) Volume shows the count of total vehicles passedthrough this short section for one period.

(ii) Speed shows the average speed of total cars passedthrough this short section for one period.

(iii) Occupancy shows the sum of the time; vehiclesoccupy the short section divided by one period time.

3.1. Linear Multiple Regression Analysis Method as a Classifier.The objective of linear multiple regression analysis (LMRA)is to define which of the independent variables are importanton predicting the model [35]. Multiple regression analysisprovides a predictive equation

Y = a + β1x1 + β2x2 + · · · + βnxn, (1)

where, a is interception constant, βi (i=1, 2, ...,n) are standard-ized partial regression coefficients (reflecting the relativeimpact on the criterion variable). xi (i=1, 2, ...,n) are the metricscores (i.e., interval or ratio data) of different independentvariables. Y is the single dependent variable structuring the

X2

X1

Optimal hyperplane

Maximum margin

Figure 1: Maximum margin of optimal hyperplane.

model. Equation (1) actually symbolizes a linear hyper plane.The purpose of the LMRA classifier is to minimize the energyfunction E for each data point, defined by LMS as in (2). Theparameters a and βi (i=1, 2,...,n) are obtained from solving thepartial derivatives of

E =∑(

yp −(a + β1x

p1 + β2x

p2 + · · · + βnx

pn

))2. (2)

for given data.

3.2. Support Vector Machine as a Classifier. Another inno-vative supervised pattern classifier technique SVM was firstproposed by Vapnik in 1995 [36]. The formulation appliedby SVM embodies the Structural Risk Minimization (SRM)principle, which has been shown to be superior to traditionalEmpirical Risk Minimization (ERM) principle [37]. WhileSRM minimizes an upper bound on the expected risk,ERM minimizes the error on the training data [38]. It isthe difference which equips SVM with a greater ability togeneralize, which is the goal in statistical learning. SVMs weredeveloped to solve the classification problem (i.e., data mustbelong to either Class 1(+1) or Class 2(−1)) but recently theyhave been extended to the domain of regression problems[39].

The decision function in (3) determines the classes of allinput vectors (x = xim) [where i is the input and m is thetuple indexes]. φ is fixed feature-space transformation, W ism dimensional weights, and b is bias term [40]:

D(x) =WTφ(x) + b. (3)

If there are multiple solutions, we should find thesmallest generalization error. So, “margin” (i.e., smallestdistance between decision boundary and any of the samples)should be found through [controlling the separability] [41]:

min∣∣∣WTx + b

∣∣∣ = 1. (4)

Although there are infinite number of solutions toseparate hyperplanes, by maximizing the margin, Figure 1shows only two decision functions satisfying (4).


StructureDescription of

decision regionsExclusive or

problemClasses with

meshed regionsGeneral region

shapesMost generalregion shapes

Single layer

Two-layer

Three-layer

Half planebounded byhyperplane

Arbitrary(complexitylimited by #of hidden

units)


units)

StructureDescription of

decision regionsExclusive or

problemClasses with

meshed regionsGeneral region

shapesMost generalregion shapes

Single layaa er

TwTT o-layaa er

Three-layaa er

Half planebounded byhyhh perplane


units)


units)

Figure 2: Classification interpretation of ANN model selecyion [51].

The margin is given by the perpendicular distance to theclosest point xi from the data set and we wish to optimizethe parameters W and b in order to maximize this distance(using target ti). Thus the maximum margin solution isfound by solving

argmaxW ,b

{1

‖W‖}

mini

[ti(WTφ(x) + b

)]. (5)

The dominating approach for solving multiclass prob-lems using SVM has been based on reducing single multiclass problem into multiple binary problems. For instance,a common method is to build a set of binary classifierswhere each classifier distinguishes between one of the labelsto the rest. This approach is a special case of using outputcodes for solving multi class problems [42]. However, whilemulti class learning using output codes provides a simple andpowerful framework it cannot capture correlations betweenthe different classes since it breaks a multi class problem intomultiple independent binary problems [43].

The idea of casting multi class problems as a singleconstrained optimization with a quadratic objective functionwas introduced by Vapnik [44] and Watkins [45]. Theseattempts to extend the binary case are achieved by addingconstraints for every class and thus the size of the quadraticoptimization is proportional to the number of categories inthe classification problems. The result is often a homoge-neous quadratic problem which is hard to solve and difficultto store. The idea of breaking constrained optimization prob-lem into smaller problems was extended by Joachims [46].Multiple class problems were then discussed by Scholkopf[47]. Today’s panorama of SVM is well summarized byCristianini [48] and more completely by Scholkopf andSmola [49].

3.3. Backpropagation Neural Network as a Classifier. Artificialneural networks (ANN) are used to serve two importantfunctions as pattern classifiers and as nonlinear adaptive

V3,4

W4,2

W1,1

V1,1Input

Hidden

X1

X2

X3

Z1

Z2

Z3

Z4

Y1

Y1

Output

Figure 3: Simple Three-layered Neural Network Structure (3-4-2).

filters. Figure 2 gives a brief overview about the ANNarchitecture for proper pattern classification [50, 51]. Sincethe data is not guaranteed to be linearly separable (i.e.,overlapped), three-layered Backpropagation NN (BPNN)architecture is chosen in this paper as classifier. This typeof architecture does not need any prior knowledge aboutdata (i.e., exemplar pattern initialization). Since there is noneed for distribution of the data (for fault-tolerance), asyn-chronous update in weights and self-classification, BPNN issuitable as supervised classifier [32, 50]. In Figure 3, simplethree-layered BPNN is indicated.

In our paper work, we used alternative learning methodsfor comparing the performances. They are Gradient Descentwith Momentum (gdm), Conjugate Gradient Descent (cgd),Scaled Conjugate Gradient (scg), and Levenberg-MarquardtGradient Descent with momentum and Scaled ConjugateGradient. In [52], you can find algorithms and related details


0

10

20

0246810120

20

40

60

80

100

120

Volu

me

Speed

Occ

upa

ncy

Congested traffic

Figure 4: An example simulator data for Congested traffic.

020

40

60

80

100

0

50

100

0

5

10

15

20

25

Volume

Flow traffic

Speed

Occ

upa

ncy

Figure 5: An example simulator data for Flow traffic.

about the formulation and theory about these learningmethods.

4. Proposed Method

4.1. Data. For each traffic condition, different values andtypes of data can be acquired through the FHWA’s (FederalHighway Agency) TSIS 5.1 software (i.e., a CORSIM-basedmicroscopic traffic simulation tool). Differentiation of datais supplied by configuring simulation tool parameters suchas distribution function of generated traffic, lane speeds, cardistances, and incident creation intervals.

20

30

40

50

2025303540455055

10

15

20

25

30

35

40

45

50

Volu

me

Dense traffic

SpeedO

ccu

pan

cy

Figure 6: An example simulator data for Dense traffic.

Each ILD sensors generates speed, volume, and occupancyparameters values for configurable one period of time (i.e.,60 seconds) for each lane on the freeway.

4.1.1. Congested Condition. Congested traffic condition canbe observed by creating long-term incidents along thefreeway. According to the number of lanes they occupy,incidents can be assorted. In our study, kinds of incidentscenarios are created for four-lane freeway. The sensors areplaced 100 feet upstream from the incident point. As seenin Figure 4 occupancy reaches the maximum values, whereasboth speed and volume reach the minimum values.

4.1.2. Flow Condition. In order to generate the flow trafficcondition, CORSIM’s average speed input is determinedfrom high values (i.e., 90 km/h) and no incidents are created.Figure 5 illustrates the flow condition. It is clear from thefigure that, when the speed is high and volume is low, theoccupancy approximately reaches to zero. Under the lowspeed and high volume conditions, occupancy rises but neverreaches to the maximum value.

4.1.3. Dense Condition. Dense traffic condition can be iden-tified by generating low speed values, short-term incidents,and decreased distances between the cars in traffic. Thiscondition is observed between the flow and congestedconditions. As seen in Figure 6 the occupancy never reachesto min or max values.

4.2. Multiple Regression Analysis. There exist two significantfacts through the basis of this method. The prior one saysit is obvious that occupancy approximates to zero whetherthe average speed of vehicles approximates to infinity in


0

50100

150020

4060

80

0

20

40

60

80

100

Speed

98.2–8.5x-2.5y

Volume

Occ

upa

ncy

Figure 7: Regression plane (red) for Congested traffic.

flow traffic condition or not. On the contrary, occupancygoes to maximum value when speed approximates to zero incongested traffic flow condition.

It is also obvious that, until the volume reaches itsmaximum, occupancy is also increasing. After the maximum,the volume is going to be monotonically decreasing whereasoccupancy is still increasing.

Under these circumstances, in flow traffic condition morevolume leads to more occupancy. However, in congestedtraffic flow, the less volume observed leads to more occupancy.

According to the facts mentioned above, occupancy isdependent both on speed and volume.

Occupancy = a + β1 ∗ Volume + β2 ∗ Speed + Error. (6)

Equation (6) is not viable in real world affairs. Oncethe traffic seems to be in congested, both speed, andvolume can show the zero, although occupancy approachesto maximum. Unless the vehicles exist along the freeway(i.e., flow condition), speed, volume, and also the occupancyindicate zero. A linear dependency of SVO for each trafficcondition with respect to the discrepancy are modified from(6) as in (7). For each flow condition a, β1, and β2 parametersare different,

Occupancycondition = a + β1 ∗ Volume + β2 ∗ Speed + Error.(7)

4.2.1. Regression Planes. Using the CORSIM data, α, β1, andβ2 values for each traffic condition are calculated using LMS(Least Mean Squares) method. The obtained results can betracked through following subtitles.

0 10 20 30 40 50 600

50

100

1500

10203040506070

70

8090

100

Volume

0.05 + 0.3x − 0.004y

Speed

Occ

upa

ncy

Figure 8: Regression plane for Flow traffic.

(a) Congested condition. The plane evaluated for congestedtraffic is depicted in Figure 7 and its equation is in

Occupancycongested = 98, 02− 8, 5∗ Volume− 2, 5∗ Speed.

(8)

Through the crosschecking (8), it can easily be seen thatoccupancy reaches max (98.2) when the speed and volumeare marked as zero. This verifies the expectations of the realtraffic condition.

(b) Flow Condition. Figure 8 shows the plane obtained fromthe flow traffic. Its equation is defined as

Occupancyflow = 0, 05 + 0, 3∗ Volume− 0, 004∗ Speed. (9)

In real traffic, if no car is detected on the freeway duringthe time period, detectors will produce the occupancy as 0whereas speed and volume also share the same value. Throughthe crosschecking (9), it can easily be seen that occupancyreaches min (0.05) when the speed and volume are markedas zero. This verifies the expectations of the real trafficcondition.

(c) Dense Condition. Figure 9 depicts the dense traffic flowcondition’s regression plane. Linear equation of the plane canbe defined as

Occupancydense = 34.5 + 0, 58∗ Volume− 0, 7∗ Speed.(10)

(d) Combined Data and Regression Planes. Data and regres-sion planes of all traffic conditions are shown in Figure 10.

Three different methods are applied, respectively, forclassification of traffic conditions using linear regression.

(1) Using Analytic geometry (AG). From the CORSIMsimulation, regression planes are defined for each trafficconditions. AG calculates the distances between all data


0

20

40

60

050

100150

Volume

34.5 + 0.58x − 0.7y

Speed

Occ

upa

ncy

0

10

20

30

40

50

60

70

80

90

100

Figure 9: Regression plane for Dense traffic.

points and each plane. The result (i.e., nearest plane) suppliesus to infer the class (status) of current traffic.

From (8), we can obtain (11) for congested traffic as

0 = 98, 02+8, 5∗Volume−2, 5∗Speed−Occupancycongested,

dcongested

=∣∣98.02−8, 5∗Volume−2, 5∗Speed−Occupancy

∣∣√8, 52 +2, 52 +12

.

(11)

From (9), we can obtain (12) for flow traffic as

0 = 0, 05+0, 3∗Volume−0, 004∗Speed−Occupancyflow;

dflow

=∣∣0, 05−0, 3∗Volume−0, 004∗Speed−Occupancy

∣∣√0, 32 +0, 0042 +12

.

(12)

From (10), we can obtain (13) for dense traffic as

0 = 34.5+0, 58∗Volume−0, 7∗Speed−Occupancydense,

ddense

=∣∣34.5−0, 58∗Volume−0, 7∗Speed−Occupancy

∣∣√0, 582 + 0, 72 + 12

.

(13)

(2) Estimating Occupancy without Error Prediction (EO). InEO, linear regression (6) is used and error is assumed as zero.Using this assumption occupancy estimator can be calculatedfor each simulated traffic condition (8), (9), (10) using thereal speed and real volume. After acquiring these estimators,they are compared with the real occupancy. The one nearest

010

2030

4050

6070

0

50

100

150

0

20

40

60

80

100

SpeedVolume

Occ

upa

ncy

Figure 10: Composite regression planes of all traffic.

to real occupancy becomes the result of our traffic flowcondition.

(3) Estimating Occupancy with Error Prediction (EOwEP):Another method for finding the traffic flow class is extendedversion of EO. This method calculates occupancy estimatorsby adding the predicted error values.

EOwEP depends on the training data values and theirresiduals. Nearest point is found by measuring L2 distancebetween real traffic data point and CORSIM simulated datapoints from each traffic condition. The residual of the foundpoint is assumed to approximate our error. As a result,occupancy estimator for that condition can be determined.So the obtained equation as

Occupancycongested = 98, 02 + 8, 5∗ Volume− 2, 5

∗ Speed + resi,(14)

where i is the nearest point, found by (15), to the real datapoint among the congested training data points

Min[

(Voli − Vol)2 +(Speedi − Speed

)2 + (Occi −Occ)2].

(15)

4.3. Support Vector Machines. As mentioned in background,SVM is generally applied to binary classification problems.Since the addressed problem is mapping the input data intoone of three flow conditions, SVM can be seen inapplicableat first. However, without changing its calculation style, SVMcan also be used in multiclassification problems. The idea laysbehind multi classification is just using more than one SVMand classifing the data according to the outputs of multipleSVMs. We use our two novel approaches for Multiple SVM.

Multiclassification with One-Against-All Method (OAAM).The designated architecture (our novel approach) for thismethod is composed of three SVMs. Each SVM representsone of the traffic flow conditions. Detector values canbe fed into each SVM, respectively. The bipolar output


BPNN Algorithm;(i) Initialize weights (vi, j , w j,k) with small random values(ii) Broadcast the input data to the input layer xi

Feed-forward phase is starting here;(i) Calculate the hidden layer unit signals

(“f ” is sigmoid function in our work)

zj in =n∑i=1xivi j

z j = f (zj in) = 21 + e−z j in

− 1

(ii) Calculate the output unit signal

yk in =p∑j=1

zjwjk

yk = f (yk in) = 21 + e−yk in

− 1

Backpropagation phase is starting here;(i) Calculate the residual. By using expected value for output signal (tk).

δk = (tk − yk)∗ f ′(yk in) = (tk − yk)∗(

12

[1 + f (yki n)][1− f (yk in)])

Δwjk = αδkzjwhere α is learning rate. In the learning phase, Δwjk is updated according to learning rule.

(ii) Then calculate reflectance of the residuals and propagate it to the input weights.

δj =m∑k=1

δkwjk ∗ f ′(zj in)

Δvi j = αδjxi(iii) All weights are updated with learning rule (i.e., gdm/scg).

wjk(new) = wjk(old) + Δwjk

vi j(new) = vi j(old) + Δvi j(iv) Test the stopping condition (i.e., reaching to “goal”; predefined Mse —mean square error- of the total residuals)

Algorithm 1

[“+1” (Class 1), “−1” (Class 2)] indicates if the input databelongs to this SVM or not.

For modeling OAAM, data is acquired from TSIS 5.1.Decision is made through. “KKT (Karush-Khun-Tuckermethod for minimizing Quadratic Problem solution forSVM) [45] ”.

min{

12‖W‖2

}. (16)

(a) Congested SVM (C-SVM). In the training, C-SVM targetis trained with “+1” whereas the other SVM targets aretrained with “−1”. Plane for C-SVM can be shown inFigure 11. Its equation is

D(vol, spd, occ

) = −0.1235∗ vol−0.0443∗ spd + 0.1052

∗occ− 0.9996.(17)

(b) Dense SVM (D-SVM). In the training, D-SVM target istrained with “+1” whereas the other SVM targets are trainedwith “−1”. Plane for D-SVM can be shown in Figure 12. Itsequation is

D(vol, spd, occ

) = 2.6577∗ vol− 1.7462∗ spd− 0.2456

∗ occ− 0.9998.(18)

(c) Flow SVM (F-SVM). In the training, F-SVM target istrained with “+1” whereas the other SVM targets are trainedwith “−1”. Plane for F-SVM can be shown in Figure 13. Itsequation is in (19);

D(vol, spd, occ

) = 0.0105∗ vol + 0.0788∗ spd − 0.3218

∗occ + 1.0003.(19)

After training is done, each member of Multiple-SVMshas its own linear planes for OAAM. In order to get thecorrect class of the queried data is just easy as to look atthe responses of each SVMs. The “+1” response of the SVMclarifies the class of the data. For instance if F-SVM’s responseis “+1” (i.e., the others are expected as “−1”) then the classof the data belongs to flow traffic.

However, this is not sufficient alone. Following casesmust be taken into account and supplemented approachesmust be applied in these conditions.

(i) If at least two SVMs produce output “+1”, then theone with the longest L2 distance from data pointto its own SVM plane can be picked out. Since thefurthest distance from the data point to plane exposesthe stricter and more accurate data for current trafficflow condition, we choose the furthest one.

(ii) If all SVMs produce output “−1”, the one with theshortest L2 distance to SVM plane must be chosen.


010

2030

4050

6070 0

50

100

150

0

20

40

60

80

100

Speed

9.5019 + 1.1739x + 0.4211y

Volume

Occ

upa

ncy

Figure 11: Congested Traffic Plane for C-SVM of OAAM.

0 10 20 30 40 50 60 70

050

1001500

102030405060708090

100

Volume

−4.0708 + 10.8212x − 7.1099y

Speed

Occ

upa

ncy

Figure 12: Dense Traffic Plane for D-SVM of OAAM.

Producing output “–1” implies that the data does notbelong to the current SVM’s traffic condition. Since,the shortest distance from data point to SVM planeis the nearest one to “+1” (Class 1), we can pick thisSVM out as our desired class.

Multiclassification with Binary Comparison (BC). Our othernovel approach applied to classify our data is again composedof three SVMs. However, it can be distinguished from theOAAM method by its architecture. In BC method, eachSVM is responsible for the following binary combinations,respectively; (Congested, Dense), (Congested, Flow), and(Flow, Dense). This multi classification method dependson basic voting principle. There exist three candidates (i.e.,Congested, Flow, and Dense traffic conditions) and threevoters (each SVM). The one among the candidates wins theelection if it gets the most votes from voters.

020

4060

0501001500

10

20

30

40

50

60

70

80

90

100

Volum

e

Speed

3.1084 + 0.0326xx + 0.2448y

Occ

upa

ncy

Figure 13: Flow Traffic Plane for F-SVM of OAAM.

(a) Congested-Dense (CD-SVM). In the training, CD-SVMtarget “+1” indicates that data belongs to Class 1 (i.e.,congested traffic). The other alternative of “−1” indicatesthat the data belongs to Class 2 (i.e., dense traffic).

So the voter gives its vote to Congested candidate if itsoutput equals “+1”. In adverse condition, it gives to Densecandidate. Figure 14 shows the decision plane for CD-SVM.The plane equation is

D(vol, spd, occ

) = −0.0642∗ vol− 0.0440∗ spd− 0.0041

∗ occ + 1.8692.(20)

(b) Congested-Flow SVM (CF-SVM). In the training, CF-SVM target “+1” indicates that data belongs to congestedtraffic. The other alternative of “−1” indicates that the databelongs to flow traffic. Figure 15 shows the decision plane forCF-SVM. The plane equation is

D(vol, spd, occ

) = −0.0044∗ vol− 0.0160∗ spd + 0.0711

∗ occ−0.9998.(21)

(c) Flow-Dense SVM (FD-SVM). In the training, FD-SVMtarget “+1” indicates that data belongs to flow traffic. Theother alternative of “−1” indicates that the data belongs todense traffic. Figure 16 shows the decision plane for FD-SVM. The plane equation is

D(vol, spd, occ

)=0.0105∗ vol + 0.0788∗ spd− 0.3218

∗ occ +1.0003.(22)

A supplemented way must be found for handling thesituation when all candidates have the equal vote. Using thelongest distance from data point to each SVM’s plane will


020

4060

0 50 100 150

0

10

20

30

40

50

60

70

80

90

100

Speed

455.9024 − 15.6585x − 10.7317y

Volume

Occ

upa

ncy

Figure 14: Plane of CD-SVM for BC.

0

500

501001500

10

20

30

40

50

60

70

80

90

100

Volume

Speed

14.0618 + 0.0618x + 0.2250y

Occ

upa

ncy

Figure 15: Plane of CF-SVM for BC.

give us the most accurate class for that data point. Then theSVM with the longest distance will choose the leader class.

Backpropagation Neural Network Architectures . The networkarchitecture in our work is selected as three-layered back-propagation neural network (BPNN). three groups of dataare used as input. The first group is fed into the network astraining data set. The next one is used as verification dataset. The last one is used as query set (i.e., test set) to measurethe classification performance of the neural net.Training dataset is obtained from TSIS 5.1 simulator. This data is verifiedthrough the verification set also gathered from the simulator.The test set is gathered from single sensor planted on the2.Ring road in Istanbul. Test set indicates real data. Ourmodel is trained and verified through the simulated data andtested with real traffic information.In the learning phase wehave used gradient descent with momentum (Gdm) update

50050100150

0

10

20

30

40

50

60

70

80

90

100

VolumeSpeed

3.1084 + 0.0326xx + 0.2448y

Occ

upa

ncy

Figure 16: Plane of FD-SVM for BC.

function in (23) (i.e., learning rule). This base rule is used forall architectures

wjk(t + 1) = wjk(t) + αδkzj + μ[wjk(t)−wjk(t − 1)

],

vi j(t + 1) = vi j(t) + αδjxi + μ[vi j(t)− vi j(t − 1)

].

(23)

In the learning phase, we have used the algorithms of onlyGdm, Levenberg-Maquardt with Gdm and Scaled ConjugateGradient with Gdm [52]. Matlab 7.6.0 (Rev.2008a) NeuralNetwork Toolbox where our related work algorithms andlearning rules are ready is used.

We have used default max iteration as 1000, Mse goalvalue as 10−4, and verification iteration count as 6. Thenetwork architecture is performed as three-layered modelsof 3-10-3, 3-20-3, and 3-50-3. Since one hidden layer usageis almost identical with two hidden layered architecture, weprefer to use single hidden layer model. For the connec-tions lying between Input/Hidden Layers are trained withLogSig or PureLin activation functions. LogSig activationfunction normalized the input and applies Sigmoid. PureLinear activation function transfers input to hidden.We givethe performance analysis of our neural net architecturesin Table 1. Gray labeled values are average performancesobtained from 100 runs of neural net models. Accordingto the average performances, our 3-20-3 backpropagationmodel gives the best result for classification.

5. Experimental Work

Turkey General Directorate of Highways has carried outworks in managing traffic in Istanbul hence traffic manage-ment center is planned to be opened at the end of December2007. The system is installed by Turkey’s leading electronicscompany ASELSAN Inc. In Istanbul traffic managementsystem, microwave radar type detectors are used for occu-pancy, speed, and traffic volume measurements. There existalmost 30 radio transmission microwave sensors along the


Table 1: Backpropagation Neural Net Performances.

GDM LM SCG

LogSig-PurLin PurLin-PurLin LogSig-PurLin PurLin-PurLin LogSig-PurLin PurLin-PurLin

BackPrp 3-10-3

Cong 100 100 0.9604 0.9472 100 0.4073

Dense 0.5733 0.5612 0.5134 0.5671 0 0

Flow 0.9502 0.9189 0.9210 0.9470 0 0.062

Total (Rate/Iteration) 0.9678/1000 0.9258/1000 0.9352/360 0.9170/5 0.0068/156 0.0615/15

BackPrp 3-20-3

Cong 0.6227 0.9472 100 0.9393 100 0.9393

Dense 0.6914 0.6878 0.4677 0.6778 0.8531 0.6725

Flow 0.9804 0.8950 0.7492 0.8967 0.9914 0.9468


BackPrp 3-50-3

Cong 0.9721 0.9393 0.9732 0.9893 0.9789 0.9472

Dense 0.4552 0.6686 0.1256 0.6800 0.6318 0.6711

Flow 0.9571 0.9869 0.9899 0.9907 0.9879 0.9867


Table 2: Estimation Performance of Our Model for Each Trafficcondition.

Method Flow Dense Congested

Analytic Geometry 99% 84% 70%

Without Error Prediction 100% 78% 35%

Error Prediction 99% 24% 46%

SVM with One-Against-All Method 100% 46% 97%

SVM with Binary Comparison 100% 74% 100%

BackPrp NN-SCG (3-20-3) 99.14% 85.31% 100%

BackPrp NN-LM (3-50-3) 99.07% 68% 98.93%

BackPrp NN-GDM (3-10-3) 95.02% 57.33% 100%

2.Ring Road. The measurements recorded at critical pointson the road are transmitted periodically to the managementcenter [53].

Occupancy, speed, and traffic volume measurementsare both recorded in the database, and the traffic statusaccording to these measurements is mapped to color codesand displayed on a large screen. Whenever any congestedstate or dense state occurs along the road, alarms withseverity levels are generated and the operators are informed.Information regarding the traffic status is also displayed onLED-based Variable Message Signs (VMS) on the road.

During one day, data from 7 different microwave sensorsis acquired every 60 seconds periodically. Occupancy, volumeand speed measurements are shown in Figure 17. Videosfrom the traffic surveillance cameras, which are capable ofwatching the places that microwave sensors are installed,are also captured in order to synchronize them with thesensors measurements. Then, measurements are grouped ascongested, dense and flow according to these videos. [Wehave used training data set of total size 18.000 x 3 (i.e., eachtraffic condition has balanced data subset size of 6000 x 3)and we have used testing data set of total size 55.000 x 3, and3 indicates SVO values]. These data sets are used to for theTable 2.

0

100

200

0 5 10 15 20 25 30 35 40 45 50

0

10

20

30

40

50

60

70

80

90

100

Speed

Real traffic (2. Ring Road Istanbul)

Volum

e

Occ

upa

ncy

Figure 17: Real Traffic data from 2.Ring Road in Istanbul (by 7sensors).

6. Conclusion

Eight methods discussed above have applied the one day datacaptured from 7 sensors on 2.Ring Road in Istanbul. Theirperformance for each traffic condition can be seen in Table 2.

Although our novel approaches OAAM and BC for Mul-tiple SVM have good performances, our BPNN architectureof 3-20-3 with SCG (3-20-3) is better than all other sevenmethods. As a result of this comparison, this method haschosen and installed for the evaluation of traffic condition inIstanbul traffic management center (at Istanbul 2.Ring road).

According to the result it calculates, our algorithm colorsthe road map for each status changes on GIS projection


screens. Once congested level is detected, it produces analarm with high severity to inform the operators. Never-theless, dense level also creates alarm with lower severity.After verification of this assessment by operator, VMS can besupplied by messages to inform drivers about the road flowcondition.

References

[1] F. L. Hall, “Traffic stream characteristics,” in USA Transporta-tion Research Board. Traffic Flow Theory, A State of the ArtReport, 1992.

[2] B. D. Greenshields, “A study of traffic capacity,” HighwayResearch Board Proceedings, vol. 14, pp. 448–477, 1935.

[3] A. Reuschel, “Fahrzeugbewegungen in der Kolonne,” Oesterre-ichisches Ingenieur, vol. 4, no. 3/4, pp. 193–215, 1950a.

[4] A. Reuschel, “Fahrzeugbewegungen in der Kolonne bei gle-ichformig beschleunigtem oder verzogertem, Leitfahrzeug,”Zeitschrift des Oesterreichischen Ingenieurund Architekten-Vereines, vol. 95, no. 9, pp. 59–62, 1950.

[5] A. Reuschel, “Fahrzeugbewegungen in der Kolonne bei gle-ichformig beschleunigtem oder verzogertem, Leitfahrzeug,”Zeitschrift des Oesterreichischen Ingenieurund Architekten-Vereines, vol. 97-98, no. 10, pp. 73–77, 1950.

[6] J. G. Wardrop, “Some theoretical aspects of road trafficresearch,” Proceedings Institute of Civil Engineers, Part II, vol.1, pp. 325–378, 1952.

[7] L. A. Pipes, “An operational analysis of traffic dynamics,”Journal of Applied Physics, vol. 24, no. 3, pp. 274–281, 1953.

[8] M. J. Lighthill and G. B. Whitham, “On kinematic waves. II.A theory of traffic flow on long crowded roads,” Proceedings ofthe Royal Society of London, pp. 317–345, 1955.

[9] G. F. Newell, “Mathematical models for freely-flowing high-way traffic,” Journal of the Operations Research Society ofAmerica, vol. 3, pp. 176–186, 1955.

[10] F. V. Webster, “Traffic signal settings,” Road Research Techni-cal, Road Research Laboratory, London, UK, 1958.

[11] L. C. Edie and R. S. Foote, “Traffic flow in tunnels,” Proceedingsof Highway Research Board, vol. 37, pp. 334–344, 1958.

[12] R. E. Chandler, R. Herman, and E. W. Montroll, “Trafficdynamics: studies in car following,” Operations Research, vol.6, pp. 165–184, 1958.

[13] D. C. Gazis, “The origins of traffic theory,” OperationsResearch, vol. 50, no. 1, pp. 69–77, 2002.

[14] J. R. Correa, A. S. Schulz, and N. E. Stier-Moses, “Selfishrouting in capacitated networks,” Mathematics of OperationsResearch, vol. 29, no. 4, pp. 961–976, 2004.

[15] H. Greenberg, “An analysis of traffic flow,” OperationsResearch, vol. 7, pp. 79–85, 1959.

[16] L. C. Edie, “Car following and steady-state for non-congestedtraffic,” Operations Research, vol. 9, pp. 66–76, 1961.

[17] R. T. Underwood, “Speed, volume, and density relationships:quality and theory of traffic flow,” Yale Bureau of HighwayTraffic, pp. 141–188, 1961.

[18] J. S. Drake, L. Schofer, and A. D. May, “A statistical analysis ofspeed density hypotheses,” Highway Research Record, vol. 154,pp. 53–87, 1967.

[19] P. Athol, “Interdependence of certain operational characteris-tics within a moving traffic stream,” Highway Research Record,vol. 72, pp. 58–87, 1965.

[20] V. F. Hurdle and P. K. Datta, “Speeds and flows on anurban freeway: some measurements and a hypothesis,” inTransportation Research Record 905, pp. 127–137, TRB, NRC,Washington, DC, USA, 1983.

[21] B. N. Persaud and V. F. Hurdle, “Some new data thatchallenge some old ideas about speed-flow relationships,” inTransportation Research Record, vol. 1194, pp. 191–198, TRB,NRC, Washington, DC, 1988.

[22] F. L. Hall and L. M. Hall, “Capacity and speed flow analysis ofthe Queue in Ontario,” in Transportation Research Record, vol.1287, pp. 108–118, TRB, NRC, Washington, DC, USA, 1990.

[23] J. H. Banks, “Freeway speed-flow concentration relation-ships: more evidence and interpretations,” in TransportationResearch Record, vol. 1225, pp. 53–60, TRB, NRC, Washington,DC, USA, 1989.

[24] W. S. Smith, F. L. Hall, and F. O. Montgomery, “Comparing thespeed-flow relationship for motorways with new data from theM6,” Transportation Research Part A, vol. 30, no. 2, pp. 89–101,1996.

[25] F. L. Hall, V. F. Hurdle, and J. H. Banks, “Synthesis of recentwork on the nature of speed-flow and flow-occupancy (ordensity) relationships on freeways,” in Transportation ResearchRecord, vol. 1365, pp. 12–18, TRB, NRC, Washington, DC,USA, 1992.

[26] E. A. Wemple, A. M. Morris, and A. D. May, “Freeway capacityand level of service concepts,” in Proceedings of InternationalSymposium on Highway Capacity, pp. 439–455, Karlsruhe,Germany, 1991.

[27] F. L. Hall and F. O. Montgomery, “Investigation of analternative interpretation of the speed-flow relationship forU.K. motorways,” Traffic Engineering and Control, vol. 34, no.9, pp. 420–425, 1993.

[28] M. Zhou and F. L. Hall, “Investigation of speed-flow relation-ship under congested conditions on a freeway,” TransportationResearch Record, no. 1678, pp. 64–72, 1999.

[29] J. H. Banks, “Investigation of some characteristics of congestedflow,” Transportation Research Record, no. 1678, pp. 128–134,1999.

[30] J. H. Banks, “Review of empirical research on congestedfreeway flow,” Transportation Research Record, no. 1802, pp.225–232, 2002.

[31] F. Yuan and R. L. Cheu, “Incident detection using supportvector machines,” Transportation Research Part C, vol. 11, no.3-4, pp. 309–328, 2003.

[32] C. Oh and S. G. Ritchie, “Recognizing vehicle classificationinformation from blade sensor signature,” Pattern RecognitionLetters, vol. 28, no. 9, pp. 1041–1049, 2007.

[33] R. Mussa, V. Kwigizile, and M. Selekwa, “Probabilistic neuralnetworks application for vehicle classification,” Journal ofTransportation Engineering, vol. 132, no. 4, pp. 293–302, 2006.

[34] E. I. Vlahogianni, J. C. Golias, and M. G. Karlaftis, “Short-term traffic forecasting: overview of objectives and methods,”Transport Reviews, vol. 24, no. 5, pp. 533–557, 2004.

[35] http://www.isical.ac.in/∼ddroy/multistat.doc.[36] V. N. Vapnik, The Nature of Statistical Learning Theory,

Springer, New York, NY, USA, 1995.[37] S. R. Gunn, M. Brown, and K. M. Bossley, “Network

performance assessment for neurofuzzy data modeling,” inIntelligent Data Analysis, vol. 1208 of Lecture Notes in Com-puter Science, pp. 313–323, 1997.

[38] S. R. Gunn, “Support vector machines for classification andregression,” Tech. Rep., Image Speech and Intelligent SystemResearch Group, University of Southampton, Southampton,UK, 1998.


[39] V. Vapnik, S. Golowich, and A. Smola, “Support vectormethod for function approximation, regression estimation,and signal processing,” in Advances in Neural InformationProcessing Systems, vol. 9, pp. 281–287, MIT Press, Cambridge,Mass, USA, 1997.

[40] S. Abe, Support Vector Machines For Pattern Recognition, 2005.

[41] C. Bishop, Pattern Recognition and Machine Learning,Springer, New York, NY, USA, 2007.

[42] T. G. Dietterich and G. Bakiri, “Solving multiclass learningproblems via error correcting output codes,” Journal ofArtificial Intelligence Research, vol. 2, pp. 263–286, 1995.

[43] K. Crammer and Y. Singer, “On the algorithmic implementa-tion of multiClass kernel based vector machines,” Journal ofMachine Learning Research, vol. 2, pp. 265–292, 2001.

[44] V. N. Vapnik, Statistical Learning Theory, John Wiley & Sons,New York, NY, USA, 1998.

[45] J. Weston and C. Watkins, “Support vector machines formulti-class pattern recognition,” in Proceedings of the 7thEuropean Symposium on Artificial Neural Networks, 1999.

[46] T. Joachims, “Making large-scale support vector machinelearning practical,” in Advances in Kernel Methods—SupportVector Learning, B. Scholkopf, C. Burges, and A. Smola, Eds.,MIT Press, 1998.

[47] B. Scholkopf, Statistical Learning and Kernel Methods,Microsoft Research, 2000.

[48] N. Cristianini and J. S. Taylor, An Introduction to SupportVector Machines, and Other Kernel-Based Learning Methods,Cambridge University Press, Cambridge, UK, 2000.

[49] B. Scholkopf and A. Smola, Learning with Kernels, MIT Press,Cambridge, Mass, USA, 2002.

[50] L. Fausett, Fundementals of Neural Networks: Architectures,Algorithms and Applications, Prentice-Hall, Upper SaddleRiver, NJ, USA, 1994.

[51] A. K. Jain, J. Mao, and K. M. Mohiuddin, “Artificial neuralnetworks: a tutorial,” Computer, vol. 29, no. 3, pp. 31–44, 1996.

[52] S. Haykin, Neural Networks: A Comprehensive Foundation,Prentice-Hall, Upper Saddle River, NJ, USA, 1998.

[53] E. Dorken, D. M. Barut, and B. T. Oranc, “Integrated elec-tronic tolling and traffic management on Turkish highways,”in Proceedings of 12th World Congress on Intelligent TransportSystems, p. 2901, San Francisco, Calif, USA, 2005.

eurasip journal

Documents