Research ArticleA Streaming Algorithm for Online Estimation of Temporal andSpatial Extent of Delays

Kittipong Hiriotappa Suttipong ThajchayapongPimwadee Chaovalit and Suporn Pongnumkul

National Electronics and Computer Technology Center (NECTEC) 112 Thailand Science Park Phahonyothin RoadKhlong Nueng Khlong Luang PathumThani 12120 Thailand

Correspondence should be addressed to SuttipongThajchayapong suttipongthajchayapongnectecorth

Received 2 July 2016 Accepted 20 September 2016 Published 10 January 2017

Academic Editor David F Llorca

Copyright copy 2017 Kittipong Hiriotappa 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

Knowing traffic congestion and its impact on travel time in advance is vital for proactive travel planning as well as advanced trafficmanagement This paper proposes a streaming algorithm to estimate temporal and spatial extent of delays online which can bedeployed with roadside sensors First the proposed algorithm uses streaming input from individual sensors to detect a deviationfrom normal traffic patterns referred to as anomalies which is used as an early indication of delay occurrence Then a group ofconsecutive sensors that detect anomalies are used to temporally and spatially estimate extent of delay associated with the detectedanomalies Performance evaluations are conducted using a real-world data set collected by roadside sensors in Bangkok Thailandand the NGSIM data set collected in California USA Using NGSIM data it is shown qualitatively that the proposed algorithm candetect consecutive occurrences of shockwaves and estimate their associated delaysThen using a data set fromThailand it is shownquantitatively that the proposed algorithm can detect and estimate delays associated with both recurring congestion and incident-induced nonrecurring congestion The proposed algorithm also outperforms the previously proposed streaming algorithm

1 Introduction

Traffic congestion has become a major problem in big citiesaround the world causing economic environmental andhealth issues Existing solutions for addressing the trafficcongestion problem are to increase road capacity to usealternative transportation mode (eg mass transit systems)to restrict usage on particular roads at particular times (egcongestion charge) and to use existing roads more efficiently[1] While the first three solutions require extensive buildingof new infrastructure andor studying of public opinionsthe last solution can be more promptly implemented asit mainly involves utilizing traffic information in a moreefficient manner

Intelligent Transport Systems (ITS) have provided effi-cient means of gathering processing and disseminatingrelevant traffic information by utilizing the capabilities ofsensor networks on roadside and onboard vehicles [2] Oneof themost effective and practical ways is to promptly provide

relevant traffic information in advance to enable driversto avoid congested routes and traffic management centersto proactively manage traffic flow This would enable notonly drivers to estimate their travel times but also trafficmanagement centers to estimate the severity of congestionand initiate appropriate strategies to return traffic to normal

Even though numerous traffic estimation and predictionmethods have been proposed in the past decades recentliterature review in [3] indicates that there are still ten chal-lenges that need research attentions One of these challengesis on how to use short-term traffic estimation to reflect amacroscopic point of view of traffic condition which is veryessential for trafficmanagement strategiesWhilemost short-term traffic estimation methods focus on estimating flowtravel time or traffic condition itself few recent studies haveshifted to incorporating new information that reflects furtherinsights on future traffic condition These new informationinclude travel time variability which reflects reliability [4] anddelay associated with anomalies which reflects severity [5ndash7]

HindawiJournal of Advanced TransportationVolume 2017 Article ID 4018409 13 pageshttpsdoiorg10115520174018409

2 Journal of Advanced Transportation

The focus of this paper is on detecting anomalies andestimating delay on a freeway segment Anomaly detectionand delay estimation are essential for both drivers andtraffic management centers as they reflect severity of trafficcongestion which in turn can be used to minimize its impact(eg by alerting drivers to change route or sending in atow truck to remove a disabled vehicle) While anomalydetection has received more attention in the past decade [8]delay estimation has not equally received the same attentionAs detailed in Section 2 below the majority of previousstudies have focused on estimating and predicting traveltime under normal traffic condition [9ndash15] while relativelyfewer methods focus particularly on traffic delay [5ndash7 10 16]Furthermore themethods proposed in [5 7] are not designedfor real-time operation while the study in [6] focuses only ontraffic delay in work zones One of the most recent studies in[16] provides insight on the capabilities of existingmodels butthe focus is on signalized intersections

The contributions of this paper can be summarized asfollows First the proposed algorithm is designed to operateonline and can be deployed with existing roadside sensorsSecond it can detect anomalies and achieve higher detectionrates and lower false positive rates (ie false alarm rates) com-pared to an existing approach in [17] Third the algorithmcan estimate delay by first estimating the spatial extent ofdelay using a group of roadside sensors and then temporallyestimate delay using empirically derived equations Finallyfor practical purposes the benefit of selecting suitable inputtime window size and prediction horizon is discussed

This paper proposes an algorithm based on a newframework which includes both anomaly detection anddelay estimation so it is a significantly enhanced versionof the algorithm in [18] which focuses only on detectingincidents First the spatial and temporal estimation of delayis incorporated (explained in detail in Section 4) Secondit is shown that the enhanced algorithm in this paper canbe used to detect the presence of shockwaves and estimatetheir associated delays (explained in Section 5) Thirdly thereal-world data sets used to assess the proposed algorithmin this paper are much more significant than the one usedin [18] with 44 more delay cases with additional six monthsof data collection (explained in Section 5) Finally to showthe potential of the proposed algorithm being deployed in adifferent location the well-known NGSIM data [19] is alsoused in the assessment which has not been done in [18]

The rest of this paper is organized as follows Relatedworkis discussed in Section 2 Research framework is presented inSection 3 Then Section 4 describes the proposed algorithmin detail Performance evaluations using real-world dataare described in Section 5 and the results are discussedin Section 6 Finally Section 7 concludes this paper anddiscusses future work

2 Related Work

The majority of previous studies have focused on estimatingand predicting travel time under normal traffic conditions[9ndash15] Among them the most widely used technologiesare probe vehicle and automatic vehicle identification (AVI)

Probe vehicle technologies are installed in the vehiclesthemselves to communicate their locations to some remoteservers where their travel times are calculated [10] Oneof the most well-known probe vehicle technologies is theGlobal Positioning System (GPS) Unlike probe vehicles AVItechnologies are designed to be installed on roadside toidentify individual vehicles in different locations [9] UsingAVI the travel time of each vehicle can be calculated asthe difference between the times that the vehicle passestwo consecutive locations [11] Well-known AVI technologiesare radio frequency identification (RFID) and license platerecognition Algorithms used to estimate travel times aresupport vector regression models [9] a three-layer neuralnetwork model [10] a low-pass adaptive filtering algorithm[11] among others Even though using travel time providesthemost accurate and convenient way to estimate and predictdelay it requires an adequate number of sensors installedin vehicles andor on roadways which are not yet widelyavailable in many areas of developing countries

Besides using probe vehicles and AVI to directly estimatetravel time a group of indirect methods have also beenproposed where travel time is estimated and predicted fromtraffic flow [12ndash15] They use similar assumptions that thereare certain degrees of relationship between flow and delayCommonly employed techniques are regression artificialintelligence and statistical models derived from historicaldata In particular it is shown that travel time can beaccurately predicted up to 20 minutes in advance usinglinear regression [12ndash14] Even though these methods havebeen shown to perform well under normal traffic conditionsthey need adequate historical data which may not alwaysbe available on every road segment Also transient changesunder congested traffic condition may have an inconsistentimpact on the estimation accuracy [15]

Relatively few methods have been designed to estimatetraffic delay that is the additional travel time that occursunder congested conditions The methods in this group relyon a primary assumption that speed does not change signif-icantly over a section of roadway An approach proposed in[20] follows this assumption when incident-induced delayscan be estimated by first establishing reference incident-freetravel profiles Based on reference profiles one can estimateincident-induced delays more accurately Another work [21]estimates incident delay by deploying both incident durationmodel and reduced capacity model into the deterministicqueuing model thereby introducing incident-induced trafficdelays of stochastic nature into an otherwise deterministictrafficThis same principle applies to calculating traffic delayscaused by precipitation where archived weather and trafficdata are used in traffic delay estimation [22] Another goodexample is the method proposed in [5] which estimates delaybased on speed measurements However the method in [5]can only work offline and requires the knowledge of thelocation of an accident The aim of the method proposed inthis paper is similar to the aim in [5] but the main differenceis that our method is designed primarily to work online thatis to process real-time speed inputs in a streaming manner

In our previous works we have proposed algorithms todetect and classify different types of traffic anomalies [8 23]

Journal of Advanced Transportation 3

to transfer the detection capability to roadside sensors at adifferent location [24] and recently to infer traffic transitionat locations where local traffic data is not available [25]The algorithm proposed in this paper focuses on anomaliesthat cause delay so it uses a different approach from ourprevious works in [8 23] More importantly the proposedalgorithm is a significant extension of our previous work in[18] where the capabilities to detect anomalies at individualsensors to detect shockwaves and to estimate traffic delay areshown

3 Research Framework

In this section we present research framework includingdefinitions of certain terminologies that are essential fordeveloping the proposed algorithm in Section 4 Firstly it isimportant to define temporal and spatial extent of delayDelayis defined as an additional travel time experienced by vehiclesdue to circumstances that impede the movement of traffic[27] Spatial extent refers to an area on the road segmentwhere the vehicles are expected to experience delay Temporalextent in this paper refers to short-term estimation that is itindicates the delay vehicles are expected to experience in thenear future Short-term is defined as a time period from a fewseconds to a few hours into the future [3] For example forthe expressway we analyzed in this paper short-term refersto the period of (0 15]minutes

Delay occurs after traffic condition changes from nor-mal to disruptive condition [28] Normal condition refersto the scenario when incoming traffic flow is well belowthe freeway segment capacity Disruptive condition can beassociated with either recurring or nonrecurring congestionRecurring congestion refers to the scenario when traffic flowexceeds freeway capacity (eg duringmorning rush hours onweekdays) Nonrecurring congestion refers to the scenarioswhere the capacity of the freeway segment is reduced byunexpected events (eg accidents and disabled vehicles) Ananomaly in this paper refers to a deviation from expectedtraffic patterns which can occur in recurring congestion (egspeed drops due to shockwave) [25] or precede nonrecurringcongestion (eg speed drops due to accidents and disabledvehicles) [8]

Figure 1 shows how the proposed algorithm can bedeployed with roadside sensors [2] on a freeway segment Atypical roadside sensors system shown in Figure 1 consists ofsensors placed consecutively on a freeway segment In eachsensor there are two main units A sensor unit is responsiblefor collecting traffic data using sensing technologies suchas magnetometers radars lasers and video cameras [29]A communication unit is tasked with sending traffic datato a data collection module using wireless communicationtechnologies such as GPRS 3G or Bluetooth depending onthe sensorsrsquo topology and which technologies are availablelocallyThen the collected traffic data are fed to the algorithmmodule where they are processed by certain algorithms forparticular applicationsThe proposed algorithm in this paperis deployed in the algorithmmodule where it detects anoma-lies and estimates delay The delay information providedby the proposed algorithm can then be used for particular

applications including advanced traveler information systemsand advanced traffic management systems

4 The Proposed Algorithm

In this section the proposed algorithm is presented Firstan overview is given in Section 41 on how each operationof the proposed algorithm processes the input traffic dataand interacts with one another Then Section 42 explainstwo important input parameters which are used to adjustthe proposed algorithm Section 43 presents the anomalydetection operation which is used to detect the occurrenceof delay Finally Section 44 describes how the proposedalgorithm estimates delay spatially and temporally

41 Overview As shown in Figure 2 the proposed algorithmoperates online in a streaming manner and uses temporalsamples of traffic data obtained from roadside sensors asinput For a given time step 119899 and an input time windowsize 119871 each operation is performed with 119871 samples of input119881119898119899 = [V119898119899 V119898119899minus1 V119898119899minus119871+1] from each roadside sensorwhere V119898119899 refers to the average speed of the vehicles thatpass through sensor119898 at time step 119899The proposed algorithmconsists of two main operations namely anomaly detection(block II in Figure 2) and delay estimation (blocks III andIV in Figure 2) The aim of anomaly detection is to detectdeviations from normal traffic patterns that cause trafficdelay Once an anomaly is detected the algorithm activatesdelay estimation operation where the aim is to estimate anadditional travel time vehicles are expected to experience ona given road segment

42 Definitions of Input Time Window Size (119871) and PredictionHorizon (119875) Before we describe further how the proposedalgorithm performs anomaly detection and delay estimationit is important to define time window size 119871 and predictionhorizon 119875 which are two important input parameters asshown in Figure 2 The input time window size 119871 defines thenumber of temporal samples of input (in unit of time) theproposed algorithm can use to detect anomaly and estimatedelay The prediction horizon 119875 defines how long into thefuture (also in unit of time) the vehicles would experiencethe delay estimated by proposed algorithm In other wordswhen a delay of 119889 is estimated by the algorithm at time 119899 thefreeway segment is expected the experience similar delay 119889during the period [119899 119899+119875] Even though it is well known thatincreasing 119875 would reduce estimation accuracy this paperassesses the degree of change of the estimation accuracy inrespect to changes in the values of 119875

43 Anomaly Detection The proposed algorithm is devel-oped based on the assumption that traffic delay is causedby anomalies so delay can be promptly identified by firstdetecting anomalies Based on our experiments as well asthe previous findings in [30] speed flow and occupancy aresuitable for identifying traffic anomalies so they are selectedas inputs As real-world datamay be incomplete inconsistentand noisy the first step is to perform data cleansing asshown in block I in Figure 2 For the proposed algorithm an

4 Journal of Advanced Transportation

Advanced traffic

management system

advanced traveler

information system

Sensor 1Sensor unit

Communication unit

AlgorithmData collection module

system

Traffi

c flow

Sensor 2Sensor unit

Communication unit

Sensor 3Sensor unit

Communication unit

Sensor nSensor unit

Communication unit

Figure 1 How the proposed algorithm can be deployed with roadside sensors on a freeway segmentThe proposed algorithm is the algorithmmodule which uses traffic data collected from sensors to detect anomalies and estimate delay

exponentialmoving average derived fromDELOS smoothing[31] is used to interpolate missing data and remove noiseThe exponentialmoving average is chosen particularly to givemore weights to the latest speed samples as they are moreassociated with delay occurrence

The anomaly detection operation is derived based onthe sliding window approach in [32] but modified to assesseach input time window individually whether it is associ-ated with normal or anomalous traffic patterns The pro-posed algorithm uses pattern matching where each real-time input pattern consisting of 119871 samples of traffic data119881119898119899 = [V119898119899 V119898119899minus1 V119898119899minus119871+1] from roadside sensorsThe input 119881119898119899 is identified as normal or anomalous byassessing its similarity with a set of basis patterns 119880119898119899 =[119906119898119899 119906119898119899minus1 119906119898119899minus119871+1] consisting of normal patterns 119880119873119898and anomalous patterns 119880119860119898 An anomaly is detected if theinput pattern 119881119898119899 is more similar to 119880119860119898 than to 119880119873119898 Inpractice 119881119898119899 and 119880119898119899 can be speed flow and occupancymeasured from both normal and anomalous conditions apriori and stored in a database

The main challenge for anomaly detection is that anoma-lies can cause speed flow and occupancy patterns to behavedifferently in time space and shape For example it isobserved that a disruption occurring between a pair ofupstream and downstream roadside sensors causes speedto drop at the upstream sensor while remain constant ata downstream sensorOn the other hand the same disruption

may cause flow to drop at both upstream and downstreamsensors Furthermore the same disruption can cause thespeed patterns to dropwithin seconds under a scenariowherevehicles are moving at relatively high speeds In contrastit may take place in several minutes in another scenariodepending on traffic flow and driversrsquo behaviors Also theshape of speed drop patterns may be linear (with differentslopes) exponential logarithmic or even sigmoid as foundin [24] To address these challenges Dynamic TimeWarpingis used in the proposed algorithm

Dynamic Time Warping (DTW) is a pattern similar-ity measurement between time series proposed in [33] toovercome limitations of traditional Euclidean distance byintroducing warping axis to adjust the error measurementFigure 3 shows distance measurement between two timeseries patterns 119881 and 119880 using DTW It can be seen inFigure 3(a) that at first 119881 and 119880 do not appear to be similarby using Euclidean distance measurement to compare pointsat the same time However as shown in Figures 3(b) and3(c) DTWoffers amore flexible and efficient approachwheresimilarity between two time series patterns can be assessedby comparing points at different times This is equivalent toldquowrappingrdquo the time series before comparing them

In practice the DTW would operate in an algorithmmodule in Figure 1 to assess similarity between an inputpattern 119881 measured by a roadside sensor and individualbasis pattern 119880 Given an input time window size 119871 acomputer programming language is used to implementDTW

Journal of Advanced Transportation 5

Anomaly detection (II)

Delay estimation (III IV)

Data cleansing (I)

Spatial estimation (III)

Temporal estimation (IV)

Sensor 1

Sensor 2

Sensor 3

I I I I

II II II II

Traffi

c flow

V2

V3

VM

V2 V1V3VM

middotmiddotmiddot

V1 = mnminusL+1 mnminus1 mn

Vm= mnminusL+1 mnminus1 mn

m = 1 2 M

middot middot middot

Sensor M

Traffic data of sensor m attime n using time window L

Predictionhorizon P

Estimated delay outcome dn

Figure 2 Overview of the proposed algorithm

V

U

(a)

V

U

(b)

V

U

(c)

Figure 3 An example of distance measurements between two time series using Dynamic Time Warping [26] (a) Euclidean distancemeasurement between two time series (b) DTW attempts to warp in time axis to find the correct measurement (c) Distance measurementusing DTW

by constructing a 119871-by-119871 matrix where the number of rowsand columns are equivalent to the length of 119881 and 119880respectively Each matrix element (119894 119895) corresponds to thealignment between points V119894 and 119906119895 Then a warping path119882 = 1199081 1199082 119908119870 119871 le 119870 lt 2119871 minus 1 is constructedas a set of matrix elements that defines a point mapping inorder to measure the distance between 119881 and 119880 Similarity

between 119881 and 119880 can be calculated based on these warpingpaths

Based onDTW the proposed algorithm is flexible enoughto cope with these pattern differences and subsequentlycompare 119881 and 119880 more efficiently If the input pattern 119881 isfound to be similar to the basis pattern 119880 associated withanomalies the proposed algorithm declares that an anomaly

6 Journal of Advanced Transportation

Traffi

c flow

Start sensor (1st sensor where anomaly is detected)

Intermediate sensor (anomaly detected)

Intermediate sensor (anomaly detected)

End sensor (1st sensor where anomaly is not detected)

Location where a lane is blocked (eg by an accident or a disabled vehicle)

Figure 4 An illustration of how the proposed algorithm finds start intermediate and end sensors from speed patterns measured by roadsidesensors Start sensor is the first downstream sensor that detects an anomaly End sensor is the first upstream sensor that an anomaly is notdetected Intermediate sensors are used to confirm that the congested area has extended upstream from start sensor to end sensor

is detected and proceeds to delay estimation We note thatanomalies can also be detected bymultiple sensors which canbe used in the delay estimation operation

44 Delay Estimation Delay estimation operation consistsof two consecutive steps (1) estimation of spatial extentof anomalies 119897119899119875 and (2) estimation of temporal extent 119889119899as shown in blocks III and IV in Figure 2 respectivelyThe spatial extent of delay 119897119899119875 is the length of the roadsegment that experiences anomalies which can be estimatedby finding a set of consecutive sensors where anomalies aredetected in Section 43 As shown in Figure 4 the roadsidesensors considered to be in the disruption area consist ofthe start sensor the intermediate sensors and the end sensorThe start sensor is the first downstream sensor where ananomaly is detected that is it is the location where delay isoriginated The intermediate sensors are the sensors locatedconsecutively upstream from the start sensor that also detectsanomalies that is they signify the continuous upstreamextension of anomalies The end sensor is the first upstreamsensor where an anomaly is not detected that is this isthe location where the disruption ceases to exist and trafficupstream from this location is still under normal conditionThen 119897119899119875 is estimated as the distance between the start andthe end sensors

The second step temporal estimation of delay in blockIV in Figure 2 uses 119897119899119875 as input to estimate delay vehiclesare expected to experience on the road segment Usingthe definition in [27] 119889119899 the estimated delay at time 119899 iscalculated as the estimated additional travel time that exceedsthe normal travel time as shown in (1) where 120591119899119875 is theestimated travel time and 120591119899119873 is the normal travel timeon the same road segment The estimated travel time 120591119899119875is calculated by dividing 119897119899119875 with the estimated mean ofspeeds measured at the start intermediate and end sensorsV119899119875

The normal travel time 120591119899119873 is calculated by dividing 119897119899119875with the normal speed V119899119873 We note that V119899119873 is calculatedas the mean of measured speeds in the previous timewindow (size 119871) just before an anomaly is detected in thecurrent time window (also size 119871)Therefore V119899119873 reflects theexpected speed drivers on the freeway segment would haveexperienced if an anomaly had not occurred

119889119899 = 120591119899119875 minus 120591119899119873 120591119899119875 =119897119899119875V119899119875 120591119899119873 =

119897119899119875V119899119873 (1)

The estimated mean of speeds V119899119875 can be calculated asfollows Based on our previous study in [24] and furtheranalysis of real-world data over a 6-month period it isobserved that the occurrences of delay are associatedwith fiveanomalous speed drop patterns namely linear exponentiallogarithmic sigmoid and unclassified Furthermore it ispossible to find empirical equations that best fit each of thesefive speed drop patterns (eg an exponential equation thatbest fits the exponential speed drop patterns) [24] A kernelfunction is used for the unclassified speed drop patterns

For a given prediction horizon 119875 and a total of119872 startintermediate and end sensors these empirical equations canbe used to temporally extrapolate speeds well into the futurefor each roadside sensor119898 that is [V119898119899+1 V119898119899+2 V119898119899+119875]Then V119899119875 can be obtained by first averaging temporallyextrapolated speed samples V119898119899119875 and then averaging spa-tially across all119872 sensors as follows

V119899119875 =sum119872119898=1 V119898119899119875119872 V119898119899119875 =

sum119875119894=1 V119898119899+119894119875 (2)

In practice delay estimation can be operated as followsOnce delay has been detected speed patterns measured byeach roadside sensor are measured and transmitted to thealgorithmmodule in Figure 1 UsingDynamicTimeWarping

Journal of Advanced Transportation 7

Daokanong

Suksawat Sathupradit Sathupradit River side

Port

1 2 3 4 5 6 7 8 9 10 11

644m1977m1078m393m390m786m1337m671m1505m1152m

2nd stage2nd stage

Figure 5 A sketch of expressway segment where real-world data was obtained

these speed patterns are compared with the basis speedpatterns associated with anomalies 119880119860119898 which have beencollected and stored a priori to find similarity Once themost similar basis speed drop pattern has been found itscorresponding empirical equation is used to calculate theestimated mean speed V119899119875 as shown in (2) which is thenused to calculate 120591119899119875 and eventually to estimate delay 119889119899 in(1) Note that the calculation in (1) is for short-term delay[3] Extending (1) to estimate vehiclersquos hour of delay or totaldelay experienced by all delayed vehicles on the road segmentcan be performed by adding traffic volume as amultiplicationfactor to (1)

5 Performance Evaluations

In this section performance evaluations are conducted usingreal-world data First Section 51 presents two real-worlddata sets used to assess the proposed algorithm ThenSection 52 describes an existing algorithm [17] which isused as benchmark In Section 53 performance evaluationparameters are presented

51 Descriptions of Real-World Data Sets

511 Real-World Data from Roadside Camera Sensors Theproposed algorithm is assessed using real-world data col-lected from a series of camera sensors installed on heavilyused inbound section ofChalerm-Maha-NakhonExpresswayin Bangkok Thailand These sensors were part of a projectjointly initiated by the Expressway Authority of Thailand(EXAT) and National Electronic and Computer TechnologyCenter (NECTEC) aiming at real-time traffic monitoringIt consists of a series of eleven camera sensors installedconsecutively along the 11 km expressway Figure 5 shows thetopology of the expressway and the locations of the camerastations where the distance between two consecutive camerasensors ranges from 644 to 1977 meters Each camera sensoris capable of detecting individual vehicles that passed theexpressway segment and produced time series of traffic dataincluding speed occupancy and flow in 1-minute intervals[34] Each camera sensor is connected to a local EXAT oper-ation office via wireless cellular network technologies (egGPRS and 3G) where operators can view real-time trafficinformation and make decisions on traffic management

A road segment within the full of view of a video camerais used as a designated area to calculate traffic data inputSpeed is measured as a distance in a vehiclersquos traverse in the

Figure 6 An image of a traffic delay occurrence captured at the 4thcamera sensor

designated area divided by the time it takes Flow ismeasuredas the rate of vehicles passing a designated area at a given timeinterval Occupancy is measured as a fraction of time that adesignated area is occupied by vehicles

The data set used in this paper was collected from thissensor network over a one year period from January 1 toDecember 31 2010 In this paper we chose the 3rd 4th 5thand 6th sensors as the experimental sites An example of traf-fic congestion viewed by a camera sensor studied in this paperis shown in Figure 6 For evaluation purposes anomaliesand their associated delay that took place were independentlyand manually logged by a team of transportation expertsA total of 16 nonrecurring congestion cases were recordedwhich are used to test the anomaly detection capability ofthe proposed algorithm Each record was associated with thetime instances prior to the disruption occurrence and after ithad ended Besides these 16 cases further investigation withanother team of experts also revealed additional 44 delaycases Therefore a total of 60 delay cases are used to test theestimation operation of the proposed algorithm

We note that the traffic data input of the proposedalgorithm is typical data that most roadside sensors forexample loop detectors already produce Therefore weanticipate that if the sensors are to be replaced with othertypes of sensors the performance of the system should notchange significantly This is worth further investigation forlarge scale deployment

512 NGSIM Data To show the potential of being deployedat a different location the proposed algorithm is also assessedusingNGSIM data [19]The selected data set was collected onthe southbound US Highway 101 in Los Angeles California

8 Journal of Advanced Transportation

USA at 0750 amndash0835 am This data set is selected becauseit was collected during the buildup of congestion and consistsof shockwaves which can potentially induce additional delayIt should be noted that the anomaly cases and the amountof delay caused by them were not labeled in this data setso the full algorithm could not be deployed for this dataset However the potential of the algorithm could be shownby shockwave detection To assess the proposed algorithmfour virtual sensors are placed approximately 01 km apart toobtain traffic data to capture the occurrence of shockwavesand their spatial extents Similar to conventional fixed trafficsensors (eg camera sensors and loop detectors) thesevirtual sensors calculate traffic data based on traffic flow andspeed that pass the sensorsrsquo locations An input time windowsize 119871 of 5 minutes is used where time series of 30-secondaverage speeds obtained from upstream and downstreamvirtual sensors are used as inputs The occurrences of shock-waves are observed and recorded by a team of transportationresearchers [25]

52 Benchmark Algorithm To assess the capability of theproposed algorithm in respect to existing approaches abenchmark algorithm is used [17]This benchmark algorithmis selected because it has a similar objective to the proposedalgorithm in detecting anomalies that cause delay Also thisalgorithm has already been implemented in a real-worldanomaly detection system [17] Furthermore composed ofa forecasting component and a detection logic componentit is designed to assess the difference between traffic dataobserved at consecutive roadside sensors

However the main difference between the benchmarkalgorithm and the proposed algorithm lies in the detectionlogic The benchmark algorithm is designed to detect sud-den significant changes of traffic inputs from the forecastcomponent values by establishing threshold values (speedand occupancy) to filter out gradual changes in traffic inputscaused by recurring congestion and an anomalous signalwill only be raised when the specified threshold values havebeen exceeded On the other hand the proposed algorithm isdesigned to detect both sudden and gradual changes whichenables it to detect different types of anomalies with fewerfalse alarms This difference will become more apparentwhen both algorithms are assessed with real-world data inSection 6

53 Performance Evaluation Parameters To numericallyevaluate the proposed algorithm we use detection rate (DR)false positive rate (FPR) mean time to detection (MTTD)andmean absolute percentage error (MAPE) as shown in (3)(4) (5) and (6) respectively DR FPR and MTTD are usedto assess the anomaly detection operation of the proposedalgorithm MAPE is used to assess the accuracy of delayestimation of a total 119863 recorded delay cases where the errorof the estimation 119889119899119895 of the 119895th recorded delay at time 119899 isnormalized by the actual delay119889119886119899119895 tominimize the possibilityof some estimated delay samples from dominating the others

DR =Number of delays detectedTotal number of delays

(3)

FPR

=Number of detected delays that are false positives

Total number of detected delays

(4)

MTTD

=Sum of time taken to detect individual delays

Total number of detected delays

(5)

MAPE = 1119863

119863

sum119895=1

10038161003816100381610038161003816119889119886119899119895 minus 119889119899119895

10038161003816100381610038161003816119889119886119899119895 (6)

The performance parameters are applicable for data setsthat contain labels for traffic anomalies and the amount ofdelay caused by the anomalies Therefore these parameterswill be evaluated for the real-world data collected inThailandbut due to the lack of anomaly labels could not be applied forthe NGSIM data set

6 Results and Discussions

In this section the proposed algorithm is assessed usingreal-world data described in Section 51 and compared tothe benchmark algorithm described in Section 52 Bothalgorithms are tested as if they operate online where eachdetection operation is performed per input time window 119871

61 Assessment Using NGSIM Data The proposed algorithmis assessed on its capability to detect the the presence ofshockwaves which refers to a phenomenon where vehicleshave to temporarily slow down below normal speed andsubsequently force the vehicles behind to slow down furtherThis phenomenon is known to create shockwave fronts ormoving speed drop areas that move backward disruptingflow and annoying drivers behind [35] The occurrence ofshockwaves during congestion not only causes additionaldelay but also increases chances of rear-end collisionsIdentifying shockwaves during congestion also presents achallenge as it occurs temporarily and requires an algorithmthat is fast and sensitive enough to detect them [25] Theoccurrence of shockwaves represents unexpected events sothey are considered as anomalies in this paper

Figure 7 shows that the proposed algorithm can detectall four occurrences (Figure 7(a)) of shockwaves whilebenchmark algorithm can detect only two (Figure 7(b))As shown in Figure 7(b) only abrupt changes in speeddifference (second and third occurrences) can be detectedby the benchmark algorithm This algorithm is designedto detect consecutive differences between upstream anddownstream traffic measurements so it operates well withconsecutive changes in speed difference that are significantenough On the other hand benchmark algorithmmisses thefirst and fourth occurrences of shockwaves in Figure 7(b) asthey are associated with relatively more gradual changes ofspeed difference Unlike benchmark algorithm the proposedalgorithm uses DTW to assess input patterns to detect delayoccurrences so it can detect all four occurrences associated

Journal of Advanced Transportation 9

Upstream sensorDownstream sensor

805 820 835750Time (hoursminutes)

01020304050607080

Spee

d (k

mh

)

(a)

One time step aheadTwo time steps ahead

805 820 835750Time (hoursminutes)

minus3minus2minus1

0123

Spee

d di

ffere

nces

(km

h)

(b)

Figure 7 Detecting delay occurrences associated with shockwaves in NGSIM data using (a) the proposed algorithm and (b) benchmarkalgorithm [17] Dashed vertical lines denote the time points where the algorithms detect delay occurrences Red squares denote the durationwhere the occurrences of shockwaves are observed

with both abrupt and gradual speed changes as shown inFigure 7(a)

Once the occurrences of shockwaves are detected theproposed algorithm proceeds to estimate delay as describedin Section 44 Using the spatial extent estimation describedin Section 44 out of four sensors three sensors are usuallyfound to detect speed drops associated with shockwaves at atime Therefore the most downstream sensor and the mostupstream sensor are selected as the start and end sensorsrespectively and 119897119899119875 is found to be 048 km

Table 1 shows the values of V119899119875 V119899119873 120591119899119875 and 120591119899119873 in (1)for each of the four shockwave occurrences For comparisonpurposes delay estimation is also applied for the second andthird occurrences that are detected by benchmark algorithm(see Figure 7) Note that since benchmark algorithm cannotdetect the first and fourth occurrences estimation operationis not activated during these times and normal speed andtravel times are used instead We further note that all fouroccurrences and their associated travel times and delays canbe calculated consecutively online by the proposed algorithm

As shown in Table 1 the four occurrences of shockwavesdetected in Figure 7 are associated with travel times (120591119899119875 in(1)) of 5682 s 5554 s 6098 s and 5379 s respectively Theproposed algorithm is able to detect all four occurrences andthese estimated travel times would all be included in the delaycalculation On the other hand if the benchmark algorithmwas used instead only the second and third occurrences andtheir associated travel times would be reported resulting inless accurate delay calculationWe note that even though thedifference between the delays obtained from the proposedalgorithm and benchmark algorithm in this example maybe small significant improvement can be seen when higherdegree of disruptions are assessed

The results shown in Table 1 have many applicationsin advanced traveler information systems and advancedtraffic management systems A direct application would beto estimate travel time of the freeway segment and report itto motorists The ability to detect occurrences of shockwavescan also be used in other applications such as calculatingshockwave propagation time [25] It can also be used to warn

motorists to adjust intervehicle spacing to avoid rear-endcollisions

62 Assessment Using Real-World Data Collected in ThailandThis section assesses the capability of the proposed algorithmand the benchmark algorithm [17] to detect the sixteenrecorded traffic anomalies in the real-world data set describedin Section 51 As inputs both algorithms use speed flowand occupancy collected by roadside sensorsThe assessmentfollows a 119896-fold cross-validation technique The real-worlddata set consists of 16 anomaly cases 1198631 1198632 11986316 soevaluation is performed 16 times At an iteration 119894 119894 =1 2 16 the anomaly case119863119894 is reserved as a test set whilethe other anomaly cases are collectively used to calibratethe proposed algorithm and benchmark algorithm That isin the first iteration 1198632 11986316 collectively serve as thetraining set which is then tested with1198631Then in the seconditeration 1198631 1198633 11986316 are used for training while 1198632 isused for testing and so on

Table 2 shows that the proposed algorithm achieveshigher detection rate and lower false alarm rate than thebenchmark algorithm Particularly compared to benchmarkalgorithm the proposed algorithm improves the detectionrate and false positive rate by approximately 10 and 7respectively Figures 8 and 9 show examples of anomaliesdetected by the proposed algorithm and found to causetraffic delay Based on the records these anomalies areassociated with disabled vehicles parking over the shoulderarea

It is found that for the benchmark algorithm miss-detection usually occurs under low traffic flow This algo-rithm always needs data from at least two sensors and themagnitude of traffic flow measured by the roadside sensorsis not large enough to trigger the benchmark algorithmWhile medium to heavy traffic flow triggers the benchmarkalgorithm it also increases false alarms This algorithm isdesigned to directly assess the difference between upstreamand downstream sensors of the roadside sensors whichsubsequently increases its sensitivity and triggers many falsealarms

10 Journal of Advanced Transportation

Table 1 Estimation results onNGSIMdata using the proposed algorithm and benchmark algorithm 119897119899119875 =048 kmNote that if the shockwaveis not detected the estimated travel time 120591119899119875 would be equal to normal travel time 120591119899119873

1st 2nd 3rd 4thProposed algorithmShockwaves detected Yes Yes Yes YesV119899119875 (see (1)) 3041 kmh 3112 kmh 2834 kmh 3213 kmh120591119899119875 (see (1)) 5682 s 5554 s 6098 s 5379 s119889119899 = 120591119899119875 minus 120591119899119873 1841 s 1893 s 2399 s 541 sBenchmark algorithm [17]Shockwaves detected No Yes Yes NoV119899119875 (see (1)) 4499 kmh 3112 kmh 2834 kmh 3572 kmh120591119899119875 (see (1)) 3841 s 5554 s 6098 s 4838 s119889119899 = 120591119899119875 minus 120591119899119873 000 s 1893 s 2399 s 000 sIf all 4 shockwaves are not detectedV119899119873 (see (1)) 4499 kmh 4721 kmh 4671 kmh 3572 kmh120591119899119873 (see (1)) 3841 s 3660 s 3700 s 4838 s

Table 2 Results on anomaly detection using data set collected inThailand

Performanceparameters

Proposedalgorithm

Benchmarkalgorithm [17]

DR () 94 83FPR () 5 12MTTD (seconds) 340 140

Figure 8 An anomaly detected at the 4th camera sensor

Unlike the benchmark algorithm the proposed algorithmuses DTW to assess traffic data from each sensor individuallyThis enables the proposed algorithm to recognize patternsassociated with anomalies more effectively including underlow traffic flow The use of DTW also reduces a numberof false alarms as it ensures that unrecognized patternswill not trigger detection at individual sensors In respectto mean time to detection (MTTD) it takes longer forthe proposed algorithm to detect anomalies compared tobenchmark algorithm as it needs to wait till enough temporalsamples of traffic data have been collected in each sensorbefore the detection operation can be triggered Howeverthe mean time to detection is still less than 6 minutes which

Figure 9 Another anomaly detected at the 4th camera sensor

should give enough time for traffic operators to response [8]To reduce mean time to detection of the proposed algorithmshorter time samples of traffic data can be used ReducingMTTD while maintaining acceptable levels of DR and FPRis worth further investigation

To deploy the proposed algorithm computation timeof the algorithm could be an added delay to the MTTDEven though the computation time was not systematicallymeasured for the performance evaluation the computationtime was in order of seconds For the computerrsquos specificationthat we use which was a personal laptop with 25 GHz CPUand 4GB of RAM the computation time was approximately[1 3) seconds As the computing time (a few seconds) wasmuch smaller than the mean time to detection (up to 6minutes) it is reasonable to consider the computing timenegligible

This section further discusses the results when the pro-posed algorithm was applied to estimate 60 delay casesdescribed in Section 51 using different input time windowsizes (119871) and prediction horizons (119875) For practical purposesit is important to select suitable input time window sizeand prediction horizon to enable the proposed method to

Journal of Advanced Transportation 11

L = 3 L = 10L = 5

P = 5

P = 10

P = 15

0

005

01

015

02

025

03

035

04M

APE

Figure 10Mean absolute percentage error (MAPE) fromestimating60 delay cases using different input time window sizes 119871 (minutes)and prediction horizons 119875 (minutes)

perform effectively Even though the 15-minute interval isoften recommended [9] to avoid strong fluctuation usinglower input timewindow sizes can providemore fine-grainedinformation for practitioners [25] Therefore in this paperwe use 119871 = 3 5 and 10 minutes for input time windowsize For prediction horizons we use 119875 = 5 10 and 15minutes as this paper is focused on short-term estimation(less than 15 minutes) Also changes in traffic condition onthis experimental site are normally found to occur in less than15 minutes

The proposed algorithm is assessed on its capability toestimate delay using MAPE Figure 10 shows MAPE fromestimating 60 delay cases in the real-world data set withdifferent input timewindow sizes (119871) and prediction horizons(119875) These results show that the estimation part of theproposed algorithm can achieve MAPE lower than 01 forthis set of real-world data This accuracy is achieved becausethe estimation part consists of both spatial and temporalestimations which enable the proposed algorithm to firstestimate the area disrupted by anomalies and then estimatethe delay itself Furthermore it can be seen that increasinginput time window size (119871) reduces MAPE as using moreinput samples enhances estimation accuracy For the impactof prediction horizons (119875) MAPE tends to increase withlarger prediction horizons as it is more difficult to estimatedelays farther ahead into the future However as shownin Figure 10 when 119875 is increased MAPE in fact increasesat a much smaller rate when larger values of 119871 are usedTherefore it is recommended for practitioners that if theobjective is to estimate delays far ahead into the future (usinglarge prediction horizons) larger input time window size isneeded

Based on the findings in this section it is importantthat practitioners select a suitable input time window sizeaccording to their practical objectives If the objective is toprovide prompt delay information small input time windowsize should be used On the other hand if the objective is toprovide accurately estimated delays for traffic managementpurposes larger input time window sizes should be chosenFurthermore based on the proposed algorithm a morecomplex and flexible system can also be further developedFor example a stepwise approach can be implemented on topof the proposed algorithm where a small input time windowsize is used as a trigger to promptly discover the occurrenceof traffic delay Then in the next step the input time windowsize can be increased to collect more traffic data to achievemore accurate detection and estimation of traffic delay

For anomalies with large scale effect beyond the 11 kmexperimental site we anticipate that the proposed algo-rithm should be able to detect those anomalies if theyoccur downstream and cause changes in traffic patterns topropagate upstream into the 11 km segment To accuratelyestimate delay basis patterns associated with such large scaleanomalies need to be incorporated to distinguish themThenan extrapolation technique needs to be incorporated to enablethe proposed algorithm to estimate delay beyond the 11 kmexpressway segment where local traffic information cannotbe directly observed [25] This is worth further investigation

7 Conclusions and Future Work

This paper proposes an algorithm to detect anomalies andestimate delay using traffic data from roadside sensors Theproposed algorithm uses traffic data from individual sensorsto detect anomalies and then uses data from consecutivedownstream and upstream sensors together to spatially andtemporally estimate delay Performance evaluations usingreal-world data show that the proposed algorithm achieveshigher detection rate and lower false positive rate when usedto detect anomalies compared to an existing benchmarkalgorithmAn investigationwithNGSIMdata also shows thatthe proposed algorithm is also able to detect the presence ofshockwaves and subsequently estimate delaymore efficientlythan benchmark algorithm Further results show that theestimation error increases at a much smaller rate whensuitable input time window size and prediction horizon areselected

There are interesting aspects that are worth furtherinvestigation One interesting aspect is to reduce the pro-cessing time while maintaining low detection and estima-tion errors A possible approach is to introduce a featureextraction method that can extract hidden patterns whensmaller input time window sizes are used [8] Alternativeapproaches include studying a set of hidden representativetraffic flow features in advance by pretraining a deep-learningmodel [36] and populating traffic features into a geneticalgorithm learning method to reduce dimensionality [37]Another interesting aspect is to improve inference capabilityby incorporating a spatial inferencemethod to estimate trafficdata itself at the locations where they cannot be directlymeasured [25] Also it would be practically interesting to

12 Journal of Advanced Transportation

assess the proposed algorithm with different road types andgeometries

In addition to improvements of the algorithm an interest-ing aspect to investigate is the applications of the algorithmThe solution provided in this paper is designed to be astreaming algorithm which can be run online thereforeit can be naturally extended to use in various real-timeapplications For example the detected anomaly and esti-mated additional travel time would be useful for both trafficoperators and travelers if the information is delivered ina timely and convenient manner Active communicationbetween roadside sensors and vehicle-to-infrastructure (V2I)could be used to promptly transmit the information from oursolution to nearby roadside message boards to display theinformation to the drivers on the roads or vehicles with V2Icommunication to adapt routing and navigation

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this article

Acknowledgments

The authors would like to thankDr Supakorn Siddhichai andhis IMG team at NECTEC for their assistance in providingreal-world data for the analysis

References

[1] N Parrado and Y Donoso ldquoCongestion based mechanism forroute discovery in a V2I-V2V system applying smart devicesand IoTrdquo Sensors vol 15 no 4 pp 7768ndash7806 2015

[2] K Nellore and G P Hancke ldquoA survey on urban trafficmanagement system using wireless sensor networksrdquo Sensorsvol 16 no 2 article 157 2016

[3] E I Vlahogianni M G Karlaftis and J C Golias ldquoShort-term traffic forecasting where we are and where wersquore goingrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 3ndash19 2014

[4] R Li and G Rose ldquoIncorporating uncertainty into short-term travel time predictionsrdquo Transportation Research Part CEmerging Technologies vol 19 no 6 pp 1006ndash1018 2011

[5] Y Chung ldquoAmethodological approach for estimating temporaland spatial extent of delays caused by freeway accidentsrdquo IEEETransactions on Intelligent Transportation Systems vol 13 no 3pp 1454ndash1461 2012

[6] J Weng and Q Meng ldquoEstimating capacity and traffic delayin work zones an overviewrdquo Transportation ResearchmdashPart CEmerging Technologies vol 35 pp 34ndash45 2013

[7] X Zhang E Onieva A Perallos E Osaba and V C SLee ldquoHierarchical fuzzy rule-based system optimized withgenetic algorithms for short term traffic congestion predictionrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 127ndash142 2014

[8] S Thajchayapong E S Garcia-Trevino and J A Barria ldquoDis-tributed classification of traffic anomalies using microscopictraffic variablesrdquo IEEETransactions on Intelligent TransportationSystems vol 14 no 1 pp 448ndash458 2013

[9] M Lippi M Bertini and P Frasconi ldquoShort-term traffic flowforecasting an experimental comparison of time-series anal-ysis and supervised learningrdquo IEEE Transactions on IntelligentTransportation Systems vol 14 no 2 pp 871ndash882 2013

[10] F Zheng andH Van Zuylen ldquoUrban link travel time estimationbased on sparse probe vehicle datardquo Transportation ResearchmdashPart C Emerging Technologies vol 31 pp 145ndash157 2013

[11] F Dion and H Rakha ldquoEstimating dynamic roadway traveltimes using automatic vehicle identification data for low sam-pling ratesrdquoTransportation Research Part BMethodological vol40 no 9 pp 745ndash766 2006

[12] D Paterson and G Rose ldquoDynamic travel time estima-tion on instrumented freewaysrdquo in Proceedings of the 6thWorld Congress on Intelligent Transportation Systems TorontoCanada 1999

[13] J Kwon B Coifman and P Bickel ldquoDay-to-day travel-timetrends and travel-time prediction from loop-detector datardquoTransportation Research Record no 1717 pp 120ndash129 2000

[14] J Rice and E Van Zwet ldquoA simple and effective method forpredicting travel times on freewaysrdquo IEEE Transactions onIntelligent Transportation Systems vol 5 no 3 pp 200ndash2072004

[15] C D R Lindveld RThijs P H L Bovy andN J van der ZijppldquoEvaluation of online travel time estimators and predictorsrdquoTransportation Research Record vol 1719 pp 45ndash53 2000

[16] H Yang S Chen YWang BWu andZWang ldquoComparison ofdelay estimation models for signalised intersections using fieldobservations in Shanghairdquo IET Intelligent Transport Systems vol10 no 3 pp 165ndash174 2016

[17] C L Mak and H S L Fan ldquoDevelopment of dual-stationautomated expressway incident detection algorithmsrdquo IEEETransactions on Intelligent Transportation Systems vol 8 no 3pp 480ndash490 2007

[18] K Hi-Ri-O-Tappa C Likitkhajorn A Poolsawat and S Tha-jchayapong ldquoTraffic incident detection system using series ofpoint detectorsrdquo in Proceedings of the 15th International IEEEConference on Intelligent Transportation Systems (ITSC rsquo12) pp182ndash187 Anchorage Alaska USA September 2012

[19] V Alexiadis J Colyar J Halkias R Hranac and G McHaleldquoThe next generation simulation programrdquo ITE Journal vol 74no 8 pp 22ndash26 2004

[20] F G Habtemichael M Cetin and K A Anuar ldquoIncident-induced delays on freeways quantificationmethod by groupingsimilar traffic patternsrdquo Transportation Research Record vol2484 pp 60ndash69 2015

[21] J Li C-J Lan and X Gu ldquoEstimation of incident delay andits uncertainty on freeway networksrdquo Transportation ResearchRecord vol 1959 pp 37ndash45 2006

[22] Y Chung ldquoAssessment of non-recurrent congestion causedby precipitation using archived weather and traffic flow datardquoTransport Policy vol 19 no 1 pp 167ndash173 2012

[23] J A Barria and SThajchayapong ldquoDetection and classificationof traffic anomalies using microscopic traffic variablesrdquo IEEETransactions on Intelligent Transportation Systems vol 12 no3 pp 695ndash704 2011

[24] K HiriOtappa and S Thajchayapong ldquoGeneralizability andtransferability of incident detection algorithm using dynamictime warpingrdquo in Proceedings of the in 19th World Congress onIntelligent Transportation Systems Vienna Austria 2012

Journal of Advanced Transportation 13

[25] S Thajchayapong and J A Barria ldquoSpatial inference of traffictransition using microndashmacro traffic variablesrdquo IEEE Transac-tions on Intelligent Transportation Systems vol 16 no 2 pp854ndash864 2015

[26] C A Ratanamahatana and E Keogh ldquoEverything you knowabout dynamic time warping is wrongrdquo in Proceedings of the3rd Workshop on Mining Temporal and Sequential Data in Con-junction with the 10th ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining (KDD rsquo04) CiteseerSeattle Wash USA August 2004

[27] Transportation Research Thesauras Transportation ResearchBoard National Academy httptrttrborgtrtaspNN=Bthd

[28] N Chiabaut C Buisson and L Leclercq ldquoFundamental dia-gram estimation through passing rate measurements in conges-tionrdquo IEEE Transactions on Intelligent Transportation Systemsvol 10 no 2 pp 355ndash359 2009

[29] J J Fernandez-Lozano M Martın-Guzman J Martın-Avilaand A Garcıa-Cerezo ldquoA wireless sensor network for urbantraffic characterization and trend monitoringrdquo Sensors vol 15no 10 pp 26143ndash26169 2015

[30] J-S Oh C Oh S G Ritchie and M Chang ldquoReal-timeestimation of accident likelihood for safety enhancementrdquoJournal of Transportation Engineering vol 131 no 5 pp 358ndash363 2005

[31] A P Chassiakos and Y J Stephanedes ldquoSmoothing algorithmsfor incident detectionrdquo Transportation Research Record vol1394 pp 8ndash16 1993

[32] V Alarcon-Aquino and J A Barria ldquoAnomaly detection incommunication networks using waveletsrdquo IEE ProceedingsmdashCommunications vol 148 no 6 pp 355ndash362 2001

[33] D Berndt and J Clifford ldquoUsing dynamic time warping tofind patterns in time seriesrdquo in Proceedings of the Workshop onKnowledge Discovery in Databases (KDD rsquo94) vol 10 pp 359ndash370 Seattle Wash USA August 1994

[34] K Kiratiratanapruk and S Siddhichai ldquoVehicle detection andtracking for traffic monitoring systemrdquo in Proceedings of theIEEE Region 10 Conference (TENCON rsquo06) pp 1ndash4 IEEE HongKong November 2006

[35] A Hegyi B De Schutter and J Hellendoorn ldquoOptimal coordi-nation of variable speed limits to suppress shock wavesrdquo IEEETransactions on Intelligent Transportation Systems vol 6 no 1pp 102ndash112 2005

[36] Y Lv Y Duan W Kang Z Li and F-Y Wang ldquoTraffic flowprediction with big data a deep learning approachrdquo IEEETransactions on Intelligent Transportation Systems vol 16 no2 pp 865ndash873 2015

[37] P Lopez-Garcia E Onieva E Osaba A D Masegosa and APerallos ldquoA hybrid method for short-term traffic congestionforecasting using genetic algorithms and cross entropyrdquo IEEETransactions on Intelligent Transportation Systems vol 17 no 2pp 557ndash569 2016

Journal of Advanced Transportation 3

to transfer the detection capability to roadside sensors at adifferent location [24] and recently to infer traffic transitionat locations where local traffic data is not available [25]The algorithm proposed in this paper focuses on anomaliesthat cause delay so it uses a different approach from ourprevious works in [8 23] More importantly the proposedalgorithm is a significant extension of our previous work in[18] where the capabilities to detect anomalies at individualsensors to detect shockwaves and to estimate traffic delay areshown

3 Research Framework

In this section we present research framework includingdefinitions of certain terminologies that are essential fordeveloping the proposed algorithm in Section 4 Firstly it isimportant to define temporal and spatial extent of delayDelayis defined as an additional travel time experienced by vehiclesdue to circumstances that impede the movement of traffic[27] Spatial extent refers to an area on the road segmentwhere the vehicles are expected to experience delay Temporalextent in this paper refers to short-term estimation that is itindicates the delay vehicles are expected to experience in thenear future Short-term is defined as a time period from a fewseconds to a few hours into the future [3] For example forthe expressway we analyzed in this paper short-term refersto the period of (0 15]minutes

Delay occurs after traffic condition changes from nor-mal to disruptive condition [28] Normal condition refersto the scenario when incoming traffic flow is well belowthe freeway segment capacity Disruptive condition can beassociated with either recurring or nonrecurring congestionRecurring congestion refers to the scenario when traffic flowexceeds freeway capacity (eg duringmorning rush hours onweekdays) Nonrecurring congestion refers to the scenarioswhere the capacity of the freeway segment is reduced byunexpected events (eg accidents and disabled vehicles) Ananomaly in this paper refers to a deviation from expectedtraffic patterns which can occur in recurring congestion (egspeed drops due to shockwave) [25] or precede nonrecurringcongestion (eg speed drops due to accidents and disabledvehicles) [8]

Figure 1 shows how the proposed algorithm can bedeployed with roadside sensors [2] on a freeway segment Atypical roadside sensors system shown in Figure 1 consists ofsensors placed consecutively on a freeway segment In eachsensor there are two main units A sensor unit is responsiblefor collecting traffic data using sensing technologies suchas magnetometers radars lasers and video cameras [29]A communication unit is tasked with sending traffic datato a data collection module using wireless communicationtechnologies such as GPRS 3G or Bluetooth depending onthe sensorsrsquo topology and which technologies are availablelocallyThen the collected traffic data are fed to the algorithmmodule where they are processed by certain algorithms forparticular applicationsThe proposed algorithm in this paperis deployed in the algorithmmodule where it detects anoma-lies and estimates delay The delay information providedby the proposed algorithm can then be used for particular

applications including advanced traveler information systemsand advanced traffic management systems

4 The Proposed Algorithm

In this section the proposed algorithm is presented Firstan overview is given in Section 41 on how each operationof the proposed algorithm processes the input traffic dataand interacts with one another Then Section 42 explainstwo important input parameters which are used to adjustthe proposed algorithm Section 43 presents the anomalydetection operation which is used to detect the occurrenceof delay Finally Section 44 describes how the proposedalgorithm estimates delay spatially and temporally

41 Overview As shown in Figure 2 the proposed algorithmoperates online in a streaming manner and uses temporalsamples of traffic data obtained from roadside sensors asinput For a given time step 119899 and an input time windowsize 119871 each operation is performed with 119871 samples of input119881119898119899 = [V119898119899 V119898119899minus1 V119898119899minus119871+1] from each roadside sensorwhere V119898119899 refers to the average speed of the vehicles thatpass through sensor119898 at time step 119899The proposed algorithmconsists of two main operations namely anomaly detection(block II in Figure 2) and delay estimation (blocks III andIV in Figure 2) The aim of anomaly detection is to detectdeviations from normal traffic patterns that cause trafficdelay Once an anomaly is detected the algorithm activatesdelay estimation operation where the aim is to estimate anadditional travel time vehicles are expected to experience ona given road segment

42 Definitions of Input Time Window Size (119871) and PredictionHorizon (119875) Before we describe further how the proposedalgorithm performs anomaly detection and delay estimationit is important to define time window size 119871 and predictionhorizon 119875 which are two important input parameters asshown in Figure 2 The input time window size 119871 defines thenumber of temporal samples of input (in unit of time) theproposed algorithm can use to detect anomaly and estimatedelay The prediction horizon 119875 defines how long into thefuture (also in unit of time) the vehicles would experiencethe delay estimated by proposed algorithm In other wordswhen a delay of 119889 is estimated by the algorithm at time 119899 thefreeway segment is expected the experience similar delay 119889during the period [119899 119899+119875] Even though it is well known thatincreasing 119875 would reduce estimation accuracy this paperassesses the degree of change of the estimation accuracy inrespect to changes in the values of 119875

43 Anomaly Detection The proposed algorithm is devel-oped based on the assumption that traffic delay is causedby anomalies so delay can be promptly identified by firstdetecting anomalies Based on our experiments as well asthe previous findings in [30] speed flow and occupancy aresuitable for identifying traffic anomalies so they are selectedas inputs As real-world datamay be incomplete inconsistentand noisy the first step is to perform data cleansing asshown in block I in Figure 2 For the proposed algorithm an

4 Journal of Advanced Transportation

Advanced traffic

management system

advanced traveler

information system

Sensor 1Sensor unit

Communication unit

AlgorithmData collection module

system

Traffi

c flow

Sensor 2Sensor unit

Communication unit

Sensor 3Sensor unit

Communication unit

Sensor nSensor unit

Communication unit

Figure 1 How the proposed algorithm can be deployed with roadside sensors on a freeway segmentThe proposed algorithm is the algorithmmodule which uses traffic data collected from sensors to detect anomalies and estimate delay

exponentialmoving average derived fromDELOS smoothing[31] is used to interpolate missing data and remove noiseThe exponentialmoving average is chosen particularly to givemore weights to the latest speed samples as they are moreassociated with delay occurrence

The anomaly detection operation is derived based onthe sliding window approach in [32] but modified to assesseach input time window individually whether it is associ-ated with normal or anomalous traffic patterns The pro-posed algorithm uses pattern matching where each real-time input pattern consisting of 119871 samples of traffic data119881119898119899 = [V119898119899 V119898119899minus1 V119898119899minus119871+1] from roadside sensorsThe input 119881119898119899 is identified as normal or anomalous byassessing its similarity with a set of basis patterns 119880119898119899 =[119906119898119899 119906119898119899minus1 119906119898119899minus119871+1] consisting of normal patterns 119880119873119898and anomalous patterns 119880119860119898 An anomaly is detected if theinput pattern 119881119898119899 is more similar to 119880119860119898 than to 119880119873119898 Inpractice 119881119898119899 and 119880119898119899 can be speed flow and occupancymeasured from both normal and anomalous conditions apriori and stored in a database

The main challenge for anomaly detection is that anoma-lies can cause speed flow and occupancy patterns to behavedifferently in time space and shape For example it isobserved that a disruption occurring between a pair ofupstream and downstream roadside sensors causes speedto drop at the upstream sensor while remain constant ata downstream sensorOn the other hand the same disruption

may cause flow to drop at both upstream and downstreamsensors Furthermore the same disruption can cause thespeed patterns to dropwithin seconds under a scenariowherevehicles are moving at relatively high speeds In contrastit may take place in several minutes in another scenariodepending on traffic flow and driversrsquo behaviors Also theshape of speed drop patterns may be linear (with differentslopes) exponential logarithmic or even sigmoid as foundin [24] To address these challenges Dynamic TimeWarpingis used in the proposed algorithm

Dynamic Time Warping (DTW) is a pattern similar-ity measurement between time series proposed in [33] toovercome limitations of traditional Euclidean distance byintroducing warping axis to adjust the error measurementFigure 3 shows distance measurement between two timeseries patterns 119881 and 119880 using DTW It can be seen inFigure 3(a) that at first 119881 and 119880 do not appear to be similarby using Euclidean distance measurement to compare pointsat the same time However as shown in Figures 3(b) and3(c) DTWoffers amore flexible and efficient approachwheresimilarity between two time series patterns can be assessedby comparing points at different times This is equivalent toldquowrappingrdquo the time series before comparing them

In practice the DTW would operate in an algorithmmodule in Figure 1 to assess similarity between an inputpattern 119881 measured by a roadside sensor and individualbasis pattern 119880 Given an input time window size 119871 acomputer programming language is used to implementDTW

Journal of Advanced Transportation 5

Anomaly detection (II)

Delay estimation (III IV)

Data cleansing (I)

Spatial estimation (III)

Temporal estimation (IV)

Sensor 1

Sensor 2

Sensor 3

I I I I

II II II II

Traffi

c flow

V2

V3

VM

V2 V1V3VM

middotmiddotmiddot

V1 = mnminusL+1 mnminus1 mn

Vm= mnminusL+1 mnminus1 mn

m = 1 2 M

middot middot middot

Sensor M

Traffic data of sensor m attime n using time window L

Predictionhorizon P

Estimated delay outcome dn

Figure 2 Overview of the proposed algorithm

V

U

(a)

V

U

(b)

V

U

(c)

Figure 3 An example of distance measurements between two time series using Dynamic Time Warping [26] (a) Euclidean distancemeasurement between two time series (b) DTW attempts to warp in time axis to find the correct measurement (c) Distance measurementusing DTW

by constructing a 119871-by-119871 matrix where the number of rowsand columns are equivalent to the length of 119881 and 119880respectively Each matrix element (119894 119895) corresponds to thealignment between points V119894 and 119906119895 Then a warping path119882 = 1199081 1199082 119908119870 119871 le 119870 lt 2119871 minus 1 is constructedas a set of matrix elements that defines a point mapping inorder to measure the distance between 119881 and 119880 Similarity

between 119881 and 119880 can be calculated based on these warpingpaths

Based onDTW the proposed algorithm is flexible enoughto cope with these pattern differences and subsequentlycompare 119881 and 119880 more efficiently If the input pattern 119881 isfound to be similar to the basis pattern 119880 associated withanomalies the proposed algorithm declares that an anomaly

6 Journal of Advanced Transportation

Traffi

c flow

Start sensor (1st sensor where anomaly is detected)

Intermediate sensor (anomaly detected)

Intermediate sensor (anomaly detected)

End sensor (1st sensor where anomaly is not detected)

Location where a lane is blocked (eg by an accident or a disabled vehicle)

Figure 4 An illustration of how the proposed algorithm finds start intermediate and end sensors from speed patterns measured by roadsidesensors Start sensor is the first downstream sensor that detects an anomaly End sensor is the first upstream sensor that an anomaly is notdetected Intermediate sensors are used to confirm that the congested area has extended upstream from start sensor to end sensor

is detected and proceeds to delay estimation We note thatanomalies can also be detected bymultiple sensors which canbe used in the delay estimation operation

44 Delay Estimation Delay estimation operation consistsof two consecutive steps (1) estimation of spatial extentof anomalies 119897119899119875 and (2) estimation of temporal extent 119889119899as shown in blocks III and IV in Figure 2 respectivelyThe spatial extent of delay 119897119899119875 is the length of the roadsegment that experiences anomalies which can be estimatedby finding a set of consecutive sensors where anomalies aredetected in Section 43 As shown in Figure 4 the roadsidesensors considered to be in the disruption area consist ofthe start sensor the intermediate sensors and the end sensorThe start sensor is the first downstream sensor where ananomaly is detected that is it is the location where delay isoriginated The intermediate sensors are the sensors locatedconsecutively upstream from the start sensor that also detectsanomalies that is they signify the continuous upstreamextension of anomalies The end sensor is the first upstreamsensor where an anomaly is not detected that is this isthe location where the disruption ceases to exist and trafficupstream from this location is still under normal conditionThen 119897119899119875 is estimated as the distance between the start andthe end sensors

The second step temporal estimation of delay in blockIV in Figure 2 uses 119897119899119875 as input to estimate delay vehiclesare expected to experience on the road segment Usingthe definition in [27] 119889119899 the estimated delay at time 119899 iscalculated as the estimated additional travel time that exceedsthe normal travel time as shown in (1) where 120591119899119875 is theestimated travel time and 120591119899119873 is the normal travel timeon the same road segment The estimated travel time 120591119899119875is calculated by dividing 119897119899119875 with the estimated mean ofspeeds measured at the start intermediate and end sensorsV119899119875

The normal travel time 120591119899119873 is calculated by dividing 119897119899119875with the normal speed V119899119873 We note that V119899119873 is calculatedas the mean of measured speeds in the previous timewindow (size 119871) just before an anomaly is detected in thecurrent time window (also size 119871)Therefore V119899119873 reflects theexpected speed drivers on the freeway segment would haveexperienced if an anomaly had not occurred

119889119899 = 120591119899119875 minus 120591119899119873 120591119899119875 =119897119899119875V119899119875 120591119899119873 =

119897119899119875V119899119873 (1)

The estimated mean of speeds V119899119875 can be calculated asfollows Based on our previous study in [24] and furtheranalysis of real-world data over a 6-month period it isobserved that the occurrences of delay are associatedwith fiveanomalous speed drop patterns namely linear exponentiallogarithmic sigmoid and unclassified Furthermore it ispossible to find empirical equations that best fit each of thesefive speed drop patterns (eg an exponential equation thatbest fits the exponential speed drop patterns) [24] A kernelfunction is used for the unclassified speed drop patterns

For a given prediction horizon 119875 and a total of119872 startintermediate and end sensors these empirical equations canbe used to temporally extrapolate speeds well into the futurefor each roadside sensor119898 that is [V119898119899+1 V119898119899+2 V119898119899+119875]Then V119899119875 can be obtained by first averaging temporallyextrapolated speed samples V119898119899119875 and then averaging spa-tially across all119872 sensors as follows

V119899119875 =sum119872119898=1 V119898119899119875119872 V119898119899119875 =

sum119875119894=1 V119898119899+119894119875 (2)

In practice delay estimation can be operated as followsOnce delay has been detected speed patterns measured byeach roadside sensor are measured and transmitted to thealgorithmmodule in Figure 1 UsingDynamicTimeWarping

Journal of Advanced Transportation 7

Daokanong

Suksawat Sathupradit Sathupradit River side

Port

1 2 3 4 5 6 7 8 9 10 11

644m1977m1078m393m390m786m1337m671m1505m1152m

2nd stage2nd stage

Figure 5 A sketch of expressway segment where real-world data was obtained

these speed patterns are compared with the basis speedpatterns associated with anomalies 119880119860119898 which have beencollected and stored a priori to find similarity Once themost similar basis speed drop pattern has been found itscorresponding empirical equation is used to calculate theestimated mean speed V119899119875 as shown in (2) which is thenused to calculate 120591119899119875 and eventually to estimate delay 119889119899 in(1) Note that the calculation in (1) is for short-term delay[3] Extending (1) to estimate vehiclersquos hour of delay or totaldelay experienced by all delayed vehicles on the road segmentcan be performed by adding traffic volume as amultiplicationfactor to (1)

5 Performance Evaluations

In this section performance evaluations are conducted usingreal-world data First Section 51 presents two real-worlddata sets used to assess the proposed algorithm ThenSection 52 describes an existing algorithm [17] which isused as benchmark In Section 53 performance evaluationparameters are presented

51 Descriptions of Real-World Data Sets

511 Real-World Data from Roadside Camera Sensors Theproposed algorithm is assessed using real-world data col-lected from a series of camera sensors installed on heavilyused inbound section ofChalerm-Maha-NakhonExpresswayin Bangkok Thailand These sensors were part of a projectjointly initiated by the Expressway Authority of Thailand(EXAT) and National Electronic and Computer TechnologyCenter (NECTEC) aiming at real-time traffic monitoringIt consists of a series of eleven camera sensors installedconsecutively along the 11 km expressway Figure 5 shows thetopology of the expressway and the locations of the camerastations where the distance between two consecutive camerasensors ranges from 644 to 1977 meters Each camera sensoris capable of detecting individual vehicles that passed theexpressway segment and produced time series of traffic dataincluding speed occupancy and flow in 1-minute intervals[34] Each camera sensor is connected to a local EXAT oper-ation office via wireless cellular network technologies (egGPRS and 3G) where operators can view real-time trafficinformation and make decisions on traffic management

A road segment within the full of view of a video camerais used as a designated area to calculate traffic data inputSpeed is measured as a distance in a vehiclersquos traverse in the

Figure 6 An image of a traffic delay occurrence captured at the 4thcamera sensor

designated area divided by the time it takes Flow ismeasuredas the rate of vehicles passing a designated area at a given timeinterval Occupancy is measured as a fraction of time that adesignated area is occupied by vehicles

The data set used in this paper was collected from thissensor network over a one year period from January 1 toDecember 31 2010 In this paper we chose the 3rd 4th 5thand 6th sensors as the experimental sites An example of traf-fic congestion viewed by a camera sensor studied in this paperis shown in Figure 6 For evaluation purposes anomaliesand their associated delay that took place were independentlyand manually logged by a team of transportation expertsA total of 16 nonrecurring congestion cases were recordedwhich are used to test the anomaly detection capability ofthe proposed algorithm Each record was associated with thetime instances prior to the disruption occurrence and after ithad ended Besides these 16 cases further investigation withanother team of experts also revealed additional 44 delaycases Therefore a total of 60 delay cases are used to test theestimation operation of the proposed algorithm

We note that the traffic data input of the proposedalgorithm is typical data that most roadside sensors forexample loop detectors already produce Therefore weanticipate that if the sensors are to be replaced with othertypes of sensors the performance of the system should notchange significantly This is worth further investigation forlarge scale deployment

512 NGSIM Data To show the potential of being deployedat a different location the proposed algorithm is also assessedusingNGSIM data [19]The selected data set was collected onthe southbound US Highway 101 in Los Angeles California

8 Journal of Advanced Transportation

USA at 0750 amndash0835 am This data set is selected becauseit was collected during the buildup of congestion and consistsof shockwaves which can potentially induce additional delayIt should be noted that the anomaly cases and the amountof delay caused by them were not labeled in this data setso the full algorithm could not be deployed for this dataset However the potential of the algorithm could be shownby shockwave detection To assess the proposed algorithmfour virtual sensors are placed approximately 01 km apart toobtain traffic data to capture the occurrence of shockwavesand their spatial extents Similar to conventional fixed trafficsensors (eg camera sensors and loop detectors) thesevirtual sensors calculate traffic data based on traffic flow andspeed that pass the sensorsrsquo locations An input time windowsize 119871 of 5 minutes is used where time series of 30-secondaverage speeds obtained from upstream and downstreamvirtual sensors are used as inputs The occurrences of shock-waves are observed and recorded by a team of transportationresearchers [25]

52 Benchmark Algorithm To assess the capability of theproposed algorithm in respect to existing approaches abenchmark algorithm is used [17]This benchmark algorithmis selected because it has a similar objective to the proposedalgorithm in detecting anomalies that cause delay Also thisalgorithm has already been implemented in a real-worldanomaly detection system [17] Furthermore composed ofa forecasting component and a detection logic componentit is designed to assess the difference between traffic dataobserved at consecutive roadside sensors

However the main difference between the benchmarkalgorithm and the proposed algorithm lies in the detectionlogic The benchmark algorithm is designed to detect sud-den significant changes of traffic inputs from the forecastcomponent values by establishing threshold values (speedand occupancy) to filter out gradual changes in traffic inputscaused by recurring congestion and an anomalous signalwill only be raised when the specified threshold values havebeen exceeded On the other hand the proposed algorithm isdesigned to detect both sudden and gradual changes whichenables it to detect different types of anomalies with fewerfalse alarms This difference will become more apparentwhen both algorithms are assessed with real-world data inSection 6

53 Performance Evaluation Parameters To numericallyevaluate the proposed algorithm we use detection rate (DR)false positive rate (FPR) mean time to detection (MTTD)andmean absolute percentage error (MAPE) as shown in (3)(4) (5) and (6) respectively DR FPR and MTTD are usedto assess the anomaly detection operation of the proposedalgorithm MAPE is used to assess the accuracy of delayestimation of a total 119863 recorded delay cases where the errorof the estimation 119889119899119895 of the 119895th recorded delay at time 119899 isnormalized by the actual delay119889119886119899119895 tominimize the possibilityof some estimated delay samples from dominating the others

DR =Number of delays detectedTotal number of delays

(3)

FPR

=Number of detected delays that are false positives

Total number of detected delays

(4)

MTTD

=Sum of time taken to detect individual delays

Total number of detected delays

(5)

MAPE = 1119863

119863

sum119895=1

10038161003816100381610038161003816119889119886119899119895 minus 119889119899119895

10038161003816100381610038161003816119889119886119899119895 (6)

The performance parameters are applicable for data setsthat contain labels for traffic anomalies and the amount ofdelay caused by the anomalies Therefore these parameterswill be evaluated for the real-world data collected inThailandbut due to the lack of anomaly labels could not be applied forthe NGSIM data set

6 Results and Discussions

In this section the proposed algorithm is assessed usingreal-world data described in Section 51 and compared tothe benchmark algorithm described in Section 52 Bothalgorithms are tested as if they operate online where eachdetection operation is performed per input time window 119871

61 Assessment Using NGSIM Data The proposed algorithmis assessed on its capability to detect the the presence ofshockwaves which refers to a phenomenon where vehicleshave to temporarily slow down below normal speed andsubsequently force the vehicles behind to slow down furtherThis phenomenon is known to create shockwave fronts ormoving speed drop areas that move backward disruptingflow and annoying drivers behind [35] The occurrence ofshockwaves during congestion not only causes additionaldelay but also increases chances of rear-end collisionsIdentifying shockwaves during congestion also presents achallenge as it occurs temporarily and requires an algorithmthat is fast and sensitive enough to detect them [25] Theoccurrence of shockwaves represents unexpected events sothey are considered as anomalies in this paper

Figure 7 shows that the proposed algorithm can detectall four occurrences (Figure 7(a)) of shockwaves whilebenchmark algorithm can detect only two (Figure 7(b))As shown in Figure 7(b) only abrupt changes in speeddifference (second and third occurrences) can be detectedby the benchmark algorithm This algorithm is designedto detect consecutive differences between upstream anddownstream traffic measurements so it operates well withconsecutive changes in speed difference that are significantenough On the other hand benchmark algorithmmisses thefirst and fourth occurrences of shockwaves in Figure 7(b) asthey are associated with relatively more gradual changes ofspeed difference Unlike benchmark algorithm the proposedalgorithm uses DTW to assess input patterns to detect delayoccurrences so it can detect all four occurrences associated

Journal of Advanced Transportation 9

Upstream sensorDownstream sensor

805 820 835750Time (hoursminutes)

01020304050607080

Spee

d (k

mh

)

(a)

One time step aheadTwo time steps ahead

805 820 835750Time (hoursminutes)

minus3minus2minus1

0123

Spee

d di

ffere

nces

(km

h)

(b)

Figure 7 Detecting delay occurrences associated with shockwaves in NGSIM data using (a) the proposed algorithm and (b) benchmarkalgorithm [17] Dashed vertical lines denote the time points where the algorithms detect delay occurrences Red squares denote the durationwhere the occurrences of shockwaves are observed

with both abrupt and gradual speed changes as shown inFigure 7(a)

Once the occurrences of shockwaves are detected theproposed algorithm proceeds to estimate delay as describedin Section 44 Using the spatial extent estimation describedin Section 44 out of four sensors three sensors are usuallyfound to detect speed drops associated with shockwaves at atime Therefore the most downstream sensor and the mostupstream sensor are selected as the start and end sensorsrespectively and 119897119899119875 is found to be 048 km

Table 1 shows the values of V119899119875 V119899119873 120591119899119875 and 120591119899119873 in (1)for each of the four shockwave occurrences For comparisonpurposes delay estimation is also applied for the second andthird occurrences that are detected by benchmark algorithm(see Figure 7) Note that since benchmark algorithm cannotdetect the first and fourth occurrences estimation operationis not activated during these times and normal speed andtravel times are used instead We further note that all fouroccurrences and their associated travel times and delays canbe calculated consecutively online by the proposed algorithm

As shown in Table 1 the four occurrences of shockwavesdetected in Figure 7 are associated with travel times (120591119899119875 in(1)) of 5682 s 5554 s 6098 s and 5379 s respectively Theproposed algorithm is able to detect all four occurrences andthese estimated travel times would all be included in the delaycalculation On the other hand if the benchmark algorithmwas used instead only the second and third occurrences andtheir associated travel times would be reported resulting inless accurate delay calculationWe note that even though thedifference between the delays obtained from the proposedalgorithm and benchmark algorithm in this example maybe small significant improvement can be seen when higherdegree of disruptions are assessed

The results shown in Table 1 have many applicationsin advanced traveler information systems and advancedtraffic management systems A direct application would beto estimate travel time of the freeway segment and report itto motorists The ability to detect occurrences of shockwavescan also be used in other applications such as calculatingshockwave propagation time [25] It can also be used to warn

motorists to adjust intervehicle spacing to avoid rear-endcollisions

62 Assessment Using Real-World Data Collected in ThailandThis section assesses the capability of the proposed algorithmand the benchmark algorithm [17] to detect the sixteenrecorded traffic anomalies in the real-world data set describedin Section 51 As inputs both algorithms use speed flowand occupancy collected by roadside sensorsThe assessmentfollows a 119896-fold cross-validation technique The real-worlddata set consists of 16 anomaly cases 1198631 1198632 11986316 soevaluation is performed 16 times At an iteration 119894 119894 =1 2 16 the anomaly case119863119894 is reserved as a test set whilethe other anomaly cases are collectively used to calibratethe proposed algorithm and benchmark algorithm That isin the first iteration 1198632 11986316 collectively serve as thetraining set which is then tested with1198631Then in the seconditeration 1198631 1198633 11986316 are used for training while 1198632 isused for testing and so on

Table 2 shows that the proposed algorithm achieveshigher detection rate and lower false alarm rate than thebenchmark algorithm Particularly compared to benchmarkalgorithm the proposed algorithm improves the detectionrate and false positive rate by approximately 10 and 7respectively Figures 8 and 9 show examples of anomaliesdetected by the proposed algorithm and found to causetraffic delay Based on the records these anomalies areassociated with disabled vehicles parking over the shoulderarea

It is found that for the benchmark algorithm miss-detection usually occurs under low traffic flow This algo-rithm always needs data from at least two sensors and themagnitude of traffic flow measured by the roadside sensorsis not large enough to trigger the benchmark algorithmWhile medium to heavy traffic flow triggers the benchmarkalgorithm it also increases false alarms This algorithm isdesigned to directly assess the difference between upstreamand downstream sensors of the roadside sensors whichsubsequently increases its sensitivity and triggers many falsealarms

10 Journal of Advanced Transportation

Table 1 Estimation results onNGSIMdata using the proposed algorithm and benchmark algorithm 119897119899119875 =048 kmNote that if the shockwaveis not detected the estimated travel time 120591119899119875 would be equal to normal travel time 120591119899119873

1st 2nd 3rd 4thProposed algorithmShockwaves detected Yes Yes Yes YesV119899119875 (see (1)) 3041 kmh 3112 kmh 2834 kmh 3213 kmh120591119899119875 (see (1)) 5682 s 5554 s 6098 s 5379 s119889119899 = 120591119899119875 minus 120591119899119873 1841 s 1893 s 2399 s 541 sBenchmark algorithm [17]Shockwaves detected No Yes Yes NoV119899119875 (see (1)) 4499 kmh 3112 kmh 2834 kmh 3572 kmh120591119899119875 (see (1)) 3841 s 5554 s 6098 s 4838 s119889119899 = 120591119899119875 minus 120591119899119873 000 s 1893 s 2399 s 000 sIf all 4 shockwaves are not detectedV119899119873 (see (1)) 4499 kmh 4721 kmh 4671 kmh 3572 kmh120591119899119873 (see (1)) 3841 s 3660 s 3700 s 4838 s

Table 2 Results on anomaly detection using data set collected inThailand

Performanceparameters

Proposedalgorithm

Benchmarkalgorithm [17]

DR () 94 83FPR () 5 12MTTD (seconds) 340 140

Figure 8 An anomaly detected at the 4th camera sensor

Unlike the benchmark algorithm the proposed algorithmuses DTW to assess traffic data from each sensor individuallyThis enables the proposed algorithm to recognize patternsassociated with anomalies more effectively including underlow traffic flow The use of DTW also reduces a numberof false alarms as it ensures that unrecognized patternswill not trigger detection at individual sensors In respectto mean time to detection (MTTD) it takes longer forthe proposed algorithm to detect anomalies compared tobenchmark algorithm as it needs to wait till enough temporalsamples of traffic data have been collected in each sensorbefore the detection operation can be triggered Howeverthe mean time to detection is still less than 6 minutes which

Figure 9 Another anomaly detected at the 4th camera sensor

should give enough time for traffic operators to response [8]To reduce mean time to detection of the proposed algorithmshorter time samples of traffic data can be used ReducingMTTD while maintaining acceptable levels of DR and FPRis worth further investigation

To deploy the proposed algorithm computation timeof the algorithm could be an added delay to the MTTDEven though the computation time was not systematicallymeasured for the performance evaluation the computationtime was in order of seconds For the computerrsquos specificationthat we use which was a personal laptop with 25 GHz CPUand 4GB of RAM the computation time was approximately[1 3) seconds As the computing time (a few seconds) wasmuch smaller than the mean time to detection (up to 6minutes) it is reasonable to consider the computing timenegligible

This section further discusses the results when the pro-posed algorithm was applied to estimate 60 delay casesdescribed in Section 51 using different input time windowsizes (119871) and prediction horizons (119875) For practical purposesit is important to select suitable input time window sizeand prediction horizon to enable the proposed method to

Journal of Advanced Transportation 11

L = 3 L = 10L = 5

P = 5

P = 10

P = 15

0

005

01

015

02

025

03

035

04M

APE

Figure 10Mean absolute percentage error (MAPE) fromestimating60 delay cases using different input time window sizes 119871 (minutes)and prediction horizons 119875 (minutes)

perform effectively Even though the 15-minute interval isoften recommended [9] to avoid strong fluctuation usinglower input timewindow sizes can providemore fine-grainedinformation for practitioners [25] Therefore in this paperwe use 119871 = 3 5 and 10 minutes for input time windowsize For prediction horizons we use 119875 = 5 10 and 15minutes as this paper is focused on short-term estimation(less than 15 minutes) Also changes in traffic condition onthis experimental site are normally found to occur in less than15 minutes

The proposed algorithm is assessed on its capability toestimate delay using MAPE Figure 10 shows MAPE fromestimating 60 delay cases in the real-world data set withdifferent input timewindow sizes (119871) and prediction horizons(119875) These results show that the estimation part of theproposed algorithm can achieve MAPE lower than 01 forthis set of real-world data This accuracy is achieved becausethe estimation part consists of both spatial and temporalestimations which enable the proposed algorithm to firstestimate the area disrupted by anomalies and then estimatethe delay itself Furthermore it can be seen that increasinginput time window size (119871) reduces MAPE as using moreinput samples enhances estimation accuracy For the impactof prediction horizons (119875) MAPE tends to increase withlarger prediction horizons as it is more difficult to estimatedelays farther ahead into the future However as shownin Figure 10 when 119875 is increased MAPE in fact increasesat a much smaller rate when larger values of 119871 are usedTherefore it is recommended for practitioners that if theobjective is to estimate delays far ahead into the future (usinglarge prediction horizons) larger input time window size isneeded

Based on the findings in this section it is importantthat practitioners select a suitable input time window sizeaccording to their practical objectives If the objective is toprovide prompt delay information small input time windowsize should be used On the other hand if the objective is toprovide accurately estimated delays for traffic managementpurposes larger input time window sizes should be chosenFurthermore based on the proposed algorithm a morecomplex and flexible system can also be further developedFor example a stepwise approach can be implemented on topof the proposed algorithm where a small input time windowsize is used as a trigger to promptly discover the occurrenceof traffic delay Then in the next step the input time windowsize can be increased to collect more traffic data to achievemore accurate detection and estimation of traffic delay

For anomalies with large scale effect beyond the 11 kmexperimental site we anticipate that the proposed algo-rithm should be able to detect those anomalies if theyoccur downstream and cause changes in traffic patterns topropagate upstream into the 11 km segment To accuratelyestimate delay basis patterns associated with such large scaleanomalies need to be incorporated to distinguish themThenan extrapolation technique needs to be incorporated to enablethe proposed algorithm to estimate delay beyond the 11 kmexpressway segment where local traffic information cannotbe directly observed [25] This is worth further investigation

7 Conclusions and Future Work

This paper proposes an algorithm to detect anomalies andestimate delay using traffic data from roadside sensors Theproposed algorithm uses traffic data from individual sensorsto detect anomalies and then uses data from consecutivedownstream and upstream sensors together to spatially andtemporally estimate delay Performance evaluations usingreal-world data show that the proposed algorithm achieveshigher detection rate and lower false positive rate when usedto detect anomalies compared to an existing benchmarkalgorithmAn investigationwithNGSIMdata also shows thatthe proposed algorithm is also able to detect the presence ofshockwaves and subsequently estimate delaymore efficientlythan benchmark algorithm Further results show that theestimation error increases at a much smaller rate whensuitable input time window size and prediction horizon areselected

There are interesting aspects that are worth furtherinvestigation One interesting aspect is to reduce the pro-cessing time while maintaining low detection and estima-tion errors A possible approach is to introduce a featureextraction method that can extract hidden patterns whensmaller input time window sizes are used [8] Alternativeapproaches include studying a set of hidden representativetraffic flow features in advance by pretraining a deep-learningmodel [36] and populating traffic features into a geneticalgorithm learning method to reduce dimensionality [37]Another interesting aspect is to improve inference capabilityby incorporating a spatial inferencemethod to estimate trafficdata itself at the locations where they cannot be directlymeasured [25] Also it would be practically interesting to

12 Journal of Advanced Transportation

assess the proposed algorithm with different road types andgeometries

In addition to improvements of the algorithm an interest-ing aspect to investigate is the applications of the algorithmThe solution provided in this paper is designed to be astreaming algorithm which can be run online thereforeit can be naturally extended to use in various real-timeapplications For example the detected anomaly and esti-mated additional travel time would be useful for both trafficoperators and travelers if the information is delivered ina timely and convenient manner Active communicationbetween roadside sensors and vehicle-to-infrastructure (V2I)could be used to promptly transmit the information from oursolution to nearby roadside message boards to display theinformation to the drivers on the roads or vehicles with V2Icommunication to adapt routing and navigation

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this article

Acknowledgments

The authors would like to thankDr Supakorn Siddhichai andhis IMG team at NECTEC for their assistance in providingreal-world data for the analysis

References

[1] N Parrado and Y Donoso ldquoCongestion based mechanism forroute discovery in a V2I-V2V system applying smart devicesand IoTrdquo Sensors vol 15 no 4 pp 7768ndash7806 2015

[2] K Nellore and G P Hancke ldquoA survey on urban trafficmanagement system using wireless sensor networksrdquo Sensorsvol 16 no 2 article 157 2016

[3] E I Vlahogianni M G Karlaftis and J C Golias ldquoShort-term traffic forecasting where we are and where wersquore goingrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 3ndash19 2014

[4] R Li and G Rose ldquoIncorporating uncertainty into short-term travel time predictionsrdquo Transportation Research Part CEmerging Technologies vol 19 no 6 pp 1006ndash1018 2011

[5] Y Chung ldquoAmethodological approach for estimating temporaland spatial extent of delays caused by freeway accidentsrdquo IEEETransactions on Intelligent Transportation Systems vol 13 no 3pp 1454ndash1461 2012

[6] J Weng and Q Meng ldquoEstimating capacity and traffic delayin work zones an overviewrdquo Transportation ResearchmdashPart CEmerging Technologies vol 35 pp 34ndash45 2013

[7] X Zhang E Onieva A Perallos E Osaba and V C SLee ldquoHierarchical fuzzy rule-based system optimized withgenetic algorithms for short term traffic congestion predictionrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 127ndash142 2014

[8] S Thajchayapong E S Garcia-Trevino and J A Barria ldquoDis-tributed classification of traffic anomalies using microscopictraffic variablesrdquo IEEETransactions on Intelligent TransportationSystems vol 14 no 1 pp 448ndash458 2013

[9] M Lippi M Bertini and P Frasconi ldquoShort-term traffic flowforecasting an experimental comparison of time-series anal-ysis and supervised learningrdquo IEEE Transactions on IntelligentTransportation Systems vol 14 no 2 pp 871ndash882 2013

[10] F Zheng andH Van Zuylen ldquoUrban link travel time estimationbased on sparse probe vehicle datardquo Transportation ResearchmdashPart C Emerging Technologies vol 31 pp 145ndash157 2013

[11] F Dion and H Rakha ldquoEstimating dynamic roadway traveltimes using automatic vehicle identification data for low sam-pling ratesrdquoTransportation Research Part BMethodological vol40 no 9 pp 745ndash766 2006

[12] D Paterson and G Rose ldquoDynamic travel time estima-tion on instrumented freewaysrdquo in Proceedings of the 6thWorld Congress on Intelligent Transportation Systems TorontoCanada 1999

[13] J Kwon B Coifman and P Bickel ldquoDay-to-day travel-timetrends and travel-time prediction from loop-detector datardquoTransportation Research Record no 1717 pp 120ndash129 2000

[14] J Rice and E Van Zwet ldquoA simple and effective method forpredicting travel times on freewaysrdquo IEEE Transactions onIntelligent Transportation Systems vol 5 no 3 pp 200ndash2072004

[15] C D R Lindveld RThijs P H L Bovy andN J van der ZijppldquoEvaluation of online travel time estimators and predictorsrdquoTransportation Research Record vol 1719 pp 45ndash53 2000

[16] H Yang S Chen YWang BWu andZWang ldquoComparison ofdelay estimation models for signalised intersections using fieldobservations in Shanghairdquo IET Intelligent Transport Systems vol10 no 3 pp 165ndash174 2016

[17] C L Mak and H S L Fan ldquoDevelopment of dual-stationautomated expressway incident detection algorithmsrdquo IEEETransactions on Intelligent Transportation Systems vol 8 no 3pp 480ndash490 2007

[18] K Hi-Ri-O-Tappa C Likitkhajorn A Poolsawat and S Tha-jchayapong ldquoTraffic incident detection system using series ofpoint detectorsrdquo in Proceedings of the 15th International IEEEConference on Intelligent Transportation Systems (ITSC rsquo12) pp182ndash187 Anchorage Alaska USA September 2012

[19] V Alexiadis J Colyar J Halkias R Hranac and G McHaleldquoThe next generation simulation programrdquo ITE Journal vol 74no 8 pp 22ndash26 2004

[20] F G Habtemichael M Cetin and K A Anuar ldquoIncident-induced delays on freeways quantificationmethod by groupingsimilar traffic patternsrdquo Transportation Research Record vol2484 pp 60ndash69 2015

[21] J Li C-J Lan and X Gu ldquoEstimation of incident delay andits uncertainty on freeway networksrdquo Transportation ResearchRecord vol 1959 pp 37ndash45 2006

[22] Y Chung ldquoAssessment of non-recurrent congestion causedby precipitation using archived weather and traffic flow datardquoTransport Policy vol 19 no 1 pp 167ndash173 2012

[23] J A Barria and SThajchayapong ldquoDetection and classificationof traffic anomalies using microscopic traffic variablesrdquo IEEETransactions on Intelligent Transportation Systems vol 12 no3 pp 695ndash704 2011

[24] K HiriOtappa and S Thajchayapong ldquoGeneralizability andtransferability of incident detection algorithm using dynamictime warpingrdquo in Proceedings of the in 19th World Congress onIntelligent Transportation Systems Vienna Austria 2012

Journal of Advanced Transportation 13

[25] S Thajchayapong and J A Barria ldquoSpatial inference of traffictransition using microndashmacro traffic variablesrdquo IEEE Transac-tions on Intelligent Transportation Systems vol 16 no 2 pp854ndash864 2015

[26] C A Ratanamahatana and E Keogh ldquoEverything you knowabout dynamic time warping is wrongrdquo in Proceedings of the3rd Workshop on Mining Temporal and Sequential Data in Con-junction with the 10th ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining (KDD rsquo04) CiteseerSeattle Wash USA August 2004

[27] Transportation Research Thesauras Transportation ResearchBoard National Academy httptrttrborgtrtaspNN=Bthd

[28] N Chiabaut C Buisson and L Leclercq ldquoFundamental dia-gram estimation through passing rate measurements in conges-tionrdquo IEEE Transactions on Intelligent Transportation Systemsvol 10 no 2 pp 355ndash359 2009

[29] J J Fernandez-Lozano M Martın-Guzman J Martın-Avilaand A Garcıa-Cerezo ldquoA wireless sensor network for urbantraffic characterization and trend monitoringrdquo Sensors vol 15no 10 pp 26143ndash26169 2015

[30] J-S Oh C Oh S G Ritchie and M Chang ldquoReal-timeestimation of accident likelihood for safety enhancementrdquoJournal of Transportation Engineering vol 131 no 5 pp 358ndash363 2005

[31] A P Chassiakos and Y J Stephanedes ldquoSmoothing algorithmsfor incident detectionrdquo Transportation Research Record vol1394 pp 8ndash16 1993

[32] V Alarcon-Aquino and J A Barria ldquoAnomaly detection incommunication networks using waveletsrdquo IEE ProceedingsmdashCommunications vol 148 no 6 pp 355ndash362 2001

[33] D Berndt and J Clifford ldquoUsing dynamic time warping tofind patterns in time seriesrdquo in Proceedings of the Workshop onKnowledge Discovery in Databases (KDD rsquo94) vol 10 pp 359ndash370 Seattle Wash USA August 1994

[34] K Kiratiratanapruk and S Siddhichai ldquoVehicle detection andtracking for traffic monitoring systemrdquo in Proceedings of theIEEE Region 10 Conference (TENCON rsquo06) pp 1ndash4 IEEE HongKong November 2006

[35] A Hegyi B De Schutter and J Hellendoorn ldquoOptimal coordi-nation of variable speed limits to suppress shock wavesrdquo IEEETransactions on Intelligent Transportation Systems vol 6 no 1pp 102ndash112 2005

[36] Y Lv Y Duan W Kang Z Li and F-Y Wang ldquoTraffic flowprediction with big data a deep learning approachrdquo IEEETransactions on Intelligent Transportation Systems vol 16 no2 pp 865ndash873 2015

[37] P Lopez-Garcia E Onieva E Osaba A D Masegosa and APerallos ldquoA hybrid method for short-term traffic congestionforecasting using genetic algorithms and cross entropyrdquo IEEETransactions on Intelligent Transportation Systems vol 17 no 2pp 557ndash569 2016

4 Journal of Advanced Transportation

Advanced traffic

management system

advanced traveler

information system

Sensor 1Sensor unit

Communication unit

AlgorithmData collection module

system

Traffi

c flow

Sensor 2Sensor unit

Communication unit

Sensor 3Sensor unit

Communication unit

Sensor nSensor unit

Communication unit

Figure 1 How the proposed algorithm can be deployed with roadside sensors on a freeway segmentThe proposed algorithm is the algorithmmodule which uses traffic data collected from sensors to detect anomalies and estimate delay

exponentialmoving average derived fromDELOS smoothing[31] is used to interpolate missing data and remove noiseThe exponentialmoving average is chosen particularly to givemore weights to the latest speed samples as they are moreassociated with delay occurrence

The anomaly detection operation is derived based onthe sliding window approach in [32] but modified to assesseach input time window individually whether it is associ-ated with normal or anomalous traffic patterns The pro-posed algorithm uses pattern matching where each real-time input pattern consisting of 119871 samples of traffic data119881119898119899 = [V119898119899 V119898119899minus1 V119898119899minus119871+1] from roadside sensorsThe input 119881119898119899 is identified as normal or anomalous byassessing its similarity with a set of basis patterns 119880119898119899 =[119906119898119899 119906119898119899minus1 119906119898119899minus119871+1] consisting of normal patterns 119880119873119898and anomalous patterns 119880119860119898 An anomaly is detected if theinput pattern 119881119898119899 is more similar to 119880119860119898 than to 119880119873119898 Inpractice 119881119898119899 and 119880119898119899 can be speed flow and occupancymeasured from both normal and anomalous conditions apriori and stored in a database

The main challenge for anomaly detection is that anoma-lies can cause speed flow and occupancy patterns to behavedifferently in time space and shape For example it isobserved that a disruption occurring between a pair ofupstream and downstream roadside sensors causes speedto drop at the upstream sensor while remain constant ata downstream sensorOn the other hand the same disruption

may cause flow to drop at both upstream and downstreamsensors Furthermore the same disruption can cause thespeed patterns to dropwithin seconds under a scenariowherevehicles are moving at relatively high speeds In contrastit may take place in several minutes in another scenariodepending on traffic flow and driversrsquo behaviors Also theshape of speed drop patterns may be linear (with differentslopes) exponential logarithmic or even sigmoid as foundin [24] To address these challenges Dynamic TimeWarpingis used in the proposed algorithm

Dynamic Time Warping (DTW) is a pattern similar-ity measurement between time series proposed in [33] toovercome limitations of traditional Euclidean distance byintroducing warping axis to adjust the error measurementFigure 3 shows distance measurement between two timeseries patterns 119881 and 119880 using DTW It can be seen inFigure 3(a) that at first 119881 and 119880 do not appear to be similarby using Euclidean distance measurement to compare pointsat the same time However as shown in Figures 3(b) and3(c) DTWoffers amore flexible and efficient approachwheresimilarity between two time series patterns can be assessedby comparing points at different times This is equivalent toldquowrappingrdquo the time series before comparing them

In practice the DTW would operate in an algorithmmodule in Figure 1 to assess similarity between an inputpattern 119881 measured by a roadside sensor and individualbasis pattern 119880 Given an input time window size 119871 acomputer programming language is used to implementDTW

Journal of Advanced Transportation 5

Anomaly detection (II)

Delay estimation (III IV)

Data cleansing (I)

Spatial estimation (III)

Temporal estimation (IV)

Sensor 1

Sensor 2

Sensor 3

I I I I

II II II II

Traffi

c flow

V2

V3

VM

V2 V1V3VM

middotmiddotmiddot

V1 = mnminusL+1 mnminus1 mn

Vm= mnminusL+1 mnminus1 mn

m = 1 2 M

middot middot middot

Sensor M

Traffic data of sensor m attime n using time window L

Predictionhorizon P

Estimated delay outcome dn

Figure 2 Overview of the proposed algorithm

V

U

(a)

V

U

(b)

V

U

(c)

Figure 3 An example of distance measurements between two time series using Dynamic Time Warping [26] (a) Euclidean distancemeasurement between two time series (b) DTW attempts to warp in time axis to find the correct measurement (c) Distance measurementusing DTW

by constructing a 119871-by-119871 matrix where the number of rowsand columns are equivalent to the length of 119881 and 119880respectively Each matrix element (119894 119895) corresponds to thealignment between points V119894 and 119906119895 Then a warping path119882 = 1199081 1199082 119908119870 119871 le 119870 lt 2119871 minus 1 is constructedas a set of matrix elements that defines a point mapping inorder to measure the distance between 119881 and 119880 Similarity

between 119881 and 119880 can be calculated based on these warpingpaths

Based onDTW the proposed algorithm is flexible enoughto cope with these pattern differences and subsequentlycompare 119881 and 119880 more efficiently If the input pattern 119881 isfound to be similar to the basis pattern 119880 associated withanomalies the proposed algorithm declares that an anomaly

6 Journal of Advanced Transportation

Traffi

c flow

Start sensor (1st sensor where anomaly is detected)

Intermediate sensor (anomaly detected)

Intermediate sensor (anomaly detected)

End sensor (1st sensor where anomaly is not detected)

Location where a lane is blocked (eg by an accident or a disabled vehicle)

Figure 4 An illustration of how the proposed algorithm finds start intermediate and end sensors from speed patterns measured by roadsidesensors Start sensor is the first downstream sensor that detects an anomaly End sensor is the first upstream sensor that an anomaly is notdetected Intermediate sensors are used to confirm that the congested area has extended upstream from start sensor to end sensor

is detected and proceeds to delay estimation We note thatanomalies can also be detected bymultiple sensors which canbe used in the delay estimation operation

44 Delay Estimation Delay estimation operation consistsof two consecutive steps (1) estimation of spatial extentof anomalies 119897119899119875 and (2) estimation of temporal extent 119889119899as shown in blocks III and IV in Figure 2 respectivelyThe spatial extent of delay 119897119899119875 is the length of the roadsegment that experiences anomalies which can be estimatedby finding a set of consecutive sensors where anomalies aredetected in Section 43 As shown in Figure 4 the roadsidesensors considered to be in the disruption area consist ofthe start sensor the intermediate sensors and the end sensorThe start sensor is the first downstream sensor where ananomaly is detected that is it is the location where delay isoriginated The intermediate sensors are the sensors locatedconsecutively upstream from the start sensor that also detectsanomalies that is they signify the continuous upstreamextension of anomalies The end sensor is the first upstreamsensor where an anomaly is not detected that is this isthe location where the disruption ceases to exist and trafficupstream from this location is still under normal conditionThen 119897119899119875 is estimated as the distance between the start andthe end sensors

The second step temporal estimation of delay in blockIV in Figure 2 uses 119897119899119875 as input to estimate delay vehiclesare expected to experience on the road segment Usingthe definition in [27] 119889119899 the estimated delay at time 119899 iscalculated as the estimated additional travel time that exceedsthe normal travel time as shown in (1) where 120591119899119875 is theestimated travel time and 120591119899119873 is the normal travel timeon the same road segment The estimated travel time 120591119899119875is calculated by dividing 119897119899119875 with the estimated mean ofspeeds measured at the start intermediate and end sensorsV119899119875

The normal travel time 120591119899119873 is calculated by dividing 119897119899119875with the normal speed V119899119873 We note that V119899119873 is calculatedas the mean of measured speeds in the previous timewindow (size 119871) just before an anomaly is detected in thecurrent time window (also size 119871)Therefore V119899119873 reflects theexpected speed drivers on the freeway segment would haveexperienced if an anomaly had not occurred

119889119899 = 120591119899119875 minus 120591119899119873 120591119899119875 =119897119899119875V119899119875 120591119899119873 =

119897119899119875V119899119873 (1)

The estimated mean of speeds V119899119875 can be calculated asfollows Based on our previous study in [24] and furtheranalysis of real-world data over a 6-month period it isobserved that the occurrences of delay are associatedwith fiveanomalous speed drop patterns namely linear exponentiallogarithmic sigmoid and unclassified Furthermore it ispossible to find empirical equations that best fit each of thesefive speed drop patterns (eg an exponential equation thatbest fits the exponential speed drop patterns) [24] A kernelfunction is used for the unclassified speed drop patterns

For a given prediction horizon 119875 and a total of119872 startintermediate and end sensors these empirical equations canbe used to temporally extrapolate speeds well into the futurefor each roadside sensor119898 that is [V119898119899+1 V119898119899+2 V119898119899+119875]Then V119899119875 can be obtained by first averaging temporallyextrapolated speed samples V119898119899119875 and then averaging spa-tially across all119872 sensors as follows

V119899119875 =sum119872119898=1 V119898119899119875119872 V119898119899119875 =

sum119875119894=1 V119898119899+119894119875 (2)

In practice delay estimation can be operated as followsOnce delay has been detected speed patterns measured byeach roadside sensor are measured and transmitted to thealgorithmmodule in Figure 1 UsingDynamicTimeWarping

Journal of Advanced Transportation 7

Daokanong

Suksawat Sathupradit Sathupradit River side

Port

1 2 3 4 5 6 7 8 9 10 11

644m1977m1078m393m390m786m1337m671m1505m1152m

2nd stage2nd stage

Figure 5 A sketch of expressway segment where real-world data was obtained

these speed patterns are compared with the basis speedpatterns associated with anomalies 119880119860119898 which have beencollected and stored a priori to find similarity Once themost similar basis speed drop pattern has been found itscorresponding empirical equation is used to calculate theestimated mean speed V119899119875 as shown in (2) which is thenused to calculate 120591119899119875 and eventually to estimate delay 119889119899 in(1) Note that the calculation in (1) is for short-term delay[3] Extending (1) to estimate vehiclersquos hour of delay or totaldelay experienced by all delayed vehicles on the road segmentcan be performed by adding traffic volume as amultiplicationfactor to (1)

5 Performance Evaluations

In this section performance evaluations are conducted usingreal-world data First Section 51 presents two real-worlddata sets used to assess the proposed algorithm ThenSection 52 describes an existing algorithm [17] which isused as benchmark In Section 53 performance evaluationparameters are presented

51 Descriptions of Real-World Data Sets

511 Real-World Data from Roadside Camera Sensors Theproposed algorithm is assessed using real-world data col-lected from a series of camera sensors installed on heavilyused inbound section ofChalerm-Maha-NakhonExpresswayin Bangkok Thailand These sensors were part of a projectjointly initiated by the Expressway Authority of Thailand(EXAT) and National Electronic and Computer TechnologyCenter (NECTEC) aiming at real-time traffic monitoringIt consists of a series of eleven camera sensors installedconsecutively along the 11 km expressway Figure 5 shows thetopology of the expressway and the locations of the camerastations where the distance between two consecutive camerasensors ranges from 644 to 1977 meters Each camera sensoris capable of detecting individual vehicles that passed theexpressway segment and produced time series of traffic dataincluding speed occupancy and flow in 1-minute intervals[34] Each camera sensor is connected to a local EXAT oper-ation office via wireless cellular network technologies (egGPRS and 3G) where operators can view real-time trafficinformation and make decisions on traffic management

A road segment within the full of view of a video camerais used as a designated area to calculate traffic data inputSpeed is measured as a distance in a vehiclersquos traverse in the

Figure 6 An image of a traffic delay occurrence captured at the 4thcamera sensor

designated area divided by the time it takes Flow ismeasuredas the rate of vehicles passing a designated area at a given timeinterval Occupancy is measured as a fraction of time that adesignated area is occupied by vehicles

The data set used in this paper was collected from thissensor network over a one year period from January 1 toDecember 31 2010 In this paper we chose the 3rd 4th 5thand 6th sensors as the experimental sites An example of traf-fic congestion viewed by a camera sensor studied in this paperis shown in Figure 6 For evaluation purposes anomaliesand their associated delay that took place were independentlyand manually logged by a team of transportation expertsA total of 16 nonrecurring congestion cases were recordedwhich are used to test the anomaly detection capability ofthe proposed algorithm Each record was associated with thetime instances prior to the disruption occurrence and after ithad ended Besides these 16 cases further investigation withanother team of experts also revealed additional 44 delaycases Therefore a total of 60 delay cases are used to test theestimation operation of the proposed algorithm

We note that the traffic data input of the proposedalgorithm is typical data that most roadside sensors forexample loop detectors already produce Therefore weanticipate that if the sensors are to be replaced with othertypes of sensors the performance of the system should notchange significantly This is worth further investigation forlarge scale deployment

512 NGSIM Data To show the potential of being deployedat a different location the proposed algorithm is also assessedusingNGSIM data [19]The selected data set was collected onthe southbound US Highway 101 in Los Angeles California

8 Journal of Advanced Transportation

USA at 0750 amndash0835 am This data set is selected becauseit was collected during the buildup of congestion and consistsof shockwaves which can potentially induce additional delayIt should be noted that the anomaly cases and the amountof delay caused by them were not labeled in this data setso the full algorithm could not be deployed for this dataset However the potential of the algorithm could be shownby shockwave detection To assess the proposed algorithmfour virtual sensors are placed approximately 01 km apart toobtain traffic data to capture the occurrence of shockwavesand their spatial extents Similar to conventional fixed trafficsensors (eg camera sensors and loop detectors) thesevirtual sensors calculate traffic data based on traffic flow andspeed that pass the sensorsrsquo locations An input time windowsize 119871 of 5 minutes is used where time series of 30-secondaverage speeds obtained from upstream and downstreamvirtual sensors are used as inputs The occurrences of shock-waves are observed and recorded by a team of transportationresearchers [25]

52 Benchmark Algorithm To assess the capability of theproposed algorithm in respect to existing approaches abenchmark algorithm is used [17]This benchmark algorithmis selected because it has a similar objective to the proposedalgorithm in detecting anomalies that cause delay Also thisalgorithm has already been implemented in a real-worldanomaly detection system [17] Furthermore composed ofa forecasting component and a detection logic componentit is designed to assess the difference between traffic dataobserved at consecutive roadside sensors

However the main difference between the benchmarkalgorithm and the proposed algorithm lies in the detectionlogic The benchmark algorithm is designed to detect sud-den significant changes of traffic inputs from the forecastcomponent values by establishing threshold values (speedand occupancy) to filter out gradual changes in traffic inputscaused by recurring congestion and an anomalous signalwill only be raised when the specified threshold values havebeen exceeded On the other hand the proposed algorithm isdesigned to detect both sudden and gradual changes whichenables it to detect different types of anomalies with fewerfalse alarms This difference will become more apparentwhen both algorithms are assessed with real-world data inSection 6

53 Performance Evaluation Parameters To numericallyevaluate the proposed algorithm we use detection rate (DR)false positive rate (FPR) mean time to detection (MTTD)andmean absolute percentage error (MAPE) as shown in (3)(4) (5) and (6) respectively DR FPR and MTTD are usedto assess the anomaly detection operation of the proposedalgorithm MAPE is used to assess the accuracy of delayestimation of a total 119863 recorded delay cases where the errorof the estimation 119889119899119895 of the 119895th recorded delay at time 119899 isnormalized by the actual delay119889119886119899119895 tominimize the possibilityof some estimated delay samples from dominating the others

DR =Number of delays detectedTotal number of delays

(3)

FPR

=Number of detected delays that are false positives

Total number of detected delays

(4)

MTTD

=Sum of time taken to detect individual delays

Total number of detected delays

(5)

MAPE = 1119863

119863

sum119895=1

10038161003816100381610038161003816119889119886119899119895 minus 119889119899119895

10038161003816100381610038161003816119889119886119899119895 (6)

The performance parameters are applicable for data setsthat contain labels for traffic anomalies and the amount ofdelay caused by the anomalies Therefore these parameterswill be evaluated for the real-world data collected inThailandbut due to the lack of anomaly labels could not be applied forthe NGSIM data set

6 Results and Discussions

In this section the proposed algorithm is assessed usingreal-world data described in Section 51 and compared tothe benchmark algorithm described in Section 52 Bothalgorithms are tested as if they operate online where eachdetection operation is performed per input time window 119871

61 Assessment Using NGSIM Data The proposed algorithmis assessed on its capability to detect the the presence ofshockwaves which refers to a phenomenon where vehicleshave to temporarily slow down below normal speed andsubsequently force the vehicles behind to slow down furtherThis phenomenon is known to create shockwave fronts ormoving speed drop areas that move backward disruptingflow and annoying drivers behind [35] The occurrence ofshockwaves during congestion not only causes additionaldelay but also increases chances of rear-end collisionsIdentifying shockwaves during congestion also presents achallenge as it occurs temporarily and requires an algorithmthat is fast and sensitive enough to detect them [25] Theoccurrence of shockwaves represents unexpected events sothey are considered as anomalies in this paper

Figure 7 shows that the proposed algorithm can detectall four occurrences (Figure 7(a)) of shockwaves whilebenchmark algorithm can detect only two (Figure 7(b))As shown in Figure 7(b) only abrupt changes in speeddifference (second and third occurrences) can be detectedby the benchmark algorithm This algorithm is designedto detect consecutive differences between upstream anddownstream traffic measurements so it operates well withconsecutive changes in speed difference that are significantenough On the other hand benchmark algorithmmisses thefirst and fourth occurrences of shockwaves in Figure 7(b) asthey are associated with relatively more gradual changes ofspeed difference Unlike benchmark algorithm the proposedalgorithm uses DTW to assess input patterns to detect delayoccurrences so it can detect all four occurrences associated

Journal of Advanced Transportation 9

Upstream sensorDownstream sensor

805 820 835750Time (hoursminutes)

01020304050607080

Spee

d (k

mh

)

(a)

One time step aheadTwo time steps ahead

805 820 835750Time (hoursminutes)

minus3minus2minus1

0123

Spee

d di

ffere

nces

(km

h)

(b)

Figure 7 Detecting delay occurrences associated with shockwaves in NGSIM data using (a) the proposed algorithm and (b) benchmarkalgorithm [17] Dashed vertical lines denote the time points where the algorithms detect delay occurrences Red squares denote the durationwhere the occurrences of shockwaves are observed

with both abrupt and gradual speed changes as shown inFigure 7(a)

Once the occurrences of shockwaves are detected theproposed algorithm proceeds to estimate delay as describedin Section 44 Using the spatial extent estimation describedin Section 44 out of four sensors three sensors are usuallyfound to detect speed drops associated with shockwaves at atime Therefore the most downstream sensor and the mostupstream sensor are selected as the start and end sensorsrespectively and 119897119899119875 is found to be 048 km

Table 1 shows the values of V119899119875 V119899119873 120591119899119875 and 120591119899119873 in (1)for each of the four shockwave occurrences For comparisonpurposes delay estimation is also applied for the second andthird occurrences that are detected by benchmark algorithm(see Figure 7) Note that since benchmark algorithm cannotdetect the first and fourth occurrences estimation operationis not activated during these times and normal speed andtravel times are used instead We further note that all fouroccurrences and their associated travel times and delays canbe calculated consecutively online by the proposed algorithm

As shown in Table 1 the four occurrences of shockwavesdetected in Figure 7 are associated with travel times (120591119899119875 in(1)) of 5682 s 5554 s 6098 s and 5379 s respectively Theproposed algorithm is able to detect all four occurrences andthese estimated travel times would all be included in the delaycalculation On the other hand if the benchmark algorithmwas used instead only the second and third occurrences andtheir associated travel times would be reported resulting inless accurate delay calculationWe note that even though thedifference between the delays obtained from the proposedalgorithm and benchmark algorithm in this example maybe small significant improvement can be seen when higherdegree of disruptions are assessed

The results shown in Table 1 have many applicationsin advanced traveler information systems and advancedtraffic management systems A direct application would beto estimate travel time of the freeway segment and report itto motorists The ability to detect occurrences of shockwavescan also be used in other applications such as calculatingshockwave propagation time [25] It can also be used to warn

motorists to adjust intervehicle spacing to avoid rear-endcollisions

62 Assessment Using Real-World Data Collected in ThailandThis section assesses the capability of the proposed algorithmand the benchmark algorithm [17] to detect the sixteenrecorded traffic anomalies in the real-world data set describedin Section 51 As inputs both algorithms use speed flowand occupancy collected by roadside sensorsThe assessmentfollows a 119896-fold cross-validation technique The real-worlddata set consists of 16 anomaly cases 1198631 1198632 11986316 soevaluation is performed 16 times At an iteration 119894 119894 =1 2 16 the anomaly case119863119894 is reserved as a test set whilethe other anomaly cases are collectively used to calibratethe proposed algorithm and benchmark algorithm That isin the first iteration 1198632 11986316 collectively serve as thetraining set which is then tested with1198631Then in the seconditeration 1198631 1198633 11986316 are used for training while 1198632 isused for testing and so on

Table 2 shows that the proposed algorithm achieveshigher detection rate and lower false alarm rate than thebenchmark algorithm Particularly compared to benchmarkalgorithm the proposed algorithm improves the detectionrate and false positive rate by approximately 10 and 7respectively Figures 8 and 9 show examples of anomaliesdetected by the proposed algorithm and found to causetraffic delay Based on the records these anomalies areassociated with disabled vehicles parking over the shoulderarea

It is found that for the benchmark algorithm miss-detection usually occurs under low traffic flow This algo-rithm always needs data from at least two sensors and themagnitude of traffic flow measured by the roadside sensorsis not large enough to trigger the benchmark algorithmWhile medium to heavy traffic flow triggers the benchmarkalgorithm it also increases false alarms This algorithm isdesigned to directly assess the difference between upstreamand downstream sensors of the roadside sensors whichsubsequently increases its sensitivity and triggers many falsealarms

10 Journal of Advanced Transportation

Table 1 Estimation results onNGSIMdata using the proposed algorithm and benchmark algorithm 119897119899119875 =048 kmNote that if the shockwaveis not detected the estimated travel time 120591119899119875 would be equal to normal travel time 120591119899119873

1st 2nd 3rd 4thProposed algorithmShockwaves detected Yes Yes Yes YesV119899119875 (see (1)) 3041 kmh 3112 kmh 2834 kmh 3213 kmh120591119899119875 (see (1)) 5682 s 5554 s 6098 s 5379 s119889119899 = 120591119899119875 minus 120591119899119873 1841 s 1893 s 2399 s 541 sBenchmark algorithm [17]Shockwaves detected No Yes Yes NoV119899119875 (see (1)) 4499 kmh 3112 kmh 2834 kmh 3572 kmh120591119899119875 (see (1)) 3841 s 5554 s 6098 s 4838 s119889119899 = 120591119899119875 minus 120591119899119873 000 s 1893 s 2399 s 000 sIf all 4 shockwaves are not detectedV119899119873 (see (1)) 4499 kmh 4721 kmh 4671 kmh 3572 kmh120591119899119873 (see (1)) 3841 s 3660 s 3700 s 4838 s

Table 2 Results on anomaly detection using data set collected inThailand

Performanceparameters

Proposedalgorithm

Benchmarkalgorithm [17]

DR () 94 83FPR () 5 12MTTD (seconds) 340 140

Figure 8 An anomaly detected at the 4th camera sensor

Unlike the benchmark algorithm the proposed algorithmuses DTW to assess traffic data from each sensor individuallyThis enables the proposed algorithm to recognize patternsassociated with anomalies more effectively including underlow traffic flow The use of DTW also reduces a numberof false alarms as it ensures that unrecognized patternswill not trigger detection at individual sensors In respectto mean time to detection (MTTD) it takes longer forthe proposed algorithm to detect anomalies compared tobenchmark algorithm as it needs to wait till enough temporalsamples of traffic data have been collected in each sensorbefore the detection operation can be triggered Howeverthe mean time to detection is still less than 6 minutes which

Figure 9 Another anomaly detected at the 4th camera sensor

should give enough time for traffic operators to response [8]To reduce mean time to detection of the proposed algorithmshorter time samples of traffic data can be used ReducingMTTD while maintaining acceptable levels of DR and FPRis worth further investigation

To deploy the proposed algorithm computation timeof the algorithm could be an added delay to the MTTDEven though the computation time was not systematicallymeasured for the performance evaluation the computationtime was in order of seconds For the computerrsquos specificationthat we use which was a personal laptop with 25 GHz CPUand 4GB of RAM the computation time was approximately[1 3) seconds As the computing time (a few seconds) wasmuch smaller than the mean time to detection (up to 6minutes) it is reasonable to consider the computing timenegligible

This section further discusses the results when the pro-posed algorithm was applied to estimate 60 delay casesdescribed in Section 51 using different input time windowsizes (119871) and prediction horizons (119875) For practical purposesit is important to select suitable input time window sizeand prediction horizon to enable the proposed method to

Journal of Advanced Transportation 11

L = 3 L = 10L = 5

P = 5

P = 10

P = 15

0

005

01

015

02

025

03

035

04M

APE

Figure 10Mean absolute percentage error (MAPE) fromestimating60 delay cases using different input time window sizes 119871 (minutes)and prediction horizons 119875 (minutes)

perform effectively Even though the 15-minute interval isoften recommended [9] to avoid strong fluctuation usinglower input timewindow sizes can providemore fine-grainedinformation for practitioners [25] Therefore in this paperwe use 119871 = 3 5 and 10 minutes for input time windowsize For prediction horizons we use 119875 = 5 10 and 15minutes as this paper is focused on short-term estimation(less than 15 minutes) Also changes in traffic condition onthis experimental site are normally found to occur in less than15 minutes

The proposed algorithm is assessed on its capability toestimate delay using MAPE Figure 10 shows MAPE fromestimating 60 delay cases in the real-world data set withdifferent input timewindow sizes (119871) and prediction horizons(119875) These results show that the estimation part of theproposed algorithm can achieve MAPE lower than 01 forthis set of real-world data This accuracy is achieved becausethe estimation part consists of both spatial and temporalestimations which enable the proposed algorithm to firstestimate the area disrupted by anomalies and then estimatethe delay itself Furthermore it can be seen that increasinginput time window size (119871) reduces MAPE as using moreinput samples enhances estimation accuracy For the impactof prediction horizons (119875) MAPE tends to increase withlarger prediction horizons as it is more difficult to estimatedelays farther ahead into the future However as shownin Figure 10 when 119875 is increased MAPE in fact increasesat a much smaller rate when larger values of 119871 are usedTherefore it is recommended for practitioners that if theobjective is to estimate delays far ahead into the future (usinglarge prediction horizons) larger input time window size isneeded

Based on the findings in this section it is importantthat practitioners select a suitable input time window sizeaccording to their practical objectives If the objective is toprovide prompt delay information small input time windowsize should be used On the other hand if the objective is toprovide accurately estimated delays for traffic managementpurposes larger input time window sizes should be chosenFurthermore based on the proposed algorithm a morecomplex and flexible system can also be further developedFor example a stepwise approach can be implemented on topof the proposed algorithm where a small input time windowsize is used as a trigger to promptly discover the occurrenceof traffic delay Then in the next step the input time windowsize can be increased to collect more traffic data to achievemore accurate detection and estimation of traffic delay

For anomalies with large scale effect beyond the 11 kmexperimental site we anticipate that the proposed algo-rithm should be able to detect those anomalies if theyoccur downstream and cause changes in traffic patterns topropagate upstream into the 11 km segment To accuratelyestimate delay basis patterns associated with such large scaleanomalies need to be incorporated to distinguish themThenan extrapolation technique needs to be incorporated to enablethe proposed algorithm to estimate delay beyond the 11 kmexpressway segment where local traffic information cannotbe directly observed [25] This is worth further investigation

7 Conclusions and Future Work

This paper proposes an algorithm to detect anomalies andestimate delay using traffic data from roadside sensors Theproposed algorithm uses traffic data from individual sensorsto detect anomalies and then uses data from consecutivedownstream and upstream sensors together to spatially andtemporally estimate delay Performance evaluations usingreal-world data show that the proposed algorithm achieveshigher detection rate and lower false positive rate when usedto detect anomalies compared to an existing benchmarkalgorithmAn investigationwithNGSIMdata also shows thatthe proposed algorithm is also able to detect the presence ofshockwaves and subsequently estimate delaymore efficientlythan benchmark algorithm Further results show that theestimation error increases at a much smaller rate whensuitable input time window size and prediction horizon areselected

There are interesting aspects that are worth furtherinvestigation One interesting aspect is to reduce the pro-cessing time while maintaining low detection and estima-tion errors A possible approach is to introduce a featureextraction method that can extract hidden patterns whensmaller input time window sizes are used [8] Alternativeapproaches include studying a set of hidden representativetraffic flow features in advance by pretraining a deep-learningmodel [36] and populating traffic features into a geneticalgorithm learning method to reduce dimensionality [37]Another interesting aspect is to improve inference capabilityby incorporating a spatial inferencemethod to estimate trafficdata itself at the locations where they cannot be directlymeasured [25] Also it would be practically interesting to

12 Journal of Advanced Transportation

assess the proposed algorithm with different road types andgeometries

In addition to improvements of the algorithm an interest-ing aspect to investigate is the applications of the algorithmThe solution provided in this paper is designed to be astreaming algorithm which can be run online thereforeit can be naturally extended to use in various real-timeapplications For example the detected anomaly and esti-mated additional travel time would be useful for both trafficoperators and travelers if the information is delivered ina timely and convenient manner Active communicationbetween roadside sensors and vehicle-to-infrastructure (V2I)could be used to promptly transmit the information from oursolution to nearby roadside message boards to display theinformation to the drivers on the roads or vehicles with V2Icommunication to adapt routing and navigation

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this article

Acknowledgments

The authors would like to thankDr Supakorn Siddhichai andhis IMG team at NECTEC for their assistance in providingreal-world data for the analysis

References

[1] N Parrado and Y Donoso ldquoCongestion based mechanism forroute discovery in a V2I-V2V system applying smart devicesand IoTrdquo Sensors vol 15 no 4 pp 7768ndash7806 2015

[2] K Nellore and G P Hancke ldquoA survey on urban trafficmanagement system using wireless sensor networksrdquo Sensorsvol 16 no 2 article 157 2016

[3] E I Vlahogianni M G Karlaftis and J C Golias ldquoShort-term traffic forecasting where we are and where wersquore goingrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 3ndash19 2014

[4] R Li and G Rose ldquoIncorporating uncertainty into short-term travel time predictionsrdquo Transportation Research Part CEmerging Technologies vol 19 no 6 pp 1006ndash1018 2011

[5] Y Chung ldquoAmethodological approach for estimating temporaland spatial extent of delays caused by freeway accidentsrdquo IEEETransactions on Intelligent Transportation Systems vol 13 no 3pp 1454ndash1461 2012

[6] J Weng and Q Meng ldquoEstimating capacity and traffic delayin work zones an overviewrdquo Transportation ResearchmdashPart CEmerging Technologies vol 35 pp 34ndash45 2013

[7] X Zhang E Onieva A Perallos E Osaba and V C SLee ldquoHierarchical fuzzy rule-based system optimized withgenetic algorithms for short term traffic congestion predictionrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 127ndash142 2014

[8] S Thajchayapong E S Garcia-Trevino and J A Barria ldquoDis-tributed classification of traffic anomalies using microscopictraffic variablesrdquo IEEETransactions on Intelligent TransportationSystems vol 14 no 1 pp 448ndash458 2013

[9] M Lippi M Bertini and P Frasconi ldquoShort-term traffic flowforecasting an experimental comparison of time-series anal-ysis and supervised learningrdquo IEEE Transactions on IntelligentTransportation Systems vol 14 no 2 pp 871ndash882 2013

[10] F Zheng andH Van Zuylen ldquoUrban link travel time estimationbased on sparse probe vehicle datardquo Transportation ResearchmdashPart C Emerging Technologies vol 31 pp 145ndash157 2013

[11] F Dion and H Rakha ldquoEstimating dynamic roadway traveltimes using automatic vehicle identification data for low sam-pling ratesrdquoTransportation Research Part BMethodological vol40 no 9 pp 745ndash766 2006

[12] D Paterson and G Rose ldquoDynamic travel time estima-tion on instrumented freewaysrdquo in Proceedings of the 6thWorld Congress on Intelligent Transportation Systems TorontoCanada 1999

[13] J Kwon B Coifman and P Bickel ldquoDay-to-day travel-timetrends and travel-time prediction from loop-detector datardquoTransportation Research Record no 1717 pp 120ndash129 2000

[14] J Rice and E Van Zwet ldquoA simple and effective method forpredicting travel times on freewaysrdquo IEEE Transactions onIntelligent Transportation Systems vol 5 no 3 pp 200ndash2072004

[15] C D R Lindveld RThijs P H L Bovy andN J van der ZijppldquoEvaluation of online travel time estimators and predictorsrdquoTransportation Research Record vol 1719 pp 45ndash53 2000

[16] H Yang S Chen YWang BWu andZWang ldquoComparison ofdelay estimation models for signalised intersections using fieldobservations in Shanghairdquo IET Intelligent Transport Systems vol10 no 3 pp 165ndash174 2016

[17] C L Mak and H S L Fan ldquoDevelopment of dual-stationautomated expressway incident detection algorithmsrdquo IEEETransactions on Intelligent Transportation Systems vol 8 no 3pp 480ndash490 2007

[18] K Hi-Ri-O-Tappa C Likitkhajorn A Poolsawat and S Tha-jchayapong ldquoTraffic incident detection system using series ofpoint detectorsrdquo in Proceedings of the 15th International IEEEConference on Intelligent Transportation Systems (ITSC rsquo12) pp182ndash187 Anchorage Alaska USA September 2012

[19] V Alexiadis J Colyar J Halkias R Hranac and G McHaleldquoThe next generation simulation programrdquo ITE Journal vol 74no 8 pp 22ndash26 2004

[20] F G Habtemichael M Cetin and K A Anuar ldquoIncident-induced delays on freeways quantificationmethod by groupingsimilar traffic patternsrdquo Transportation Research Record vol2484 pp 60ndash69 2015

[21] J Li C-J Lan and X Gu ldquoEstimation of incident delay andits uncertainty on freeway networksrdquo Transportation ResearchRecord vol 1959 pp 37ndash45 2006

[22] Y Chung ldquoAssessment of non-recurrent congestion causedby precipitation using archived weather and traffic flow datardquoTransport Policy vol 19 no 1 pp 167ndash173 2012

[23] J A Barria and SThajchayapong ldquoDetection and classificationof traffic anomalies using microscopic traffic variablesrdquo IEEETransactions on Intelligent Transportation Systems vol 12 no3 pp 695ndash704 2011

[24] K HiriOtappa and S Thajchayapong ldquoGeneralizability andtransferability of incident detection algorithm using dynamictime warpingrdquo in Proceedings of the in 19th World Congress onIntelligent Transportation Systems Vienna Austria 2012

Journal of Advanced Transportation 13

[25] S Thajchayapong and J A Barria ldquoSpatial inference of traffictransition using microndashmacro traffic variablesrdquo IEEE Transac-tions on Intelligent Transportation Systems vol 16 no 2 pp854ndash864 2015

[26] C A Ratanamahatana and E Keogh ldquoEverything you knowabout dynamic time warping is wrongrdquo in Proceedings of the3rd Workshop on Mining Temporal and Sequential Data in Con-junction with the 10th ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining (KDD rsquo04) CiteseerSeattle Wash USA August 2004

[27] Transportation Research Thesauras Transportation ResearchBoard National Academy httptrttrborgtrtaspNN=Bthd

[28] N Chiabaut C Buisson and L Leclercq ldquoFundamental dia-gram estimation through passing rate measurements in conges-tionrdquo IEEE Transactions on Intelligent Transportation Systemsvol 10 no 2 pp 355ndash359 2009

[29] J J Fernandez-Lozano M Martın-Guzman J Martın-Avilaand A Garcıa-Cerezo ldquoA wireless sensor network for urbantraffic characterization and trend monitoringrdquo Sensors vol 15no 10 pp 26143ndash26169 2015

[30] J-S Oh C Oh S G Ritchie and M Chang ldquoReal-timeestimation of accident likelihood for safety enhancementrdquoJournal of Transportation Engineering vol 131 no 5 pp 358ndash363 2005

[31] A P Chassiakos and Y J Stephanedes ldquoSmoothing algorithmsfor incident detectionrdquo Transportation Research Record vol1394 pp 8ndash16 1993

[32] V Alarcon-Aquino and J A Barria ldquoAnomaly detection incommunication networks using waveletsrdquo IEE ProceedingsmdashCommunications vol 148 no 6 pp 355ndash362 2001

[33] D Berndt and J Clifford ldquoUsing dynamic time warping tofind patterns in time seriesrdquo in Proceedings of the Workshop onKnowledge Discovery in Databases (KDD rsquo94) vol 10 pp 359ndash370 Seattle Wash USA August 1994

[34] K Kiratiratanapruk and S Siddhichai ldquoVehicle detection andtracking for traffic monitoring systemrdquo in Proceedings of theIEEE Region 10 Conference (TENCON rsquo06) pp 1ndash4 IEEE HongKong November 2006

[35] A Hegyi B De Schutter and J Hellendoorn ldquoOptimal coordi-nation of variable speed limits to suppress shock wavesrdquo IEEETransactions on Intelligent Transportation Systems vol 6 no 1pp 102ndash112 2005

[36] Y Lv Y Duan W Kang Z Li and F-Y Wang ldquoTraffic flowprediction with big data a deep learning approachrdquo IEEETransactions on Intelligent Transportation Systems vol 16 no2 pp 865ndash873 2015

[37] P Lopez-Garcia E Onieva E Osaba A D Masegosa and APerallos ldquoA hybrid method for short-term traffic congestionforecasting using genetic algorithms and cross entropyrdquo IEEETransactions on Intelligent Transportation Systems vol 17 no 2pp 557ndash569 2016

Journal of Advanced Transportation 5

Anomaly detection (II)

Delay estimation (III IV)

Data cleansing (I)

Spatial estimation (III)

Temporal estimation (IV)

Sensor 1

Sensor 2

Sensor 3

I I I I

II II II II

Traffi

c flow

V2

V3

VM

V2 V1V3VM

middotmiddotmiddot

V1 = mnminusL+1 mnminus1 mn

Vm= mnminusL+1 mnminus1 mn

m = 1 2 M

middot middot middot

Sensor M

Traffic data of sensor m attime n using time window L

Predictionhorizon P

Estimated delay outcome dn

Figure 2 Overview of the proposed algorithm

V

U

(a)

V

U

(b)

V

U

(c)

Figure 3 An example of distance measurements between two time series using Dynamic Time Warping [26] (a) Euclidean distancemeasurement between two time series (b) DTW attempts to warp in time axis to find the correct measurement (c) Distance measurementusing DTW

by constructing a 119871-by-119871 matrix where the number of rowsand columns are equivalent to the length of 119881 and 119880respectively Each matrix element (119894 119895) corresponds to thealignment between points V119894 and 119906119895 Then a warping path119882 = 1199081 1199082 119908119870 119871 le 119870 lt 2119871 minus 1 is constructedas a set of matrix elements that defines a point mapping inorder to measure the distance between 119881 and 119880 Similarity

between 119881 and 119880 can be calculated based on these warpingpaths

Based onDTW the proposed algorithm is flexible enoughto cope with these pattern differences and subsequentlycompare 119881 and 119880 more efficiently If the input pattern 119881 isfound to be similar to the basis pattern 119880 associated withanomalies the proposed algorithm declares that an anomaly

6 Journal of Advanced Transportation

Traffi

c flow

Start sensor (1st sensor where anomaly is detected)

Intermediate sensor (anomaly detected)

Intermediate sensor (anomaly detected)

End sensor (1st sensor where anomaly is not detected)

Location where a lane is blocked (eg by an accident or a disabled vehicle)

Figure 4 An illustration of how the proposed algorithm finds start intermediate and end sensors from speed patterns measured by roadsidesensors Start sensor is the first downstream sensor that detects an anomaly End sensor is the first upstream sensor that an anomaly is notdetected Intermediate sensors are used to confirm that the congested area has extended upstream from start sensor to end sensor

is detected and proceeds to delay estimation We note thatanomalies can also be detected bymultiple sensors which canbe used in the delay estimation operation

44 Delay Estimation Delay estimation operation consistsof two consecutive steps (1) estimation of spatial extentof anomalies 119897119899119875 and (2) estimation of temporal extent 119889119899as shown in blocks III and IV in Figure 2 respectivelyThe spatial extent of delay 119897119899119875 is the length of the roadsegment that experiences anomalies which can be estimatedby finding a set of consecutive sensors where anomalies aredetected in Section 43 As shown in Figure 4 the roadsidesensors considered to be in the disruption area consist ofthe start sensor the intermediate sensors and the end sensorThe start sensor is the first downstream sensor where ananomaly is detected that is it is the location where delay isoriginated The intermediate sensors are the sensors locatedconsecutively upstream from the start sensor that also detectsanomalies that is they signify the continuous upstreamextension of anomalies The end sensor is the first upstreamsensor where an anomaly is not detected that is this isthe location where the disruption ceases to exist and trafficupstream from this location is still under normal conditionThen 119897119899119875 is estimated as the distance between the start andthe end sensors

The second step temporal estimation of delay in blockIV in Figure 2 uses 119897119899119875 as input to estimate delay vehiclesare expected to experience on the road segment Usingthe definition in [27] 119889119899 the estimated delay at time 119899 iscalculated as the estimated additional travel time that exceedsthe normal travel time as shown in (1) where 120591119899119875 is theestimated travel time and 120591119899119873 is the normal travel timeon the same road segment The estimated travel time 120591119899119875is calculated by dividing 119897119899119875 with the estimated mean ofspeeds measured at the start intermediate and end sensorsV119899119875

The normal travel time 120591119899119873 is calculated by dividing 119897119899119875with the normal speed V119899119873 We note that V119899119873 is calculatedas the mean of measured speeds in the previous timewindow (size 119871) just before an anomaly is detected in thecurrent time window (also size 119871)Therefore V119899119873 reflects theexpected speed drivers on the freeway segment would haveexperienced if an anomaly had not occurred

119889119899 = 120591119899119875 minus 120591119899119873 120591119899119875 =119897119899119875V119899119875 120591119899119873 =

119897119899119875V119899119873 (1)

The estimated mean of speeds V119899119875 can be calculated asfollows Based on our previous study in [24] and furtheranalysis of real-world data over a 6-month period it isobserved that the occurrences of delay are associatedwith fiveanomalous speed drop patterns namely linear exponentiallogarithmic sigmoid and unclassified Furthermore it ispossible to find empirical equations that best fit each of thesefive speed drop patterns (eg an exponential equation thatbest fits the exponential speed drop patterns) [24] A kernelfunction is used for the unclassified speed drop patterns

For a given prediction horizon 119875 and a total of119872 startintermediate and end sensors these empirical equations canbe used to temporally extrapolate speeds well into the futurefor each roadside sensor119898 that is [V119898119899+1 V119898119899+2 V119898119899+119875]Then V119899119875 can be obtained by first averaging temporallyextrapolated speed samples V119898119899119875 and then averaging spa-tially across all119872 sensors as follows

V119899119875 =sum119872119898=1 V119898119899119875119872 V119898119899119875 =

sum119875119894=1 V119898119899+119894119875 (2)

In practice delay estimation can be operated as followsOnce delay has been detected speed patterns measured byeach roadside sensor are measured and transmitted to thealgorithmmodule in Figure 1 UsingDynamicTimeWarping

Journal of Advanced Transportation 7

Daokanong

Suksawat Sathupradit Sathupradit River side

Port

1 2 3 4 5 6 7 8 9 10 11

644m1977m1078m393m390m786m1337m671m1505m1152m

2nd stage2nd stage

Figure 5 A sketch of expressway segment where real-world data was obtained

these speed patterns are compared with the basis speedpatterns associated with anomalies 119880119860119898 which have beencollected and stored a priori to find similarity Once themost similar basis speed drop pattern has been found itscorresponding empirical equation is used to calculate theestimated mean speed V119899119875 as shown in (2) which is thenused to calculate 120591119899119875 and eventually to estimate delay 119889119899 in(1) Note that the calculation in (1) is for short-term delay[3] Extending (1) to estimate vehiclersquos hour of delay or totaldelay experienced by all delayed vehicles on the road segmentcan be performed by adding traffic volume as amultiplicationfactor to (1)

5 Performance Evaluations

In this section performance evaluations are conducted usingreal-world data First Section 51 presents two real-worlddata sets used to assess the proposed algorithm ThenSection 52 describes an existing algorithm [17] which isused as benchmark In Section 53 performance evaluationparameters are presented

51 Descriptions of Real-World Data Sets

511 Real-World Data from Roadside Camera Sensors Theproposed algorithm is assessed using real-world data col-lected from a series of camera sensors installed on heavilyused inbound section ofChalerm-Maha-NakhonExpresswayin Bangkok Thailand These sensors were part of a projectjointly initiated by the Expressway Authority of Thailand(EXAT) and National Electronic and Computer TechnologyCenter (NECTEC) aiming at real-time traffic monitoringIt consists of a series of eleven camera sensors installedconsecutively along the 11 km expressway Figure 5 shows thetopology of the expressway and the locations of the camerastations where the distance between two consecutive camerasensors ranges from 644 to 1977 meters Each camera sensoris capable of detecting individual vehicles that passed theexpressway segment and produced time series of traffic dataincluding speed occupancy and flow in 1-minute intervals[34] Each camera sensor is connected to a local EXAT oper-ation office via wireless cellular network technologies (egGPRS and 3G) where operators can view real-time trafficinformation and make decisions on traffic management

A road segment within the full of view of a video camerais used as a designated area to calculate traffic data inputSpeed is measured as a distance in a vehiclersquos traverse in the

Figure 6 An image of a traffic delay occurrence captured at the 4thcamera sensor

designated area divided by the time it takes Flow ismeasuredas the rate of vehicles passing a designated area at a given timeinterval Occupancy is measured as a fraction of time that adesignated area is occupied by vehicles

The data set used in this paper was collected from thissensor network over a one year period from January 1 toDecember 31 2010 In this paper we chose the 3rd 4th 5thand 6th sensors as the experimental sites An example of traf-fic congestion viewed by a camera sensor studied in this paperis shown in Figure 6 For evaluation purposes anomaliesand their associated delay that took place were independentlyand manually logged by a team of transportation expertsA total of 16 nonrecurring congestion cases were recordedwhich are used to test the anomaly detection capability ofthe proposed algorithm Each record was associated with thetime instances prior to the disruption occurrence and after ithad ended Besides these 16 cases further investigation withanother team of experts also revealed additional 44 delaycases Therefore a total of 60 delay cases are used to test theestimation operation of the proposed algorithm

We note that the traffic data input of the proposedalgorithm is typical data that most roadside sensors forexample loop detectors already produce Therefore weanticipate that if the sensors are to be replaced with othertypes of sensors the performance of the system should notchange significantly This is worth further investigation forlarge scale deployment

512 NGSIM Data To show the potential of being deployedat a different location the proposed algorithm is also assessedusingNGSIM data [19]The selected data set was collected onthe southbound US Highway 101 in Los Angeles California

8 Journal of Advanced Transportation

USA at 0750 amndash0835 am This data set is selected becauseit was collected during the buildup of congestion and consistsof shockwaves which can potentially induce additional delayIt should be noted that the anomaly cases and the amountof delay caused by them were not labeled in this data setso the full algorithm could not be deployed for this dataset However the potential of the algorithm could be shownby shockwave detection To assess the proposed algorithmfour virtual sensors are placed approximately 01 km apart toobtain traffic data to capture the occurrence of shockwavesand their spatial extents Similar to conventional fixed trafficsensors (eg camera sensors and loop detectors) thesevirtual sensors calculate traffic data based on traffic flow andspeed that pass the sensorsrsquo locations An input time windowsize 119871 of 5 minutes is used where time series of 30-secondaverage speeds obtained from upstream and downstreamvirtual sensors are used as inputs The occurrences of shock-waves are observed and recorded by a team of transportationresearchers [25]

52 Benchmark Algorithm To assess the capability of theproposed algorithm in respect to existing approaches abenchmark algorithm is used [17]This benchmark algorithmis selected because it has a similar objective to the proposedalgorithm in detecting anomalies that cause delay Also thisalgorithm has already been implemented in a real-worldanomaly detection system [17] Furthermore composed ofa forecasting component and a detection logic componentit is designed to assess the difference between traffic dataobserved at consecutive roadside sensors

However the main difference between the benchmarkalgorithm and the proposed algorithm lies in the detectionlogic The benchmark algorithm is designed to detect sud-den significant changes of traffic inputs from the forecastcomponent values by establishing threshold values (speedand occupancy) to filter out gradual changes in traffic inputscaused by recurring congestion and an anomalous signalwill only be raised when the specified threshold values havebeen exceeded On the other hand the proposed algorithm isdesigned to detect both sudden and gradual changes whichenables it to detect different types of anomalies with fewerfalse alarms This difference will become more apparentwhen both algorithms are assessed with real-world data inSection 6

53 Performance Evaluation Parameters To numericallyevaluate the proposed algorithm we use detection rate (DR)false positive rate (FPR) mean time to detection (MTTD)andmean absolute percentage error (MAPE) as shown in (3)(4) (5) and (6) respectively DR FPR and MTTD are usedto assess the anomaly detection operation of the proposedalgorithm MAPE is used to assess the accuracy of delayestimation of a total 119863 recorded delay cases where the errorof the estimation 119889119899119895 of the 119895th recorded delay at time 119899 isnormalized by the actual delay119889119886119899119895 tominimize the possibilityof some estimated delay samples from dominating the others

DR =Number of delays detectedTotal number of delays

(3)

FPR

=Number of detected delays that are false positives

Total number of detected delays

(4)

MTTD

=Sum of time taken to detect individual delays

Total number of detected delays

(5)

MAPE = 1119863

119863

sum119895=1

10038161003816100381610038161003816119889119886119899119895 minus 119889119899119895

10038161003816100381610038161003816119889119886119899119895 (6)

The performance parameters are applicable for data setsthat contain labels for traffic anomalies and the amount ofdelay caused by the anomalies Therefore these parameterswill be evaluated for the real-world data collected inThailandbut due to the lack of anomaly labels could not be applied forthe NGSIM data set

6 Results and Discussions

In this section the proposed algorithm is assessed usingreal-world data described in Section 51 and compared tothe benchmark algorithm described in Section 52 Bothalgorithms are tested as if they operate online where eachdetection operation is performed per input time window 119871

61 Assessment Using NGSIM Data The proposed algorithmis assessed on its capability to detect the the presence ofshockwaves which refers to a phenomenon where vehicleshave to temporarily slow down below normal speed andsubsequently force the vehicles behind to slow down furtherThis phenomenon is known to create shockwave fronts ormoving speed drop areas that move backward disruptingflow and annoying drivers behind [35] The occurrence ofshockwaves during congestion not only causes additionaldelay but also increases chances of rear-end collisionsIdentifying shockwaves during congestion also presents achallenge as it occurs temporarily and requires an algorithmthat is fast and sensitive enough to detect them [25] Theoccurrence of shockwaves represents unexpected events sothey are considered as anomalies in this paper

Figure 7 shows that the proposed algorithm can detectall four occurrences (Figure 7(a)) of shockwaves whilebenchmark algorithm can detect only two (Figure 7(b))As shown in Figure 7(b) only abrupt changes in speeddifference (second and third occurrences) can be detectedby the benchmark algorithm This algorithm is designedto detect consecutive differences between upstream anddownstream traffic measurements so it operates well withconsecutive changes in speed difference that are significantenough On the other hand benchmark algorithmmisses thefirst and fourth occurrences of shockwaves in Figure 7(b) asthey are associated with relatively more gradual changes ofspeed difference Unlike benchmark algorithm the proposedalgorithm uses DTW to assess input patterns to detect delayoccurrences so it can detect all four occurrences associated

Journal of Advanced Transportation 9

Upstream sensorDownstream sensor

805 820 835750Time (hoursminutes)

01020304050607080

Spee

d (k

mh

)

(a)

One time step aheadTwo time steps ahead

805 820 835750Time (hoursminutes)

minus3minus2minus1

0123

Spee

d di

ffere

nces

(km

h)

(b)

Figure 7 Detecting delay occurrences associated with shockwaves in NGSIM data using (a) the proposed algorithm and (b) benchmarkalgorithm [17] Dashed vertical lines denote the time points where the algorithms detect delay occurrences Red squares denote the durationwhere the occurrences of shockwaves are observed

with both abrupt and gradual speed changes as shown inFigure 7(a)

Once the occurrences of shockwaves are detected theproposed algorithm proceeds to estimate delay as describedin Section 44 Using the spatial extent estimation describedin Section 44 out of four sensors three sensors are usuallyfound to detect speed drops associated with shockwaves at atime Therefore the most downstream sensor and the mostupstream sensor are selected as the start and end sensorsrespectively and 119897119899119875 is found to be 048 km

Table 1 shows the values of V119899119875 V119899119873 120591119899119875 and 120591119899119873 in (1)for each of the four shockwave occurrences For comparisonpurposes delay estimation is also applied for the second andthird occurrences that are detected by benchmark algorithm(see Figure 7) Note that since benchmark algorithm cannotdetect the first and fourth occurrences estimation operationis not activated during these times and normal speed andtravel times are used instead We further note that all fouroccurrences and their associated travel times and delays canbe calculated consecutively online by the proposed algorithm

As shown in Table 1 the four occurrences of shockwavesdetected in Figure 7 are associated with travel times (120591119899119875 in(1)) of 5682 s 5554 s 6098 s and 5379 s respectively Theproposed algorithm is able to detect all four occurrences andthese estimated travel times would all be included in the delaycalculation On the other hand if the benchmark algorithmwas used instead only the second and third occurrences andtheir associated travel times would be reported resulting inless accurate delay calculationWe note that even though thedifference between the delays obtained from the proposedalgorithm and benchmark algorithm in this example maybe small significant improvement can be seen when higherdegree of disruptions are assessed

The results shown in Table 1 have many applicationsin advanced traveler information systems and advancedtraffic management systems A direct application would beto estimate travel time of the freeway segment and report itto motorists The ability to detect occurrences of shockwavescan also be used in other applications such as calculatingshockwave propagation time [25] It can also be used to warn

motorists to adjust intervehicle spacing to avoid rear-endcollisions

62 Assessment Using Real-World Data Collected in ThailandThis section assesses the capability of the proposed algorithmand the benchmark algorithm [17] to detect the sixteenrecorded traffic anomalies in the real-world data set describedin Section 51 As inputs both algorithms use speed flowand occupancy collected by roadside sensorsThe assessmentfollows a 119896-fold cross-validation technique The real-worlddata set consists of 16 anomaly cases 1198631 1198632 11986316 soevaluation is performed 16 times At an iteration 119894 119894 =1 2 16 the anomaly case119863119894 is reserved as a test set whilethe other anomaly cases are collectively used to calibratethe proposed algorithm and benchmark algorithm That isin the first iteration 1198632 11986316 collectively serve as thetraining set which is then tested with1198631Then in the seconditeration 1198631 1198633 11986316 are used for training while 1198632 isused for testing and so on

Table 2 shows that the proposed algorithm achieveshigher detection rate and lower false alarm rate than thebenchmark algorithm Particularly compared to benchmarkalgorithm the proposed algorithm improves the detectionrate and false positive rate by approximately 10 and 7respectively Figures 8 and 9 show examples of anomaliesdetected by the proposed algorithm and found to causetraffic delay Based on the records these anomalies areassociated with disabled vehicles parking over the shoulderarea

It is found that for the benchmark algorithm miss-detection usually occurs under low traffic flow This algo-rithm always needs data from at least two sensors and themagnitude of traffic flow measured by the roadside sensorsis not large enough to trigger the benchmark algorithmWhile medium to heavy traffic flow triggers the benchmarkalgorithm it also increases false alarms This algorithm isdesigned to directly assess the difference between upstreamand downstream sensors of the roadside sensors whichsubsequently increases its sensitivity and triggers many falsealarms

10 Journal of Advanced Transportation

Table 1 Estimation results onNGSIMdata using the proposed algorithm and benchmark algorithm 119897119899119875 =048 kmNote that if the shockwaveis not detected the estimated travel time 120591119899119875 would be equal to normal travel time 120591119899119873

1st 2nd 3rd 4thProposed algorithmShockwaves detected Yes Yes Yes YesV119899119875 (see (1)) 3041 kmh 3112 kmh 2834 kmh 3213 kmh120591119899119875 (see (1)) 5682 s 5554 s 6098 s 5379 s119889119899 = 120591119899119875 minus 120591119899119873 1841 s 1893 s 2399 s 541 sBenchmark algorithm [17]Shockwaves detected No Yes Yes NoV119899119875 (see (1)) 4499 kmh 3112 kmh 2834 kmh 3572 kmh120591119899119875 (see (1)) 3841 s 5554 s 6098 s 4838 s119889119899 = 120591119899119875 minus 120591119899119873 000 s 1893 s 2399 s 000 sIf all 4 shockwaves are not detectedV119899119873 (see (1)) 4499 kmh 4721 kmh 4671 kmh 3572 kmh120591119899119873 (see (1)) 3841 s 3660 s 3700 s 4838 s

Table 2 Results on anomaly detection using data set collected inThailand

Performanceparameters

Proposedalgorithm

Benchmarkalgorithm [17]

DR () 94 83FPR () 5 12MTTD (seconds) 340 140

Figure 8 An anomaly detected at the 4th camera sensor

Unlike the benchmark algorithm the proposed algorithmuses DTW to assess traffic data from each sensor individuallyThis enables the proposed algorithm to recognize patternsassociated with anomalies more effectively including underlow traffic flow The use of DTW also reduces a numberof false alarms as it ensures that unrecognized patternswill not trigger detection at individual sensors In respectto mean time to detection (MTTD) it takes longer forthe proposed algorithm to detect anomalies compared tobenchmark algorithm as it needs to wait till enough temporalsamples of traffic data have been collected in each sensorbefore the detection operation can be triggered Howeverthe mean time to detection is still less than 6 minutes which

Figure 9 Another anomaly detected at the 4th camera sensor

should give enough time for traffic operators to response [8]To reduce mean time to detection of the proposed algorithmshorter time samples of traffic data can be used ReducingMTTD while maintaining acceptable levels of DR and FPRis worth further investigation

To deploy the proposed algorithm computation timeof the algorithm could be an added delay to the MTTDEven though the computation time was not systematicallymeasured for the performance evaluation the computationtime was in order of seconds For the computerrsquos specificationthat we use which was a personal laptop with 25 GHz CPUand 4GB of RAM the computation time was approximately[1 3) seconds As the computing time (a few seconds) wasmuch smaller than the mean time to detection (up to 6minutes) it is reasonable to consider the computing timenegligible

This section further discusses the results when the pro-posed algorithm was applied to estimate 60 delay casesdescribed in Section 51 using different input time windowsizes (119871) and prediction horizons (119875) For practical purposesit is important to select suitable input time window sizeand prediction horizon to enable the proposed method to

Journal of Advanced Transportation 11

L = 3 L = 10L = 5

P = 5

P = 10

P = 15

0

005

01

015

02

025

03

035

04M

APE

Figure 10Mean absolute percentage error (MAPE) fromestimating60 delay cases using different input time window sizes 119871 (minutes)and prediction horizons 119875 (minutes)

perform effectively Even though the 15-minute interval isoften recommended [9] to avoid strong fluctuation usinglower input timewindow sizes can providemore fine-grainedinformation for practitioners [25] Therefore in this paperwe use 119871 = 3 5 and 10 minutes for input time windowsize For prediction horizons we use 119875 = 5 10 and 15minutes as this paper is focused on short-term estimation(less than 15 minutes) Also changes in traffic condition onthis experimental site are normally found to occur in less than15 minutes

The proposed algorithm is assessed on its capability toestimate delay using MAPE Figure 10 shows MAPE fromestimating 60 delay cases in the real-world data set withdifferent input timewindow sizes (119871) and prediction horizons(119875) These results show that the estimation part of theproposed algorithm can achieve MAPE lower than 01 forthis set of real-world data This accuracy is achieved becausethe estimation part consists of both spatial and temporalestimations which enable the proposed algorithm to firstestimate the area disrupted by anomalies and then estimatethe delay itself Furthermore it can be seen that increasinginput time window size (119871) reduces MAPE as using moreinput samples enhances estimation accuracy For the impactof prediction horizons (119875) MAPE tends to increase withlarger prediction horizons as it is more difficult to estimatedelays farther ahead into the future However as shownin Figure 10 when 119875 is increased MAPE in fact increasesat a much smaller rate when larger values of 119871 are usedTherefore it is recommended for practitioners that if theobjective is to estimate delays far ahead into the future (usinglarge prediction horizons) larger input time window size isneeded

Based on the findings in this section it is importantthat practitioners select a suitable input time window sizeaccording to their practical objectives If the objective is toprovide prompt delay information small input time windowsize should be used On the other hand if the objective is toprovide accurately estimated delays for traffic managementpurposes larger input time window sizes should be chosenFurthermore based on the proposed algorithm a morecomplex and flexible system can also be further developedFor example a stepwise approach can be implemented on topof the proposed algorithm where a small input time windowsize is used as a trigger to promptly discover the occurrenceof traffic delay Then in the next step the input time windowsize can be increased to collect more traffic data to achievemore accurate detection and estimation of traffic delay

For anomalies with large scale effect beyond the 11 kmexperimental site we anticipate that the proposed algo-rithm should be able to detect those anomalies if theyoccur downstream and cause changes in traffic patterns topropagate upstream into the 11 km segment To accuratelyestimate delay basis patterns associated with such large scaleanomalies need to be incorporated to distinguish themThenan extrapolation technique needs to be incorporated to enablethe proposed algorithm to estimate delay beyond the 11 kmexpressway segment where local traffic information cannotbe directly observed [25] This is worth further investigation

7 Conclusions and Future Work

This paper proposes an algorithm to detect anomalies andestimate delay using traffic data from roadside sensors Theproposed algorithm uses traffic data from individual sensorsto detect anomalies and then uses data from consecutivedownstream and upstream sensors together to spatially andtemporally estimate delay Performance evaluations usingreal-world data show that the proposed algorithm achieveshigher detection rate and lower false positive rate when usedto detect anomalies compared to an existing benchmarkalgorithmAn investigationwithNGSIMdata also shows thatthe proposed algorithm is also able to detect the presence ofshockwaves and subsequently estimate delaymore efficientlythan benchmark algorithm Further results show that theestimation error increases at a much smaller rate whensuitable input time window size and prediction horizon areselected

There are interesting aspects that are worth furtherinvestigation One interesting aspect is to reduce the pro-cessing time while maintaining low detection and estima-tion errors A possible approach is to introduce a featureextraction method that can extract hidden patterns whensmaller input time window sizes are used [8] Alternativeapproaches include studying a set of hidden representativetraffic flow features in advance by pretraining a deep-learningmodel [36] and populating traffic features into a geneticalgorithm learning method to reduce dimensionality [37]Another interesting aspect is to improve inference capabilityby incorporating a spatial inferencemethod to estimate trafficdata itself at the locations where they cannot be directlymeasured [25] Also it would be practically interesting to

12 Journal of Advanced Transportation

assess the proposed algorithm with different road types andgeometries

In addition to improvements of the algorithm an interest-ing aspect to investigate is the applications of the algorithmThe solution provided in this paper is designed to be astreaming algorithm which can be run online thereforeit can be naturally extended to use in various real-timeapplications For example the detected anomaly and esti-mated additional travel time would be useful for both trafficoperators and travelers if the information is delivered ina timely and convenient manner Active communicationbetween roadside sensors and vehicle-to-infrastructure (V2I)could be used to promptly transmit the information from oursolution to nearby roadside message boards to display theinformation to the drivers on the roads or vehicles with V2Icommunication to adapt routing and navigation

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this article

Acknowledgments

The authors would like to thankDr Supakorn Siddhichai andhis IMG team at NECTEC for their assistance in providingreal-world data for the analysis

References

[1] N Parrado and Y Donoso ldquoCongestion based mechanism forroute discovery in a V2I-V2V system applying smart devicesand IoTrdquo Sensors vol 15 no 4 pp 7768ndash7806 2015

[2] K Nellore and G P Hancke ldquoA survey on urban trafficmanagement system using wireless sensor networksrdquo Sensorsvol 16 no 2 article 157 2016

[3] E I Vlahogianni M G Karlaftis and J C Golias ldquoShort-term traffic forecasting where we are and where wersquore goingrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 3ndash19 2014

[4] R Li and G Rose ldquoIncorporating uncertainty into short-term travel time predictionsrdquo Transportation Research Part CEmerging Technologies vol 19 no 6 pp 1006ndash1018 2011

[5] Y Chung ldquoAmethodological approach for estimating temporaland spatial extent of delays caused by freeway accidentsrdquo IEEETransactions on Intelligent Transportation Systems vol 13 no 3pp 1454ndash1461 2012

[6] J Weng and Q Meng ldquoEstimating capacity and traffic delayin work zones an overviewrdquo Transportation ResearchmdashPart CEmerging Technologies vol 35 pp 34ndash45 2013

[7] X Zhang E Onieva A Perallos E Osaba and V C SLee ldquoHierarchical fuzzy rule-based system optimized withgenetic algorithms for short term traffic congestion predictionrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 127ndash142 2014

[8] S Thajchayapong E S Garcia-Trevino and J A Barria ldquoDis-tributed classification of traffic anomalies using microscopictraffic variablesrdquo IEEETransactions on Intelligent TransportationSystems vol 14 no 1 pp 448ndash458 2013

[9] M Lippi M Bertini and P Frasconi ldquoShort-term traffic flowforecasting an experimental comparison of time-series anal-ysis and supervised learningrdquo IEEE Transactions on IntelligentTransportation Systems vol 14 no 2 pp 871ndash882 2013

[10] F Zheng andH Van Zuylen ldquoUrban link travel time estimationbased on sparse probe vehicle datardquo Transportation ResearchmdashPart C Emerging Technologies vol 31 pp 145ndash157 2013

[11] F Dion and H Rakha ldquoEstimating dynamic roadway traveltimes using automatic vehicle identification data for low sam-pling ratesrdquoTransportation Research Part BMethodological vol40 no 9 pp 745ndash766 2006

[12] D Paterson and G Rose ldquoDynamic travel time estima-tion on instrumented freewaysrdquo in Proceedings of the 6thWorld Congress on Intelligent Transportation Systems TorontoCanada 1999

[13] J Kwon B Coifman and P Bickel ldquoDay-to-day travel-timetrends and travel-time prediction from loop-detector datardquoTransportation Research Record no 1717 pp 120ndash129 2000

[14] J Rice and E Van Zwet ldquoA simple and effective method forpredicting travel times on freewaysrdquo IEEE Transactions onIntelligent Transportation Systems vol 5 no 3 pp 200ndash2072004

[15] C D R Lindveld RThijs P H L Bovy andN J van der ZijppldquoEvaluation of online travel time estimators and predictorsrdquoTransportation Research Record vol 1719 pp 45ndash53 2000

[16] H Yang S Chen YWang BWu andZWang ldquoComparison ofdelay estimation models for signalised intersections using fieldobservations in Shanghairdquo IET Intelligent Transport Systems vol10 no 3 pp 165ndash174 2016

[17] C L Mak and H S L Fan ldquoDevelopment of dual-stationautomated expressway incident detection algorithmsrdquo IEEETransactions on Intelligent Transportation Systems vol 8 no 3pp 480ndash490 2007

[18] K Hi-Ri-O-Tappa C Likitkhajorn A Poolsawat and S Tha-jchayapong ldquoTraffic incident detection system using series ofpoint detectorsrdquo in Proceedings of the 15th International IEEEConference on Intelligent Transportation Systems (ITSC rsquo12) pp182ndash187 Anchorage Alaska USA September 2012

[19] V Alexiadis J Colyar J Halkias R Hranac and G McHaleldquoThe next generation simulation programrdquo ITE Journal vol 74no 8 pp 22ndash26 2004

[20] F G Habtemichael M Cetin and K A Anuar ldquoIncident-induced delays on freeways quantificationmethod by groupingsimilar traffic patternsrdquo Transportation Research Record vol2484 pp 60ndash69 2015

[21] J Li C-J Lan and X Gu ldquoEstimation of incident delay andits uncertainty on freeway networksrdquo Transportation ResearchRecord vol 1959 pp 37ndash45 2006

[22] Y Chung ldquoAssessment of non-recurrent congestion causedby precipitation using archived weather and traffic flow datardquoTransport Policy vol 19 no 1 pp 167ndash173 2012

[23] J A Barria and SThajchayapong ldquoDetection and classificationof traffic anomalies using microscopic traffic variablesrdquo IEEETransactions on Intelligent Transportation Systems vol 12 no3 pp 695ndash704 2011

[24] K HiriOtappa and S Thajchayapong ldquoGeneralizability andtransferability of incident detection algorithm using dynamictime warpingrdquo in Proceedings of the in 19th World Congress onIntelligent Transportation Systems Vienna Austria 2012

Journal of Advanced Transportation 13

[25] S Thajchayapong and J A Barria ldquoSpatial inference of traffictransition using microndashmacro traffic variablesrdquo IEEE Transac-tions on Intelligent Transportation Systems vol 16 no 2 pp854ndash864 2015

[26] C A Ratanamahatana and E Keogh ldquoEverything you knowabout dynamic time warping is wrongrdquo in Proceedings of the3rd Workshop on Mining Temporal and Sequential Data in Con-junction with the 10th ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining (KDD rsquo04) CiteseerSeattle Wash USA August 2004

[27] Transportation Research Thesauras Transportation ResearchBoard National Academy httptrttrborgtrtaspNN=Bthd

[28] N Chiabaut C Buisson and L Leclercq ldquoFundamental dia-gram estimation through passing rate measurements in conges-tionrdquo IEEE Transactions on Intelligent Transportation Systemsvol 10 no 2 pp 355ndash359 2009

[29] J J Fernandez-Lozano M Martın-Guzman J Martın-Avilaand A Garcıa-Cerezo ldquoA wireless sensor network for urbantraffic characterization and trend monitoringrdquo Sensors vol 15no 10 pp 26143ndash26169 2015

[30] J-S Oh C Oh S G Ritchie and M Chang ldquoReal-timeestimation of accident likelihood for safety enhancementrdquoJournal of Transportation Engineering vol 131 no 5 pp 358ndash363 2005

[31] A P Chassiakos and Y J Stephanedes ldquoSmoothing algorithmsfor incident detectionrdquo Transportation Research Record vol1394 pp 8ndash16 1993

[32] V Alarcon-Aquino and J A Barria ldquoAnomaly detection incommunication networks using waveletsrdquo IEE ProceedingsmdashCommunications vol 148 no 6 pp 355ndash362 2001

[33] D Berndt and J Clifford ldquoUsing dynamic time warping tofind patterns in time seriesrdquo in Proceedings of the Workshop onKnowledge Discovery in Databases (KDD rsquo94) vol 10 pp 359ndash370 Seattle Wash USA August 1994

[34] K Kiratiratanapruk and S Siddhichai ldquoVehicle detection andtracking for traffic monitoring systemrdquo in Proceedings of theIEEE Region 10 Conference (TENCON rsquo06) pp 1ndash4 IEEE HongKong November 2006

[35] A Hegyi B De Schutter and J Hellendoorn ldquoOptimal coordi-nation of variable speed limits to suppress shock wavesrdquo IEEETransactions on Intelligent Transportation Systems vol 6 no 1pp 102ndash112 2005

[36] Y Lv Y Duan W Kang Z Li and F-Y Wang ldquoTraffic flowprediction with big data a deep learning approachrdquo IEEETransactions on Intelligent Transportation Systems vol 16 no2 pp 865ndash873 2015

[37] P Lopez-Garcia E Onieva E Osaba A D Masegosa and APerallos ldquoA hybrid method for short-term traffic congestionforecasting using genetic algorithms and cross entropyrdquo IEEETransactions on Intelligent Transportation Systems vol 17 no 2pp 557ndash569 2016

6 Journal of Advanced Transportation

Traffi

c flow

Start sensor (1st sensor where anomaly is detected)

Intermediate sensor (anomaly detected)

Intermediate sensor (anomaly detected)

End sensor (1st sensor where anomaly is not detected)

Location where a lane is blocked (eg by an accident or a disabled vehicle)

Figure 4 An illustration of how the proposed algorithm finds start intermediate and end sensors from speed patterns measured by roadsidesensors Start sensor is the first downstream sensor that detects an anomaly End sensor is the first upstream sensor that an anomaly is notdetected Intermediate sensors are used to confirm that the congested area has extended upstream from start sensor to end sensor

is detected and proceeds to delay estimation We note thatanomalies can also be detected bymultiple sensors which canbe used in the delay estimation operation

44 Delay Estimation Delay estimation operation consistsof two consecutive steps (1) estimation of spatial extentof anomalies 119897119899119875 and (2) estimation of temporal extent 119889119899as shown in blocks III and IV in Figure 2 respectivelyThe spatial extent of delay 119897119899119875 is the length of the roadsegment that experiences anomalies which can be estimatedby finding a set of consecutive sensors where anomalies aredetected in Section 43 As shown in Figure 4 the roadsidesensors considered to be in the disruption area consist ofthe start sensor the intermediate sensors and the end sensorThe start sensor is the first downstream sensor where ananomaly is detected that is it is the location where delay isoriginated The intermediate sensors are the sensors locatedconsecutively upstream from the start sensor that also detectsanomalies that is they signify the continuous upstreamextension of anomalies The end sensor is the first upstreamsensor where an anomaly is not detected that is this isthe location where the disruption ceases to exist and trafficupstream from this location is still under normal conditionThen 119897119899119875 is estimated as the distance between the start andthe end sensors

The second step temporal estimation of delay in blockIV in Figure 2 uses 119897119899119875 as input to estimate delay vehiclesare expected to experience on the road segment Usingthe definition in [27] 119889119899 the estimated delay at time 119899 iscalculated as the estimated additional travel time that exceedsthe normal travel time as shown in (1) where 120591119899119875 is theestimated travel time and 120591119899119873 is the normal travel timeon the same road segment The estimated travel time 120591119899119875is calculated by dividing 119897119899119875 with the estimated mean ofspeeds measured at the start intermediate and end sensorsV119899119875

The normal travel time 120591119899119873 is calculated by dividing 119897119899119875with the normal speed V119899119873 We note that V119899119873 is calculatedas the mean of measured speeds in the previous timewindow (size 119871) just before an anomaly is detected in thecurrent time window (also size 119871)Therefore V119899119873 reflects theexpected speed drivers on the freeway segment would haveexperienced if an anomaly had not occurred

119889119899 = 120591119899119875 minus 120591119899119873 120591119899119875 =119897119899119875V119899119875 120591119899119873 =

119897119899119875V119899119873 (1)

The estimated mean of speeds V119899119875 can be calculated asfollows Based on our previous study in [24] and furtheranalysis of real-world data over a 6-month period it isobserved that the occurrences of delay are associatedwith fiveanomalous speed drop patterns namely linear exponentiallogarithmic sigmoid and unclassified Furthermore it ispossible to find empirical equations that best fit each of thesefive speed drop patterns (eg an exponential equation thatbest fits the exponential speed drop patterns) [24] A kernelfunction is used for the unclassified speed drop patterns

For a given prediction horizon 119875 and a total of119872 startintermediate and end sensors these empirical equations canbe used to temporally extrapolate speeds well into the futurefor each roadside sensor119898 that is [V119898119899+1 V119898119899+2 V119898119899+119875]Then V119899119875 can be obtained by first averaging temporallyextrapolated speed samples V119898119899119875 and then averaging spa-tially across all119872 sensors as follows

V119899119875 =sum119872119898=1 V119898119899119875119872 V119898119899119875 =

sum119875119894=1 V119898119899+119894119875 (2)

In practice delay estimation can be operated as followsOnce delay has been detected speed patterns measured byeach roadside sensor are measured and transmitted to thealgorithmmodule in Figure 1 UsingDynamicTimeWarping

Journal of Advanced Transportation 7

Daokanong

Suksawat Sathupradit Sathupradit River side

Port

1 2 3 4 5 6 7 8 9 10 11

644m1977m1078m393m390m786m1337m671m1505m1152m

2nd stage2nd stage

Figure 5 A sketch of expressway segment where real-world data was obtained

these speed patterns are compared with the basis speedpatterns associated with anomalies 119880119860119898 which have beencollected and stored a priori to find similarity Once themost similar basis speed drop pattern has been found itscorresponding empirical equation is used to calculate theestimated mean speed V119899119875 as shown in (2) which is thenused to calculate 120591119899119875 and eventually to estimate delay 119889119899 in(1) Note that the calculation in (1) is for short-term delay[3] Extending (1) to estimate vehiclersquos hour of delay or totaldelay experienced by all delayed vehicles on the road segmentcan be performed by adding traffic volume as amultiplicationfactor to (1)

5 Performance Evaluations

In this section performance evaluations are conducted usingreal-world data First Section 51 presents two real-worlddata sets used to assess the proposed algorithm ThenSection 52 describes an existing algorithm [17] which isused as benchmark In Section 53 performance evaluationparameters are presented

51 Descriptions of Real-World Data Sets

511 Real-World Data from Roadside Camera Sensors Theproposed algorithm is assessed using real-world data col-lected from a series of camera sensors installed on heavilyused inbound section ofChalerm-Maha-NakhonExpresswayin Bangkok Thailand These sensors were part of a projectjointly initiated by the Expressway Authority of Thailand(EXAT) and National Electronic and Computer TechnologyCenter (NECTEC) aiming at real-time traffic monitoringIt consists of a series of eleven camera sensors installedconsecutively along the 11 km expressway Figure 5 shows thetopology of the expressway and the locations of the camerastations where the distance between two consecutive camerasensors ranges from 644 to 1977 meters Each camera sensoris capable of detecting individual vehicles that passed theexpressway segment and produced time series of traffic dataincluding speed occupancy and flow in 1-minute intervals[34] Each camera sensor is connected to a local EXAT oper-ation office via wireless cellular network technologies (egGPRS and 3G) where operators can view real-time trafficinformation and make decisions on traffic management

A road segment within the full of view of a video camerais used as a designated area to calculate traffic data inputSpeed is measured as a distance in a vehiclersquos traverse in the

Figure 6 An image of a traffic delay occurrence captured at the 4thcamera sensor

designated area divided by the time it takes Flow ismeasuredas the rate of vehicles passing a designated area at a given timeinterval Occupancy is measured as a fraction of time that adesignated area is occupied by vehicles

The data set used in this paper was collected from thissensor network over a one year period from January 1 toDecember 31 2010 In this paper we chose the 3rd 4th 5thand 6th sensors as the experimental sites An example of traf-fic congestion viewed by a camera sensor studied in this paperis shown in Figure 6 For evaluation purposes anomaliesand their associated delay that took place were independentlyand manually logged by a team of transportation expertsA total of 16 nonrecurring congestion cases were recordedwhich are used to test the anomaly detection capability ofthe proposed algorithm Each record was associated with thetime instances prior to the disruption occurrence and after ithad ended Besides these 16 cases further investigation withanother team of experts also revealed additional 44 delaycases Therefore a total of 60 delay cases are used to test theestimation operation of the proposed algorithm

We note that the traffic data input of the proposedalgorithm is typical data that most roadside sensors forexample loop detectors already produce Therefore weanticipate that if the sensors are to be replaced with othertypes of sensors the performance of the system should notchange significantly This is worth further investigation forlarge scale deployment

512 NGSIM Data To show the potential of being deployedat a different location the proposed algorithm is also assessedusingNGSIM data [19]The selected data set was collected onthe southbound US Highway 101 in Los Angeles California

8 Journal of Advanced Transportation

USA at 0750 amndash0835 am This data set is selected becauseit was collected during the buildup of congestion and consistsof shockwaves which can potentially induce additional delayIt should be noted that the anomaly cases and the amountof delay caused by them were not labeled in this data setso the full algorithm could not be deployed for this dataset However the potential of the algorithm could be shownby shockwave detection To assess the proposed algorithmfour virtual sensors are placed approximately 01 km apart toobtain traffic data to capture the occurrence of shockwavesand their spatial extents Similar to conventional fixed trafficsensors (eg camera sensors and loop detectors) thesevirtual sensors calculate traffic data based on traffic flow andspeed that pass the sensorsrsquo locations An input time windowsize 119871 of 5 minutes is used where time series of 30-secondaverage speeds obtained from upstream and downstreamvirtual sensors are used as inputs The occurrences of shock-waves are observed and recorded by a team of transportationresearchers [25]

52 Benchmark Algorithm To assess the capability of theproposed algorithm in respect to existing approaches abenchmark algorithm is used [17]This benchmark algorithmis selected because it has a similar objective to the proposedalgorithm in detecting anomalies that cause delay Also thisalgorithm has already been implemented in a real-worldanomaly detection system [17] Furthermore composed ofa forecasting component and a detection logic componentit is designed to assess the difference between traffic dataobserved at consecutive roadside sensors

However the main difference between the benchmarkalgorithm and the proposed algorithm lies in the detectionlogic The benchmark algorithm is designed to detect sud-den significant changes of traffic inputs from the forecastcomponent values by establishing threshold values (speedand occupancy) to filter out gradual changes in traffic inputscaused by recurring congestion and an anomalous signalwill only be raised when the specified threshold values havebeen exceeded On the other hand the proposed algorithm isdesigned to detect both sudden and gradual changes whichenables it to detect different types of anomalies with fewerfalse alarms This difference will become more apparentwhen both algorithms are assessed with real-world data inSection 6

53 Performance Evaluation Parameters To numericallyevaluate the proposed algorithm we use detection rate (DR)false positive rate (FPR) mean time to detection (MTTD)andmean absolute percentage error (MAPE) as shown in (3)(4) (5) and (6) respectively DR FPR and MTTD are usedto assess the anomaly detection operation of the proposedalgorithm MAPE is used to assess the accuracy of delayestimation of a total 119863 recorded delay cases where the errorof the estimation 119889119899119895 of the 119895th recorded delay at time 119899 isnormalized by the actual delay119889119886119899119895 tominimize the possibilityof some estimated delay samples from dominating the others

DR =Number of delays detectedTotal number of delays

(3)

FPR

=Number of detected delays that are false positives

Total number of detected delays

(4)

MTTD

=Sum of time taken to detect individual delays

Total number of detected delays

(5)

MAPE = 1119863

119863

sum119895=1

10038161003816100381610038161003816119889119886119899119895 minus 119889119899119895

10038161003816100381610038161003816119889119886119899119895 (6)

The performance parameters are applicable for data setsthat contain labels for traffic anomalies and the amount ofdelay caused by the anomalies Therefore these parameterswill be evaluated for the real-world data collected inThailandbut due to the lack of anomaly labels could not be applied forthe NGSIM data set

6 Results and Discussions

In this section the proposed algorithm is assessed usingreal-world data described in Section 51 and compared tothe benchmark algorithm described in Section 52 Bothalgorithms are tested as if they operate online where eachdetection operation is performed per input time window 119871

61 Assessment Using NGSIM Data The proposed algorithmis assessed on its capability to detect the the presence ofshockwaves which refers to a phenomenon where vehicleshave to temporarily slow down below normal speed andsubsequently force the vehicles behind to slow down furtherThis phenomenon is known to create shockwave fronts ormoving speed drop areas that move backward disruptingflow and annoying drivers behind [35] The occurrence ofshockwaves during congestion not only causes additionaldelay but also increases chances of rear-end collisionsIdentifying shockwaves during congestion also presents achallenge as it occurs temporarily and requires an algorithmthat is fast and sensitive enough to detect them [25] Theoccurrence of shockwaves represents unexpected events sothey are considered as anomalies in this paper

Figure 7 shows that the proposed algorithm can detectall four occurrences (Figure 7(a)) of shockwaves whilebenchmark algorithm can detect only two (Figure 7(b))As shown in Figure 7(b) only abrupt changes in speeddifference (second and third occurrences) can be detectedby the benchmark algorithm This algorithm is designedto detect consecutive differences between upstream anddownstream traffic measurements so it operates well withconsecutive changes in speed difference that are significantenough On the other hand benchmark algorithmmisses thefirst and fourth occurrences of shockwaves in Figure 7(b) asthey are associated with relatively more gradual changes ofspeed difference Unlike benchmark algorithm the proposedalgorithm uses DTW to assess input patterns to detect delayoccurrences so it can detect all four occurrences associated

Journal of Advanced Transportation 9

Upstream sensorDownstream sensor

805 820 835750Time (hoursminutes)

01020304050607080

Spee

d (k

mh

)

(a)

One time step aheadTwo time steps ahead

805 820 835750Time (hoursminutes)

minus3minus2minus1

0123

Spee

d di

ffere

nces

(km

h)

(b)

Figure 7 Detecting delay occurrences associated with shockwaves in NGSIM data using (a) the proposed algorithm and (b) benchmarkalgorithm [17] Dashed vertical lines denote the time points where the algorithms detect delay occurrences Red squares denote the durationwhere the occurrences of shockwaves are observed

with both abrupt and gradual speed changes as shown inFigure 7(a)

Once the occurrences of shockwaves are detected theproposed algorithm proceeds to estimate delay as describedin Section 44 Using the spatial extent estimation describedin Section 44 out of four sensors three sensors are usuallyfound to detect speed drops associated with shockwaves at atime Therefore the most downstream sensor and the mostupstream sensor are selected as the start and end sensorsrespectively and 119897119899119875 is found to be 048 km

Table 1 shows the values of V119899119875 V119899119873 120591119899119875 and 120591119899119873 in (1)for each of the four shockwave occurrences For comparisonpurposes delay estimation is also applied for the second andthird occurrences that are detected by benchmark algorithm(see Figure 7) Note that since benchmark algorithm cannotdetect the first and fourth occurrences estimation operationis not activated during these times and normal speed andtravel times are used instead We further note that all fouroccurrences and their associated travel times and delays canbe calculated consecutively online by the proposed algorithm

As shown in Table 1 the four occurrences of shockwavesdetected in Figure 7 are associated with travel times (120591119899119875 in(1)) of 5682 s 5554 s 6098 s and 5379 s respectively Theproposed algorithm is able to detect all four occurrences andthese estimated travel times would all be included in the delaycalculation On the other hand if the benchmark algorithmwas used instead only the second and third occurrences andtheir associated travel times would be reported resulting inless accurate delay calculationWe note that even though thedifference between the delays obtained from the proposedalgorithm and benchmark algorithm in this example maybe small significant improvement can be seen when higherdegree of disruptions are assessed

The results shown in Table 1 have many applicationsin advanced traveler information systems and advancedtraffic management systems A direct application would beto estimate travel time of the freeway segment and report itto motorists The ability to detect occurrences of shockwavescan also be used in other applications such as calculatingshockwave propagation time [25] It can also be used to warn

motorists to adjust intervehicle spacing to avoid rear-endcollisions

62 Assessment Using Real-World Data Collected in ThailandThis section assesses the capability of the proposed algorithmand the benchmark algorithm [17] to detect the sixteenrecorded traffic anomalies in the real-world data set describedin Section 51 As inputs both algorithms use speed flowand occupancy collected by roadside sensorsThe assessmentfollows a 119896-fold cross-validation technique The real-worlddata set consists of 16 anomaly cases 1198631 1198632 11986316 soevaluation is performed 16 times At an iteration 119894 119894 =1 2 16 the anomaly case119863119894 is reserved as a test set whilethe other anomaly cases are collectively used to calibratethe proposed algorithm and benchmark algorithm That isin the first iteration 1198632 11986316 collectively serve as thetraining set which is then tested with1198631Then in the seconditeration 1198631 1198633 11986316 are used for training while 1198632 isused for testing and so on

Table 2 shows that the proposed algorithm achieveshigher detection rate and lower false alarm rate than thebenchmark algorithm Particularly compared to benchmarkalgorithm the proposed algorithm improves the detectionrate and false positive rate by approximately 10 and 7respectively Figures 8 and 9 show examples of anomaliesdetected by the proposed algorithm and found to causetraffic delay Based on the records these anomalies areassociated with disabled vehicles parking over the shoulderarea

It is found that for the benchmark algorithm miss-detection usually occurs under low traffic flow This algo-rithm always needs data from at least two sensors and themagnitude of traffic flow measured by the roadside sensorsis not large enough to trigger the benchmark algorithmWhile medium to heavy traffic flow triggers the benchmarkalgorithm it also increases false alarms This algorithm isdesigned to directly assess the difference between upstreamand downstream sensors of the roadside sensors whichsubsequently increases its sensitivity and triggers many falsealarms

10 Journal of Advanced Transportation

Table 1 Estimation results onNGSIMdata using the proposed algorithm and benchmark algorithm 119897119899119875 =048 kmNote that if the shockwaveis not detected the estimated travel time 120591119899119875 would be equal to normal travel time 120591119899119873

1st 2nd 3rd 4thProposed algorithmShockwaves detected Yes Yes Yes YesV119899119875 (see (1)) 3041 kmh 3112 kmh 2834 kmh 3213 kmh120591119899119875 (see (1)) 5682 s 5554 s 6098 s 5379 s119889119899 = 120591119899119875 minus 120591119899119873 1841 s 1893 s 2399 s 541 sBenchmark algorithm [17]Shockwaves detected No Yes Yes NoV119899119875 (see (1)) 4499 kmh 3112 kmh 2834 kmh 3572 kmh120591119899119875 (see (1)) 3841 s 5554 s 6098 s 4838 s119889119899 = 120591119899119875 minus 120591119899119873 000 s 1893 s 2399 s 000 sIf all 4 shockwaves are not detectedV119899119873 (see (1)) 4499 kmh 4721 kmh 4671 kmh 3572 kmh120591119899119873 (see (1)) 3841 s 3660 s 3700 s 4838 s

Table 2 Results on anomaly detection using data set collected inThailand

Performanceparameters

Proposedalgorithm

Benchmarkalgorithm [17]

DR () 94 83FPR () 5 12MTTD (seconds) 340 140

Figure 8 An anomaly detected at the 4th camera sensor

Unlike the benchmark algorithm the proposed algorithmuses DTW to assess traffic data from each sensor individuallyThis enables the proposed algorithm to recognize patternsassociated with anomalies more effectively including underlow traffic flow The use of DTW also reduces a numberof false alarms as it ensures that unrecognized patternswill not trigger detection at individual sensors In respectto mean time to detection (MTTD) it takes longer forthe proposed algorithm to detect anomalies compared tobenchmark algorithm as it needs to wait till enough temporalsamples of traffic data have been collected in each sensorbefore the detection operation can be triggered Howeverthe mean time to detection is still less than 6 minutes which

Figure 9 Another anomaly detected at the 4th camera sensor

should give enough time for traffic operators to response [8]To reduce mean time to detection of the proposed algorithmshorter time samples of traffic data can be used ReducingMTTD while maintaining acceptable levels of DR and FPRis worth further investigation

To deploy the proposed algorithm computation timeof the algorithm could be an added delay to the MTTDEven though the computation time was not systematicallymeasured for the performance evaluation the computationtime was in order of seconds For the computerrsquos specificationthat we use which was a personal laptop with 25 GHz CPUand 4GB of RAM the computation time was approximately[1 3) seconds As the computing time (a few seconds) wasmuch smaller than the mean time to detection (up to 6minutes) it is reasonable to consider the computing timenegligible

This section further discusses the results when the pro-posed algorithm was applied to estimate 60 delay casesdescribed in Section 51 using different input time windowsizes (119871) and prediction horizons (119875) For practical purposesit is important to select suitable input time window sizeand prediction horizon to enable the proposed method to

Journal of Advanced Transportation 11

L = 3 L = 10L = 5

P = 5

P = 10

P = 15

0

005

01

015

02

025

03

035

04M

APE

Figure 10Mean absolute percentage error (MAPE) fromestimating60 delay cases using different input time window sizes 119871 (minutes)and prediction horizons 119875 (minutes)

perform effectively Even though the 15-minute interval isoften recommended [9] to avoid strong fluctuation usinglower input timewindow sizes can providemore fine-grainedinformation for practitioners [25] Therefore in this paperwe use 119871 = 3 5 and 10 minutes for input time windowsize For prediction horizons we use 119875 = 5 10 and 15minutes as this paper is focused on short-term estimation(less than 15 minutes) Also changes in traffic condition onthis experimental site are normally found to occur in less than15 minutes

The proposed algorithm is assessed on its capability toestimate delay using MAPE Figure 10 shows MAPE fromestimating 60 delay cases in the real-world data set withdifferent input timewindow sizes (119871) and prediction horizons(119875) These results show that the estimation part of theproposed algorithm can achieve MAPE lower than 01 forthis set of real-world data This accuracy is achieved becausethe estimation part consists of both spatial and temporalestimations which enable the proposed algorithm to firstestimate the area disrupted by anomalies and then estimatethe delay itself Furthermore it can be seen that increasinginput time window size (119871) reduces MAPE as using moreinput samples enhances estimation accuracy For the impactof prediction horizons (119875) MAPE tends to increase withlarger prediction horizons as it is more difficult to estimatedelays farther ahead into the future However as shownin Figure 10 when 119875 is increased MAPE in fact increasesat a much smaller rate when larger values of 119871 are usedTherefore it is recommended for practitioners that if theobjective is to estimate delays far ahead into the future (usinglarge prediction horizons) larger input time window size isneeded

Based on the findings in this section it is importantthat practitioners select a suitable input time window sizeaccording to their practical objectives If the objective is toprovide prompt delay information small input time windowsize should be used On the other hand if the objective is toprovide accurately estimated delays for traffic managementpurposes larger input time window sizes should be chosenFurthermore based on the proposed algorithm a morecomplex and flexible system can also be further developedFor example a stepwise approach can be implemented on topof the proposed algorithm where a small input time windowsize is used as a trigger to promptly discover the occurrenceof traffic delay Then in the next step the input time windowsize can be increased to collect more traffic data to achievemore accurate detection and estimation of traffic delay

For anomalies with large scale effect beyond the 11 kmexperimental site we anticipate that the proposed algo-rithm should be able to detect those anomalies if theyoccur downstream and cause changes in traffic patterns topropagate upstream into the 11 km segment To accuratelyestimate delay basis patterns associated with such large scaleanomalies need to be incorporated to distinguish themThenan extrapolation technique needs to be incorporated to enablethe proposed algorithm to estimate delay beyond the 11 kmexpressway segment where local traffic information cannotbe directly observed [25] This is worth further investigation

7 Conclusions and Future Work

This paper proposes an algorithm to detect anomalies andestimate delay using traffic data from roadside sensors Theproposed algorithm uses traffic data from individual sensorsto detect anomalies and then uses data from consecutivedownstream and upstream sensors together to spatially andtemporally estimate delay Performance evaluations usingreal-world data show that the proposed algorithm achieveshigher detection rate and lower false positive rate when usedto detect anomalies compared to an existing benchmarkalgorithmAn investigationwithNGSIMdata also shows thatthe proposed algorithm is also able to detect the presence ofshockwaves and subsequently estimate delaymore efficientlythan benchmark algorithm Further results show that theestimation error increases at a much smaller rate whensuitable input time window size and prediction horizon areselected

There are interesting aspects that are worth furtherinvestigation One interesting aspect is to reduce the pro-cessing time while maintaining low detection and estima-tion errors A possible approach is to introduce a featureextraction method that can extract hidden patterns whensmaller input time window sizes are used [8] Alternativeapproaches include studying a set of hidden representativetraffic flow features in advance by pretraining a deep-learningmodel [36] and populating traffic features into a geneticalgorithm learning method to reduce dimensionality [37]Another interesting aspect is to improve inference capabilityby incorporating a spatial inferencemethod to estimate trafficdata itself at the locations where they cannot be directlymeasured [25] Also it would be practically interesting to

12 Journal of Advanced Transportation

assess the proposed algorithm with different road types andgeometries

In addition to improvements of the algorithm an interest-ing aspect to investigate is the applications of the algorithmThe solution provided in this paper is designed to be astreaming algorithm which can be run online thereforeit can be naturally extended to use in various real-timeapplications For example the detected anomaly and esti-mated additional travel time would be useful for both trafficoperators and travelers if the information is delivered ina timely and convenient manner Active communicationbetween roadside sensors and vehicle-to-infrastructure (V2I)could be used to promptly transmit the information from oursolution to nearby roadside message boards to display theinformation to the drivers on the roads or vehicles with V2Icommunication to adapt routing and navigation

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this article

Acknowledgments

The authors would like to thankDr Supakorn Siddhichai andhis IMG team at NECTEC for their assistance in providingreal-world data for the analysis

References

[1] N Parrado and Y Donoso ldquoCongestion based mechanism forroute discovery in a V2I-V2V system applying smart devicesand IoTrdquo Sensors vol 15 no 4 pp 7768ndash7806 2015

[2] K Nellore and G P Hancke ldquoA survey on urban trafficmanagement system using wireless sensor networksrdquo Sensorsvol 16 no 2 article 157 2016

[3] E I Vlahogianni M G Karlaftis and J C Golias ldquoShort-term traffic forecasting where we are and where wersquore goingrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 3ndash19 2014

[4] R Li and G Rose ldquoIncorporating uncertainty into short-term travel time predictionsrdquo Transportation Research Part CEmerging Technologies vol 19 no 6 pp 1006ndash1018 2011

[5] Y Chung ldquoAmethodological approach for estimating temporaland spatial extent of delays caused by freeway accidentsrdquo IEEETransactions on Intelligent Transportation Systems vol 13 no 3pp 1454ndash1461 2012

[6] J Weng and Q Meng ldquoEstimating capacity and traffic delayin work zones an overviewrdquo Transportation ResearchmdashPart CEmerging Technologies vol 35 pp 34ndash45 2013

[7] X Zhang E Onieva A Perallos E Osaba and V C SLee ldquoHierarchical fuzzy rule-based system optimized withgenetic algorithms for short term traffic congestion predictionrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 127ndash142 2014

[8] S Thajchayapong E S Garcia-Trevino and J A Barria ldquoDis-tributed classification of traffic anomalies using microscopictraffic variablesrdquo IEEETransactions on Intelligent TransportationSystems vol 14 no 1 pp 448ndash458 2013

[9] M Lippi M Bertini and P Frasconi ldquoShort-term traffic flowforecasting an experimental comparison of time-series anal-ysis and supervised learningrdquo IEEE Transactions on IntelligentTransportation Systems vol 14 no 2 pp 871ndash882 2013

[10] F Zheng andH Van Zuylen ldquoUrban link travel time estimationbased on sparse probe vehicle datardquo Transportation ResearchmdashPart C Emerging Technologies vol 31 pp 145ndash157 2013

[11] F Dion and H Rakha ldquoEstimating dynamic roadway traveltimes using automatic vehicle identification data for low sam-pling ratesrdquoTransportation Research Part BMethodological vol40 no 9 pp 745ndash766 2006

[12] D Paterson and G Rose ldquoDynamic travel time estima-tion on instrumented freewaysrdquo in Proceedings of the 6thWorld Congress on Intelligent Transportation Systems TorontoCanada 1999

[13] J Kwon B Coifman and P Bickel ldquoDay-to-day travel-timetrends and travel-time prediction from loop-detector datardquoTransportation Research Record no 1717 pp 120ndash129 2000

[14] J Rice and E Van Zwet ldquoA simple and effective method forpredicting travel times on freewaysrdquo IEEE Transactions onIntelligent Transportation Systems vol 5 no 3 pp 200ndash2072004

[15] C D R Lindveld RThijs P H L Bovy andN J van der ZijppldquoEvaluation of online travel time estimators and predictorsrdquoTransportation Research Record vol 1719 pp 45ndash53 2000

[16] H Yang S Chen YWang BWu andZWang ldquoComparison ofdelay estimation models for signalised intersections using fieldobservations in Shanghairdquo IET Intelligent Transport Systems vol10 no 3 pp 165ndash174 2016

[17] C L Mak and H S L Fan ldquoDevelopment of dual-stationautomated expressway incident detection algorithmsrdquo IEEETransactions on Intelligent Transportation Systems vol 8 no 3pp 480ndash490 2007

[18] K Hi-Ri-O-Tappa C Likitkhajorn A Poolsawat and S Tha-jchayapong ldquoTraffic incident detection system using series ofpoint detectorsrdquo in Proceedings of the 15th International IEEEConference on Intelligent Transportation Systems (ITSC rsquo12) pp182ndash187 Anchorage Alaska USA September 2012

[19] V Alexiadis J Colyar J Halkias R Hranac and G McHaleldquoThe next generation simulation programrdquo ITE Journal vol 74no 8 pp 22ndash26 2004

[20] F G Habtemichael M Cetin and K A Anuar ldquoIncident-induced delays on freeways quantificationmethod by groupingsimilar traffic patternsrdquo Transportation Research Record vol2484 pp 60ndash69 2015

[21] J Li C-J Lan and X Gu ldquoEstimation of incident delay andits uncertainty on freeway networksrdquo Transportation ResearchRecord vol 1959 pp 37ndash45 2006

[22] Y Chung ldquoAssessment of non-recurrent congestion causedby precipitation using archived weather and traffic flow datardquoTransport Policy vol 19 no 1 pp 167ndash173 2012

[23] J A Barria and SThajchayapong ldquoDetection and classificationof traffic anomalies using microscopic traffic variablesrdquo IEEETransactions on Intelligent Transportation Systems vol 12 no3 pp 695ndash704 2011

[24] K HiriOtappa and S Thajchayapong ldquoGeneralizability andtransferability of incident detection algorithm using dynamictime warpingrdquo in Proceedings of the in 19th World Congress onIntelligent Transportation Systems Vienna Austria 2012

Journal of Advanced Transportation 13

[25] S Thajchayapong and J A Barria ldquoSpatial inference of traffictransition using microndashmacro traffic variablesrdquo IEEE Transac-tions on Intelligent Transportation Systems vol 16 no 2 pp854ndash864 2015

[26] C A Ratanamahatana and E Keogh ldquoEverything you knowabout dynamic time warping is wrongrdquo in Proceedings of the3rd Workshop on Mining Temporal and Sequential Data in Con-junction with the 10th ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining (KDD rsquo04) CiteseerSeattle Wash USA August 2004

[27] Transportation Research Thesauras Transportation ResearchBoard National Academy httptrttrborgtrtaspNN=Bthd

[28] N Chiabaut C Buisson and L Leclercq ldquoFundamental dia-gram estimation through passing rate measurements in conges-tionrdquo IEEE Transactions on Intelligent Transportation Systemsvol 10 no 2 pp 355ndash359 2009

[29] J J Fernandez-Lozano M Martın-Guzman J Martın-Avilaand A Garcıa-Cerezo ldquoA wireless sensor network for urbantraffic characterization and trend monitoringrdquo Sensors vol 15no 10 pp 26143ndash26169 2015

[30] J-S Oh C Oh S G Ritchie and M Chang ldquoReal-timeestimation of accident likelihood for safety enhancementrdquoJournal of Transportation Engineering vol 131 no 5 pp 358ndash363 2005

[31] A P Chassiakos and Y J Stephanedes ldquoSmoothing algorithmsfor incident detectionrdquo Transportation Research Record vol1394 pp 8ndash16 1993

[32] V Alarcon-Aquino and J A Barria ldquoAnomaly detection incommunication networks using waveletsrdquo IEE ProceedingsmdashCommunications vol 148 no 6 pp 355ndash362 2001

[33] D Berndt and J Clifford ldquoUsing dynamic time warping tofind patterns in time seriesrdquo in Proceedings of the Workshop onKnowledge Discovery in Databases (KDD rsquo94) vol 10 pp 359ndash370 Seattle Wash USA August 1994

[34] K Kiratiratanapruk and S Siddhichai ldquoVehicle detection andtracking for traffic monitoring systemrdquo in Proceedings of theIEEE Region 10 Conference (TENCON rsquo06) pp 1ndash4 IEEE HongKong November 2006

[35] A Hegyi B De Schutter and J Hellendoorn ldquoOptimal coordi-nation of variable speed limits to suppress shock wavesrdquo IEEETransactions on Intelligent Transportation Systems vol 6 no 1pp 102ndash112 2005

[36] Y Lv Y Duan W Kang Z Li and F-Y Wang ldquoTraffic flowprediction with big data a deep learning approachrdquo IEEETransactions on Intelligent Transportation Systems vol 16 no2 pp 865ndash873 2015

[37] P Lopez-Garcia E Onieva E Osaba A D Masegosa and APerallos ldquoA hybrid method for short-term traffic congestionforecasting using genetic algorithms and cross entropyrdquo IEEETransactions on Intelligent Transportation Systems vol 17 no 2pp 557ndash569 2016

Journal of Advanced Transportation 7

Daokanong

Suksawat Sathupradit Sathupradit River side

Port

1 2 3 4 5 6 7 8 9 10 11

644m1977m1078m393m390m786m1337m671m1505m1152m

2nd stage2nd stage

Figure 5 A sketch of expressway segment where real-world data was obtained

these speed patterns are compared with the basis speedpatterns associated with anomalies 119880119860119898 which have beencollected and stored a priori to find similarity Once themost similar basis speed drop pattern has been found itscorresponding empirical equation is used to calculate theestimated mean speed V119899119875 as shown in (2) which is thenused to calculate 120591119899119875 and eventually to estimate delay 119889119899 in(1) Note that the calculation in (1) is for short-term delay[3] Extending (1) to estimate vehiclersquos hour of delay or totaldelay experienced by all delayed vehicles on the road segmentcan be performed by adding traffic volume as amultiplicationfactor to (1)

5 Performance Evaluations

In this section performance evaluations are conducted usingreal-world data First Section 51 presents two real-worlddata sets used to assess the proposed algorithm ThenSection 52 describes an existing algorithm [17] which isused as benchmark In Section 53 performance evaluationparameters are presented

51 Descriptions of Real-World Data Sets

511 Real-World Data from Roadside Camera Sensors Theproposed algorithm is assessed using real-world data col-lected from a series of camera sensors installed on heavilyused inbound section ofChalerm-Maha-NakhonExpresswayin Bangkok Thailand These sensors were part of a projectjointly initiated by the Expressway Authority of Thailand(EXAT) and National Electronic and Computer TechnologyCenter (NECTEC) aiming at real-time traffic monitoringIt consists of a series of eleven camera sensors installedconsecutively along the 11 km expressway Figure 5 shows thetopology of the expressway and the locations of the camerastations where the distance between two consecutive camerasensors ranges from 644 to 1977 meters Each camera sensoris capable of detecting individual vehicles that passed theexpressway segment and produced time series of traffic dataincluding speed occupancy and flow in 1-minute intervals[34] Each camera sensor is connected to a local EXAT oper-ation office via wireless cellular network technologies (egGPRS and 3G) where operators can view real-time trafficinformation and make decisions on traffic management

A road segment within the full of view of a video camerais used as a designated area to calculate traffic data inputSpeed is measured as a distance in a vehiclersquos traverse in the

Figure 6 An image of a traffic delay occurrence captured at the 4thcamera sensor

designated area divided by the time it takes Flow ismeasuredas the rate of vehicles passing a designated area at a given timeinterval Occupancy is measured as a fraction of time that adesignated area is occupied by vehicles

The data set used in this paper was collected from thissensor network over a one year period from January 1 toDecember 31 2010 In this paper we chose the 3rd 4th 5thand 6th sensors as the experimental sites An example of traf-fic congestion viewed by a camera sensor studied in this paperis shown in Figure 6 For evaluation purposes anomaliesand their associated delay that took place were independentlyand manually logged by a team of transportation expertsA total of 16 nonrecurring congestion cases were recordedwhich are used to test the anomaly detection capability ofthe proposed algorithm Each record was associated with thetime instances prior to the disruption occurrence and after ithad ended Besides these 16 cases further investigation withanother team of experts also revealed additional 44 delaycases Therefore a total of 60 delay cases are used to test theestimation operation of the proposed algorithm

We note that the traffic data input of the proposedalgorithm is typical data that most roadside sensors forexample loop detectors already produce Therefore weanticipate that if the sensors are to be replaced with othertypes of sensors the performance of the system should notchange significantly This is worth further investigation forlarge scale deployment

512 NGSIM Data To show the potential of being deployedat a different location the proposed algorithm is also assessedusingNGSIM data [19]The selected data set was collected onthe southbound US Highway 101 in Los Angeles California

8 Journal of Advanced Transportation

USA at 0750 amndash0835 am This data set is selected becauseit was collected during the buildup of congestion and consistsof shockwaves which can potentially induce additional delayIt should be noted that the anomaly cases and the amountof delay caused by them were not labeled in this data setso the full algorithm could not be deployed for this dataset However the potential of the algorithm could be shownby shockwave detection To assess the proposed algorithmfour virtual sensors are placed approximately 01 km apart toobtain traffic data to capture the occurrence of shockwavesand their spatial extents Similar to conventional fixed trafficsensors (eg camera sensors and loop detectors) thesevirtual sensors calculate traffic data based on traffic flow andspeed that pass the sensorsrsquo locations An input time windowsize 119871 of 5 minutes is used where time series of 30-secondaverage speeds obtained from upstream and downstreamvirtual sensors are used as inputs The occurrences of shock-waves are observed and recorded by a team of transportationresearchers [25]

52 Benchmark Algorithm To assess the capability of theproposed algorithm in respect to existing approaches abenchmark algorithm is used [17]This benchmark algorithmis selected because it has a similar objective to the proposedalgorithm in detecting anomalies that cause delay Also thisalgorithm has already been implemented in a real-worldanomaly detection system [17] Furthermore composed ofa forecasting component and a detection logic componentit is designed to assess the difference between traffic dataobserved at consecutive roadside sensors

However the main difference between the benchmarkalgorithm and the proposed algorithm lies in the detectionlogic The benchmark algorithm is designed to detect sud-den significant changes of traffic inputs from the forecastcomponent values by establishing threshold values (speedand occupancy) to filter out gradual changes in traffic inputscaused by recurring congestion and an anomalous signalwill only be raised when the specified threshold values havebeen exceeded On the other hand the proposed algorithm isdesigned to detect both sudden and gradual changes whichenables it to detect different types of anomalies with fewerfalse alarms This difference will become more apparentwhen both algorithms are assessed with real-world data inSection 6

53 Performance Evaluation Parameters To numericallyevaluate the proposed algorithm we use detection rate (DR)false positive rate (FPR) mean time to detection (MTTD)andmean absolute percentage error (MAPE) as shown in (3)(4) (5) and (6) respectively DR FPR and MTTD are usedto assess the anomaly detection operation of the proposedalgorithm MAPE is used to assess the accuracy of delayestimation of a total 119863 recorded delay cases where the errorof the estimation 119889119899119895 of the 119895th recorded delay at time 119899 isnormalized by the actual delay119889119886119899119895 tominimize the possibilityof some estimated delay samples from dominating the others

DR =Number of delays detectedTotal number of delays

(3)

FPR

=Number of detected delays that are false positives

Total number of detected delays

(4)

MTTD

=Sum of time taken to detect individual delays

Total number of detected delays

(5)

MAPE = 1119863

119863

sum119895=1

10038161003816100381610038161003816119889119886119899119895 minus 119889119899119895

10038161003816100381610038161003816119889119886119899119895 (6)

The performance parameters are applicable for data setsthat contain labels for traffic anomalies and the amount ofdelay caused by the anomalies Therefore these parameterswill be evaluated for the real-world data collected inThailandbut due to the lack of anomaly labels could not be applied forthe NGSIM data set

6 Results and Discussions

In this section the proposed algorithm is assessed usingreal-world data described in Section 51 and compared tothe benchmark algorithm described in Section 52 Bothalgorithms are tested as if they operate online where eachdetection operation is performed per input time window 119871

61 Assessment Using NGSIM Data The proposed algorithmis assessed on its capability to detect the the presence ofshockwaves which refers to a phenomenon where vehicleshave to temporarily slow down below normal speed andsubsequently force the vehicles behind to slow down furtherThis phenomenon is known to create shockwave fronts ormoving speed drop areas that move backward disruptingflow and annoying drivers behind [35] The occurrence ofshockwaves during congestion not only causes additionaldelay but also increases chances of rear-end collisionsIdentifying shockwaves during congestion also presents achallenge as it occurs temporarily and requires an algorithmthat is fast and sensitive enough to detect them [25] Theoccurrence of shockwaves represents unexpected events sothey are considered as anomalies in this paper

Figure 7 shows that the proposed algorithm can detectall four occurrences (Figure 7(a)) of shockwaves whilebenchmark algorithm can detect only two (Figure 7(b))As shown in Figure 7(b) only abrupt changes in speeddifference (second and third occurrences) can be detectedby the benchmark algorithm This algorithm is designedto detect consecutive differences between upstream anddownstream traffic measurements so it operates well withconsecutive changes in speed difference that are significantenough On the other hand benchmark algorithmmisses thefirst and fourth occurrences of shockwaves in Figure 7(b) asthey are associated with relatively more gradual changes ofspeed difference Unlike benchmark algorithm the proposedalgorithm uses DTW to assess input patterns to detect delayoccurrences so it can detect all four occurrences associated

Journal of Advanced Transportation 9

Upstream sensorDownstream sensor

805 820 835750Time (hoursminutes)

01020304050607080

Spee

d (k

mh

)

(a)

One time step aheadTwo time steps ahead

805 820 835750Time (hoursminutes)

minus3minus2minus1

0123

Spee

d di

ffere

nces

(km

h)

(b)

Figure 7 Detecting delay occurrences associated with shockwaves in NGSIM data using (a) the proposed algorithm and (b) benchmarkalgorithm [17] Dashed vertical lines denote the time points where the algorithms detect delay occurrences Red squares denote the durationwhere the occurrences of shockwaves are observed

with both abrupt and gradual speed changes as shown inFigure 7(a)

Once the occurrences of shockwaves are detected theproposed algorithm proceeds to estimate delay as describedin Section 44 Using the spatial extent estimation describedin Section 44 out of four sensors three sensors are usuallyfound to detect speed drops associated with shockwaves at atime Therefore the most downstream sensor and the mostupstream sensor are selected as the start and end sensorsrespectively and 119897119899119875 is found to be 048 km

Table 1 shows the values of V119899119875 V119899119873 120591119899119875 and 120591119899119873 in (1)for each of the four shockwave occurrences For comparisonpurposes delay estimation is also applied for the second andthird occurrences that are detected by benchmark algorithm(see Figure 7) Note that since benchmark algorithm cannotdetect the first and fourth occurrences estimation operationis not activated during these times and normal speed andtravel times are used instead We further note that all fouroccurrences and their associated travel times and delays canbe calculated consecutively online by the proposed algorithm

As shown in Table 1 the four occurrences of shockwavesdetected in Figure 7 are associated with travel times (120591119899119875 in(1)) of 5682 s 5554 s 6098 s and 5379 s respectively Theproposed algorithm is able to detect all four occurrences andthese estimated travel times would all be included in the delaycalculation On the other hand if the benchmark algorithmwas used instead only the second and third occurrences andtheir associated travel times would be reported resulting inless accurate delay calculationWe note that even though thedifference between the delays obtained from the proposedalgorithm and benchmark algorithm in this example maybe small significant improvement can be seen when higherdegree of disruptions are assessed

The results shown in Table 1 have many applicationsin advanced traveler information systems and advancedtraffic management systems A direct application would beto estimate travel time of the freeway segment and report itto motorists The ability to detect occurrences of shockwavescan also be used in other applications such as calculatingshockwave propagation time [25] It can also be used to warn

motorists to adjust intervehicle spacing to avoid rear-endcollisions

62 Assessment Using Real-World Data Collected in ThailandThis section assesses the capability of the proposed algorithmand the benchmark algorithm [17] to detect the sixteenrecorded traffic anomalies in the real-world data set describedin Section 51 As inputs both algorithms use speed flowand occupancy collected by roadside sensorsThe assessmentfollows a 119896-fold cross-validation technique The real-worlddata set consists of 16 anomaly cases 1198631 1198632 11986316 soevaluation is performed 16 times At an iteration 119894 119894 =1 2 16 the anomaly case119863119894 is reserved as a test set whilethe other anomaly cases are collectively used to calibratethe proposed algorithm and benchmark algorithm That isin the first iteration 1198632 11986316 collectively serve as thetraining set which is then tested with1198631Then in the seconditeration 1198631 1198633 11986316 are used for training while 1198632 isused for testing and so on

Table 2 shows that the proposed algorithm achieveshigher detection rate and lower false alarm rate than thebenchmark algorithm Particularly compared to benchmarkalgorithm the proposed algorithm improves the detectionrate and false positive rate by approximately 10 and 7respectively Figures 8 and 9 show examples of anomaliesdetected by the proposed algorithm and found to causetraffic delay Based on the records these anomalies areassociated with disabled vehicles parking over the shoulderarea

It is found that for the benchmark algorithm miss-detection usually occurs under low traffic flow This algo-rithm always needs data from at least two sensors and themagnitude of traffic flow measured by the roadside sensorsis not large enough to trigger the benchmark algorithmWhile medium to heavy traffic flow triggers the benchmarkalgorithm it also increases false alarms This algorithm isdesigned to directly assess the difference between upstreamand downstream sensors of the roadside sensors whichsubsequently increases its sensitivity and triggers many falsealarms

10 Journal of Advanced Transportation

Table 1 Estimation results onNGSIMdata using the proposed algorithm and benchmark algorithm 119897119899119875 =048 kmNote that if the shockwaveis not detected the estimated travel time 120591119899119875 would be equal to normal travel time 120591119899119873

1st 2nd 3rd 4thProposed algorithmShockwaves detected Yes Yes Yes YesV119899119875 (see (1)) 3041 kmh 3112 kmh 2834 kmh 3213 kmh120591119899119875 (see (1)) 5682 s 5554 s 6098 s 5379 s119889119899 = 120591119899119875 minus 120591119899119873 1841 s 1893 s 2399 s 541 sBenchmark algorithm [17]Shockwaves detected No Yes Yes NoV119899119875 (see (1)) 4499 kmh 3112 kmh 2834 kmh 3572 kmh120591119899119875 (see (1)) 3841 s 5554 s 6098 s 4838 s119889119899 = 120591119899119875 minus 120591119899119873 000 s 1893 s 2399 s 000 sIf all 4 shockwaves are not detectedV119899119873 (see (1)) 4499 kmh 4721 kmh 4671 kmh 3572 kmh120591119899119873 (see (1)) 3841 s 3660 s 3700 s 4838 s

Table 2 Results on anomaly detection using data set collected inThailand

Performanceparameters

Proposedalgorithm

Benchmarkalgorithm [17]

DR () 94 83FPR () 5 12MTTD (seconds) 340 140

Figure 8 An anomaly detected at the 4th camera sensor

Unlike the benchmark algorithm the proposed algorithmuses DTW to assess traffic data from each sensor individuallyThis enables the proposed algorithm to recognize patternsassociated with anomalies more effectively including underlow traffic flow The use of DTW also reduces a numberof false alarms as it ensures that unrecognized patternswill not trigger detection at individual sensors In respectto mean time to detection (MTTD) it takes longer forthe proposed algorithm to detect anomalies compared tobenchmark algorithm as it needs to wait till enough temporalsamples of traffic data have been collected in each sensorbefore the detection operation can be triggered Howeverthe mean time to detection is still less than 6 minutes which

Figure 9 Another anomaly detected at the 4th camera sensor

should give enough time for traffic operators to response [8]To reduce mean time to detection of the proposed algorithmshorter time samples of traffic data can be used ReducingMTTD while maintaining acceptable levels of DR and FPRis worth further investigation

To deploy the proposed algorithm computation timeof the algorithm could be an added delay to the MTTDEven though the computation time was not systematicallymeasured for the performance evaluation the computationtime was in order of seconds For the computerrsquos specificationthat we use which was a personal laptop with 25 GHz CPUand 4GB of RAM the computation time was approximately[1 3) seconds As the computing time (a few seconds) wasmuch smaller than the mean time to detection (up to 6minutes) it is reasonable to consider the computing timenegligible

This section further discusses the results when the pro-posed algorithm was applied to estimate 60 delay casesdescribed in Section 51 using different input time windowsizes (119871) and prediction horizons (119875) For practical purposesit is important to select suitable input time window sizeand prediction horizon to enable the proposed method to

Journal of Advanced Transportation 11

L = 3 L = 10L = 5

P = 5

P = 10

P = 15

0

005

01

015

02

025

03

035

04M

APE

Figure 10Mean absolute percentage error (MAPE) fromestimating60 delay cases using different input time window sizes 119871 (minutes)and prediction horizons 119875 (minutes)

perform effectively Even though the 15-minute interval isoften recommended [9] to avoid strong fluctuation usinglower input timewindow sizes can providemore fine-grainedinformation for practitioners [25] Therefore in this paperwe use 119871 = 3 5 and 10 minutes for input time windowsize For prediction horizons we use 119875 = 5 10 and 15minutes as this paper is focused on short-term estimation(less than 15 minutes) Also changes in traffic condition onthis experimental site are normally found to occur in less than15 minutes

The proposed algorithm is assessed on its capability toestimate delay using MAPE Figure 10 shows MAPE fromestimating 60 delay cases in the real-world data set withdifferent input timewindow sizes (119871) and prediction horizons(119875) These results show that the estimation part of theproposed algorithm can achieve MAPE lower than 01 forthis set of real-world data This accuracy is achieved becausethe estimation part consists of both spatial and temporalestimations which enable the proposed algorithm to firstestimate the area disrupted by anomalies and then estimatethe delay itself Furthermore it can be seen that increasinginput time window size (119871) reduces MAPE as using moreinput samples enhances estimation accuracy For the impactof prediction horizons (119875) MAPE tends to increase withlarger prediction horizons as it is more difficult to estimatedelays farther ahead into the future However as shownin Figure 10 when 119875 is increased MAPE in fact increasesat a much smaller rate when larger values of 119871 are usedTherefore it is recommended for practitioners that if theobjective is to estimate delays far ahead into the future (usinglarge prediction horizons) larger input time window size isneeded

Based on the findings in this section it is importantthat practitioners select a suitable input time window sizeaccording to their practical objectives If the objective is toprovide prompt delay information small input time windowsize should be used On the other hand if the objective is toprovide accurately estimated delays for traffic managementpurposes larger input time window sizes should be chosenFurthermore based on the proposed algorithm a morecomplex and flexible system can also be further developedFor example a stepwise approach can be implemented on topof the proposed algorithm where a small input time windowsize is used as a trigger to promptly discover the occurrenceof traffic delay Then in the next step the input time windowsize can be increased to collect more traffic data to achievemore accurate detection and estimation of traffic delay

For anomalies with large scale effect beyond the 11 kmexperimental site we anticipate that the proposed algo-rithm should be able to detect those anomalies if theyoccur downstream and cause changes in traffic patterns topropagate upstream into the 11 km segment To accuratelyestimate delay basis patterns associated with such large scaleanomalies need to be incorporated to distinguish themThenan extrapolation technique needs to be incorporated to enablethe proposed algorithm to estimate delay beyond the 11 kmexpressway segment where local traffic information cannotbe directly observed [25] This is worth further investigation

7 Conclusions and Future Work

This paper proposes an algorithm to detect anomalies andestimate delay using traffic data from roadside sensors Theproposed algorithm uses traffic data from individual sensorsto detect anomalies and then uses data from consecutivedownstream and upstream sensors together to spatially andtemporally estimate delay Performance evaluations usingreal-world data show that the proposed algorithm achieveshigher detection rate and lower false positive rate when usedto detect anomalies compared to an existing benchmarkalgorithmAn investigationwithNGSIMdata also shows thatthe proposed algorithm is also able to detect the presence ofshockwaves and subsequently estimate delaymore efficientlythan benchmark algorithm Further results show that theestimation error increases at a much smaller rate whensuitable input time window size and prediction horizon areselected

There are interesting aspects that are worth furtherinvestigation One interesting aspect is to reduce the pro-cessing time while maintaining low detection and estima-tion errors A possible approach is to introduce a featureextraction method that can extract hidden patterns whensmaller input time window sizes are used [8] Alternativeapproaches include studying a set of hidden representativetraffic flow features in advance by pretraining a deep-learningmodel [36] and populating traffic features into a geneticalgorithm learning method to reduce dimensionality [37]Another interesting aspect is to improve inference capabilityby incorporating a spatial inferencemethod to estimate trafficdata itself at the locations where they cannot be directlymeasured [25] Also it would be practically interesting to

12 Journal of Advanced Transportation

assess the proposed algorithm with different road types andgeometries

In addition to improvements of the algorithm an interest-ing aspect to investigate is the applications of the algorithmThe solution provided in this paper is designed to be astreaming algorithm which can be run online thereforeit can be naturally extended to use in various real-timeapplications For example the detected anomaly and esti-mated additional travel time would be useful for both trafficoperators and travelers if the information is delivered ina timely and convenient manner Active communicationbetween roadside sensors and vehicle-to-infrastructure (V2I)could be used to promptly transmit the information from oursolution to nearby roadside message boards to display theinformation to the drivers on the roads or vehicles with V2Icommunication to adapt routing and navigation

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this article

Acknowledgments

The authors would like to thankDr Supakorn Siddhichai andhis IMG team at NECTEC for their assistance in providingreal-world data for the analysis

References

[1] N Parrado and Y Donoso ldquoCongestion based mechanism forroute discovery in a V2I-V2V system applying smart devicesand IoTrdquo Sensors vol 15 no 4 pp 7768ndash7806 2015

[2] K Nellore and G P Hancke ldquoA survey on urban trafficmanagement system using wireless sensor networksrdquo Sensorsvol 16 no 2 article 157 2016

[3] E I Vlahogianni M G Karlaftis and J C Golias ldquoShort-term traffic forecasting where we are and where wersquore goingrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 3ndash19 2014

[4] R Li and G Rose ldquoIncorporating uncertainty into short-term travel time predictionsrdquo Transportation Research Part CEmerging Technologies vol 19 no 6 pp 1006ndash1018 2011

[5] Y Chung ldquoAmethodological approach for estimating temporaland spatial extent of delays caused by freeway accidentsrdquo IEEETransactions on Intelligent Transportation Systems vol 13 no 3pp 1454ndash1461 2012

[6] J Weng and Q Meng ldquoEstimating capacity and traffic delayin work zones an overviewrdquo Transportation ResearchmdashPart CEmerging Technologies vol 35 pp 34ndash45 2013

[7] X Zhang E Onieva A Perallos E Osaba and V C SLee ldquoHierarchical fuzzy rule-based system optimized withgenetic algorithms for short term traffic congestion predictionrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 127ndash142 2014

[8] S Thajchayapong E S Garcia-Trevino and J A Barria ldquoDis-tributed classification of traffic anomalies using microscopictraffic variablesrdquo IEEETransactions on Intelligent TransportationSystems vol 14 no 1 pp 448ndash458 2013

[9] M Lippi M Bertini and P Frasconi ldquoShort-term traffic flowforecasting an experimental comparison of time-series anal-ysis and supervised learningrdquo IEEE Transactions on IntelligentTransportation Systems vol 14 no 2 pp 871ndash882 2013

[10] F Zheng andH Van Zuylen ldquoUrban link travel time estimationbased on sparse probe vehicle datardquo Transportation ResearchmdashPart C Emerging Technologies vol 31 pp 145ndash157 2013

[11] F Dion and H Rakha ldquoEstimating dynamic roadway traveltimes using automatic vehicle identification data for low sam-pling ratesrdquoTransportation Research Part BMethodological vol40 no 9 pp 745ndash766 2006

[12] D Paterson and G Rose ldquoDynamic travel time estima-tion on instrumented freewaysrdquo in Proceedings of the 6thWorld Congress on Intelligent Transportation Systems TorontoCanada 1999

[13] J Kwon B Coifman and P Bickel ldquoDay-to-day travel-timetrends and travel-time prediction from loop-detector datardquoTransportation Research Record no 1717 pp 120ndash129 2000

[14] J Rice and E Van Zwet ldquoA simple and effective method forpredicting travel times on freewaysrdquo IEEE Transactions onIntelligent Transportation Systems vol 5 no 3 pp 200ndash2072004

[15] C D R Lindveld RThijs P H L Bovy andN J van der ZijppldquoEvaluation of online travel time estimators and predictorsrdquoTransportation Research Record vol 1719 pp 45ndash53 2000

[16] H Yang S Chen YWang BWu andZWang ldquoComparison ofdelay estimation models for signalised intersections using fieldobservations in Shanghairdquo IET Intelligent Transport Systems vol10 no 3 pp 165ndash174 2016

[17] C L Mak and H S L Fan ldquoDevelopment of dual-stationautomated expressway incident detection algorithmsrdquo IEEETransactions on Intelligent Transportation Systems vol 8 no 3pp 480ndash490 2007

[18] K Hi-Ri-O-Tappa C Likitkhajorn A Poolsawat and S Tha-jchayapong ldquoTraffic incident detection system using series ofpoint detectorsrdquo in Proceedings of the 15th International IEEEConference on Intelligent Transportation Systems (ITSC rsquo12) pp182ndash187 Anchorage Alaska USA September 2012

[19] V Alexiadis J Colyar J Halkias R Hranac and G McHaleldquoThe next generation simulation programrdquo ITE Journal vol 74no 8 pp 22ndash26 2004

[20] F G Habtemichael M Cetin and K A Anuar ldquoIncident-induced delays on freeways quantificationmethod by groupingsimilar traffic patternsrdquo Transportation Research Record vol2484 pp 60ndash69 2015

[21] J Li C-J Lan and X Gu ldquoEstimation of incident delay andits uncertainty on freeway networksrdquo Transportation ResearchRecord vol 1959 pp 37ndash45 2006

[22] Y Chung ldquoAssessment of non-recurrent congestion causedby precipitation using archived weather and traffic flow datardquoTransport Policy vol 19 no 1 pp 167ndash173 2012

[23] J A Barria and SThajchayapong ldquoDetection and classificationof traffic anomalies using microscopic traffic variablesrdquo IEEETransactions on Intelligent Transportation Systems vol 12 no3 pp 695ndash704 2011

[24] K HiriOtappa and S Thajchayapong ldquoGeneralizability andtransferability of incident detection algorithm using dynamictime warpingrdquo in Proceedings of the in 19th World Congress onIntelligent Transportation Systems Vienna Austria 2012

Journal of Advanced Transportation 13

[25] S Thajchayapong and J A Barria ldquoSpatial inference of traffictransition using microndashmacro traffic variablesrdquo IEEE Transac-tions on Intelligent Transportation Systems vol 16 no 2 pp854ndash864 2015

[26] C A Ratanamahatana and E Keogh ldquoEverything you knowabout dynamic time warping is wrongrdquo in Proceedings of the3rd Workshop on Mining Temporal and Sequential Data in Con-junction with the 10th ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining (KDD rsquo04) CiteseerSeattle Wash USA August 2004

[27] Transportation Research Thesauras Transportation ResearchBoard National Academy httptrttrborgtrtaspNN=Bthd

[28] N Chiabaut C Buisson and L Leclercq ldquoFundamental dia-gram estimation through passing rate measurements in conges-tionrdquo IEEE Transactions on Intelligent Transportation Systemsvol 10 no 2 pp 355ndash359 2009

[29] J J Fernandez-Lozano M Martın-Guzman J Martın-Avilaand A Garcıa-Cerezo ldquoA wireless sensor network for urbantraffic characterization and trend monitoringrdquo Sensors vol 15no 10 pp 26143ndash26169 2015

[30] J-S Oh C Oh S G Ritchie and M Chang ldquoReal-timeestimation of accident likelihood for safety enhancementrdquoJournal of Transportation Engineering vol 131 no 5 pp 358ndash363 2005

[31] A P Chassiakos and Y J Stephanedes ldquoSmoothing algorithmsfor incident detectionrdquo Transportation Research Record vol1394 pp 8ndash16 1993

[32] V Alarcon-Aquino and J A Barria ldquoAnomaly detection incommunication networks using waveletsrdquo IEE ProceedingsmdashCommunications vol 148 no 6 pp 355ndash362 2001

[33] D Berndt and J Clifford ldquoUsing dynamic time warping tofind patterns in time seriesrdquo in Proceedings of the Workshop onKnowledge Discovery in Databases (KDD rsquo94) vol 10 pp 359ndash370 Seattle Wash USA August 1994

[34] K Kiratiratanapruk and S Siddhichai ldquoVehicle detection andtracking for traffic monitoring systemrdquo in Proceedings of theIEEE Region 10 Conference (TENCON rsquo06) pp 1ndash4 IEEE HongKong November 2006

[35] A Hegyi B De Schutter and J Hellendoorn ldquoOptimal coordi-nation of variable speed limits to suppress shock wavesrdquo IEEETransactions on Intelligent Transportation Systems vol 6 no 1pp 102ndash112 2005

[36] Y Lv Y Duan W Kang Z Li and F-Y Wang ldquoTraffic flowprediction with big data a deep learning approachrdquo IEEETransactions on Intelligent Transportation Systems vol 16 no2 pp 865ndash873 2015

[37] P Lopez-Garcia E Onieva E Osaba A D Masegosa and APerallos ldquoA hybrid method for short-term traffic congestionforecasting using genetic algorithms and cross entropyrdquo IEEETransactions on Intelligent Transportation Systems vol 17 no 2pp 557ndash569 2016

8 Journal of Advanced Transportation

USA at 0750 amndash0835 am This data set is selected becauseit was collected during the buildup of congestion and consistsof shockwaves which can potentially induce additional delayIt should be noted that the anomaly cases and the amountof delay caused by them were not labeled in this data setso the full algorithm could not be deployed for this dataset However the potential of the algorithm could be shownby shockwave detection To assess the proposed algorithmfour virtual sensors are placed approximately 01 km apart toobtain traffic data to capture the occurrence of shockwavesand their spatial extents Similar to conventional fixed trafficsensors (eg camera sensors and loop detectors) thesevirtual sensors calculate traffic data based on traffic flow andspeed that pass the sensorsrsquo locations An input time windowsize 119871 of 5 minutes is used where time series of 30-secondaverage speeds obtained from upstream and downstreamvirtual sensors are used as inputs The occurrences of shock-waves are observed and recorded by a team of transportationresearchers [25]

52 Benchmark Algorithm To assess the capability of theproposed algorithm in respect to existing approaches abenchmark algorithm is used [17]This benchmark algorithmis selected because it has a similar objective to the proposedalgorithm in detecting anomalies that cause delay Also thisalgorithm has already been implemented in a real-worldanomaly detection system [17] Furthermore composed ofa forecasting component and a detection logic componentit is designed to assess the difference between traffic dataobserved at consecutive roadside sensors

However the main difference between the benchmarkalgorithm and the proposed algorithm lies in the detectionlogic The benchmark algorithm is designed to detect sud-den significant changes of traffic inputs from the forecastcomponent values by establishing threshold values (speedand occupancy) to filter out gradual changes in traffic inputscaused by recurring congestion and an anomalous signalwill only be raised when the specified threshold values havebeen exceeded On the other hand the proposed algorithm isdesigned to detect both sudden and gradual changes whichenables it to detect different types of anomalies with fewerfalse alarms This difference will become more apparentwhen both algorithms are assessed with real-world data inSection 6

53 Performance Evaluation Parameters To numericallyevaluate the proposed algorithm we use detection rate (DR)false positive rate (FPR) mean time to detection (MTTD)andmean absolute percentage error (MAPE) as shown in (3)(4) (5) and (6) respectively DR FPR and MTTD are usedto assess the anomaly detection operation of the proposedalgorithm MAPE is used to assess the accuracy of delayestimation of a total 119863 recorded delay cases where the errorof the estimation 119889119899119895 of the 119895th recorded delay at time 119899 isnormalized by the actual delay119889119886119899119895 tominimize the possibilityof some estimated delay samples from dominating the others

DR =Number of delays detectedTotal number of delays

(3)

FPR

=Number of detected delays that are false positives

Total number of detected delays

(4)

MTTD

=Sum of time taken to detect individual delays

Total number of detected delays

(5)

MAPE = 1119863

119863

sum119895=1

10038161003816100381610038161003816119889119886119899119895 minus 119889119899119895

10038161003816100381610038161003816119889119886119899119895 (6)

The performance parameters are applicable for data setsthat contain labels for traffic anomalies and the amount ofdelay caused by the anomalies Therefore these parameterswill be evaluated for the real-world data collected inThailandbut due to the lack of anomaly labels could not be applied forthe NGSIM data set

6 Results and Discussions

In this section the proposed algorithm is assessed usingreal-world data described in Section 51 and compared tothe benchmark algorithm described in Section 52 Bothalgorithms are tested as if they operate online where eachdetection operation is performed per input time window 119871

61 Assessment Using NGSIM Data The proposed algorithmis assessed on its capability to detect the the presence ofshockwaves which refers to a phenomenon where vehicleshave to temporarily slow down below normal speed andsubsequently force the vehicles behind to slow down furtherThis phenomenon is known to create shockwave fronts ormoving speed drop areas that move backward disruptingflow and annoying drivers behind [35] The occurrence ofshockwaves during congestion not only causes additionaldelay but also increases chances of rear-end collisionsIdentifying shockwaves during congestion also presents achallenge as it occurs temporarily and requires an algorithmthat is fast and sensitive enough to detect them [25] Theoccurrence of shockwaves represents unexpected events sothey are considered as anomalies in this paper

Figure 7 shows that the proposed algorithm can detectall four occurrences (Figure 7(a)) of shockwaves whilebenchmark algorithm can detect only two (Figure 7(b))As shown in Figure 7(b) only abrupt changes in speeddifference (second and third occurrences) can be detectedby the benchmark algorithm This algorithm is designedto detect consecutive differences between upstream anddownstream traffic measurements so it operates well withconsecutive changes in speed difference that are significantenough On the other hand benchmark algorithmmisses thefirst and fourth occurrences of shockwaves in Figure 7(b) asthey are associated with relatively more gradual changes ofspeed difference Unlike benchmark algorithm the proposedalgorithm uses DTW to assess input patterns to detect delayoccurrences so it can detect all four occurrences associated

Journal of Advanced Transportation 9

Upstream sensorDownstream sensor

805 820 835750Time (hoursminutes)

01020304050607080

Spee

d (k

mh

)

(a)

One time step aheadTwo time steps ahead

805 820 835750Time (hoursminutes)

minus3minus2minus1

0123

Spee

d di

ffere

nces

(km

h)

(b)

Figure 7 Detecting delay occurrences associated with shockwaves in NGSIM data using (a) the proposed algorithm and (b) benchmarkalgorithm [17] Dashed vertical lines denote the time points where the algorithms detect delay occurrences Red squares denote the durationwhere the occurrences of shockwaves are observed

with both abrupt and gradual speed changes as shown inFigure 7(a)

Once the occurrences of shockwaves are detected theproposed algorithm proceeds to estimate delay as describedin Section 44 Using the spatial extent estimation describedin Section 44 out of four sensors three sensors are usuallyfound to detect speed drops associated with shockwaves at atime Therefore the most downstream sensor and the mostupstream sensor are selected as the start and end sensorsrespectively and 119897119899119875 is found to be 048 km

Table 1 shows the values of V119899119875 V119899119873 120591119899119875 and 120591119899119873 in (1)for each of the four shockwave occurrences For comparisonpurposes delay estimation is also applied for the second andthird occurrences that are detected by benchmark algorithm(see Figure 7) Note that since benchmark algorithm cannotdetect the first and fourth occurrences estimation operationis not activated during these times and normal speed andtravel times are used instead We further note that all fouroccurrences and their associated travel times and delays canbe calculated consecutively online by the proposed algorithm

As shown in Table 1 the four occurrences of shockwavesdetected in Figure 7 are associated with travel times (120591119899119875 in(1)) of 5682 s 5554 s 6098 s and 5379 s respectively Theproposed algorithm is able to detect all four occurrences andthese estimated travel times would all be included in the delaycalculation On the other hand if the benchmark algorithmwas used instead only the second and third occurrences andtheir associated travel times would be reported resulting inless accurate delay calculationWe note that even though thedifference between the delays obtained from the proposedalgorithm and benchmark algorithm in this example maybe small significant improvement can be seen when higherdegree of disruptions are assessed

The results shown in Table 1 have many applicationsin advanced traveler information systems and advancedtraffic management systems A direct application would beto estimate travel time of the freeway segment and report itto motorists The ability to detect occurrences of shockwavescan also be used in other applications such as calculatingshockwave propagation time [25] It can also be used to warn

motorists to adjust intervehicle spacing to avoid rear-endcollisions

62 Assessment Using Real-World Data Collected in ThailandThis section assesses the capability of the proposed algorithmand the benchmark algorithm [17] to detect the sixteenrecorded traffic anomalies in the real-world data set describedin Section 51 As inputs both algorithms use speed flowand occupancy collected by roadside sensorsThe assessmentfollows a 119896-fold cross-validation technique The real-worlddata set consists of 16 anomaly cases 1198631 1198632 11986316 soevaluation is performed 16 times At an iteration 119894 119894 =1 2 16 the anomaly case119863119894 is reserved as a test set whilethe other anomaly cases are collectively used to calibratethe proposed algorithm and benchmark algorithm That isin the first iteration 1198632 11986316 collectively serve as thetraining set which is then tested with1198631Then in the seconditeration 1198631 1198633 11986316 are used for training while 1198632 isused for testing and so on

Table 2 shows that the proposed algorithm achieveshigher detection rate and lower false alarm rate than thebenchmark algorithm Particularly compared to benchmarkalgorithm the proposed algorithm improves the detectionrate and false positive rate by approximately 10 and 7respectively Figures 8 and 9 show examples of anomaliesdetected by the proposed algorithm and found to causetraffic delay Based on the records these anomalies areassociated with disabled vehicles parking over the shoulderarea

It is found that for the benchmark algorithm miss-detection usually occurs under low traffic flow This algo-rithm always needs data from at least two sensors and themagnitude of traffic flow measured by the roadside sensorsis not large enough to trigger the benchmark algorithmWhile medium to heavy traffic flow triggers the benchmarkalgorithm it also increases false alarms This algorithm isdesigned to directly assess the difference between upstreamand downstream sensors of the roadside sensors whichsubsequently increases its sensitivity and triggers many falsealarms

10 Journal of Advanced Transportation

Table 1 Estimation results onNGSIMdata using the proposed algorithm and benchmark algorithm 119897119899119875 =048 kmNote that if the shockwaveis not detected the estimated travel time 120591119899119875 would be equal to normal travel time 120591119899119873

1st 2nd 3rd 4thProposed algorithmShockwaves detected Yes Yes Yes YesV119899119875 (see (1)) 3041 kmh 3112 kmh 2834 kmh 3213 kmh120591119899119875 (see (1)) 5682 s 5554 s 6098 s 5379 s119889119899 = 120591119899119875 minus 120591119899119873 1841 s 1893 s 2399 s 541 sBenchmark algorithm [17]Shockwaves detected No Yes Yes NoV119899119875 (see (1)) 4499 kmh 3112 kmh 2834 kmh 3572 kmh120591119899119875 (see (1)) 3841 s 5554 s 6098 s 4838 s119889119899 = 120591119899119875 minus 120591119899119873 000 s 1893 s 2399 s 000 sIf all 4 shockwaves are not detectedV119899119873 (see (1)) 4499 kmh 4721 kmh 4671 kmh 3572 kmh120591119899119873 (see (1)) 3841 s 3660 s 3700 s 4838 s

Table 2 Results on anomaly detection using data set collected inThailand

Performanceparameters

Proposedalgorithm

Benchmarkalgorithm [17]

DR () 94 83FPR () 5 12MTTD (seconds) 340 140

Figure 8 An anomaly detected at the 4th camera sensor

Unlike the benchmark algorithm the proposed algorithmuses DTW to assess traffic data from each sensor individuallyThis enables the proposed algorithm to recognize patternsassociated with anomalies more effectively including underlow traffic flow The use of DTW also reduces a numberof false alarms as it ensures that unrecognized patternswill not trigger detection at individual sensors In respectto mean time to detection (MTTD) it takes longer forthe proposed algorithm to detect anomalies compared tobenchmark algorithm as it needs to wait till enough temporalsamples of traffic data have been collected in each sensorbefore the detection operation can be triggered Howeverthe mean time to detection is still less than 6 minutes which

Figure 9 Another anomaly detected at the 4th camera sensor

should give enough time for traffic operators to response [8]To reduce mean time to detection of the proposed algorithmshorter time samples of traffic data can be used ReducingMTTD while maintaining acceptable levels of DR and FPRis worth further investigation

To deploy the proposed algorithm computation timeof the algorithm could be an added delay to the MTTDEven though the computation time was not systematicallymeasured for the performance evaluation the computationtime was in order of seconds For the computerrsquos specificationthat we use which was a personal laptop with 25 GHz CPUand 4GB of RAM the computation time was approximately[1 3) seconds As the computing time (a few seconds) wasmuch smaller than the mean time to detection (up to 6minutes) it is reasonable to consider the computing timenegligible

This section further discusses the results when the pro-posed algorithm was applied to estimate 60 delay casesdescribed in Section 51 using different input time windowsizes (119871) and prediction horizons (119875) For practical purposesit is important to select suitable input time window sizeand prediction horizon to enable the proposed method to

Journal of Advanced Transportation 11

L = 3 L = 10L = 5

P = 5

P = 10

P = 15

0

005

01

015

02

025

03

035

04M

APE

Figure 10Mean absolute percentage error (MAPE) fromestimating60 delay cases using different input time window sizes 119871 (minutes)and prediction horizons 119875 (minutes)

perform effectively Even though the 15-minute interval isoften recommended [9] to avoid strong fluctuation usinglower input timewindow sizes can providemore fine-grainedinformation for practitioners [25] Therefore in this paperwe use 119871 = 3 5 and 10 minutes for input time windowsize For prediction horizons we use 119875 = 5 10 and 15minutes as this paper is focused on short-term estimation(less than 15 minutes) Also changes in traffic condition onthis experimental site are normally found to occur in less than15 minutes

The proposed algorithm is assessed on its capability toestimate delay using MAPE Figure 10 shows MAPE fromestimating 60 delay cases in the real-world data set withdifferent input timewindow sizes (119871) and prediction horizons(119875) These results show that the estimation part of theproposed algorithm can achieve MAPE lower than 01 forthis set of real-world data This accuracy is achieved becausethe estimation part consists of both spatial and temporalestimations which enable the proposed algorithm to firstestimate the area disrupted by anomalies and then estimatethe delay itself Furthermore it can be seen that increasinginput time window size (119871) reduces MAPE as using moreinput samples enhances estimation accuracy For the impactof prediction horizons (119875) MAPE tends to increase withlarger prediction horizons as it is more difficult to estimatedelays farther ahead into the future However as shownin Figure 10 when 119875 is increased MAPE in fact increasesat a much smaller rate when larger values of 119871 are usedTherefore it is recommended for practitioners that if theobjective is to estimate delays far ahead into the future (usinglarge prediction horizons) larger input time window size isneeded

Based on the findings in this section it is importantthat practitioners select a suitable input time window sizeaccording to their practical objectives If the objective is toprovide prompt delay information small input time windowsize should be used On the other hand if the objective is toprovide accurately estimated delays for traffic managementpurposes larger input time window sizes should be chosenFurthermore based on the proposed algorithm a morecomplex and flexible system can also be further developedFor example a stepwise approach can be implemented on topof the proposed algorithm where a small input time windowsize is used as a trigger to promptly discover the occurrenceof traffic delay Then in the next step the input time windowsize can be increased to collect more traffic data to achievemore accurate detection and estimation of traffic delay

For anomalies with large scale effect beyond the 11 kmexperimental site we anticipate that the proposed algo-rithm should be able to detect those anomalies if theyoccur downstream and cause changes in traffic patterns topropagate upstream into the 11 km segment To accuratelyestimate delay basis patterns associated with such large scaleanomalies need to be incorporated to distinguish themThenan extrapolation technique needs to be incorporated to enablethe proposed algorithm to estimate delay beyond the 11 kmexpressway segment where local traffic information cannotbe directly observed [25] This is worth further investigation

7 Conclusions and Future Work

This paper proposes an algorithm to detect anomalies andestimate delay using traffic data from roadside sensors Theproposed algorithm uses traffic data from individual sensorsto detect anomalies and then uses data from consecutivedownstream and upstream sensors together to spatially andtemporally estimate delay Performance evaluations usingreal-world data show that the proposed algorithm achieveshigher detection rate and lower false positive rate when usedto detect anomalies compared to an existing benchmarkalgorithmAn investigationwithNGSIMdata also shows thatthe proposed algorithm is also able to detect the presence ofshockwaves and subsequently estimate delaymore efficientlythan benchmark algorithm Further results show that theestimation error increases at a much smaller rate whensuitable input time window size and prediction horizon areselected

There are interesting aspects that are worth furtherinvestigation One interesting aspect is to reduce the pro-cessing time while maintaining low detection and estima-tion errors A possible approach is to introduce a featureextraction method that can extract hidden patterns whensmaller input time window sizes are used [8] Alternativeapproaches include studying a set of hidden representativetraffic flow features in advance by pretraining a deep-learningmodel [36] and populating traffic features into a geneticalgorithm learning method to reduce dimensionality [37]Another interesting aspect is to improve inference capabilityby incorporating a spatial inferencemethod to estimate trafficdata itself at the locations where they cannot be directlymeasured [25] Also it would be practically interesting to

12 Journal of Advanced Transportation

assess the proposed algorithm with different road types andgeometries

In addition to improvements of the algorithm an interest-ing aspect to investigate is the applications of the algorithmThe solution provided in this paper is designed to be astreaming algorithm which can be run online thereforeit can be naturally extended to use in various real-timeapplications For example the detected anomaly and esti-mated additional travel time would be useful for both trafficoperators and travelers if the information is delivered ina timely and convenient manner Active communicationbetween roadside sensors and vehicle-to-infrastructure (V2I)could be used to promptly transmit the information from oursolution to nearby roadside message boards to display theinformation to the drivers on the roads or vehicles with V2Icommunication to adapt routing and navigation

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this article

Acknowledgments

The authors would like to thankDr Supakorn Siddhichai andhis IMG team at NECTEC for their assistance in providingreal-world data for the analysis

References

[1] N Parrado and Y Donoso ldquoCongestion based mechanism forroute discovery in a V2I-V2V system applying smart devicesand IoTrdquo Sensors vol 15 no 4 pp 7768ndash7806 2015

[2] K Nellore and G P Hancke ldquoA survey on urban trafficmanagement system using wireless sensor networksrdquo Sensorsvol 16 no 2 article 157 2016

[3] E I Vlahogianni M G Karlaftis and J C Golias ldquoShort-term traffic forecasting where we are and where wersquore goingrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 3ndash19 2014

[4] R Li and G Rose ldquoIncorporating uncertainty into short-term travel time predictionsrdquo Transportation Research Part CEmerging Technologies vol 19 no 6 pp 1006ndash1018 2011

[5] Y Chung ldquoAmethodological approach for estimating temporaland spatial extent of delays caused by freeway accidentsrdquo IEEETransactions on Intelligent Transportation Systems vol 13 no 3pp 1454ndash1461 2012

[6] J Weng and Q Meng ldquoEstimating capacity and traffic delayin work zones an overviewrdquo Transportation ResearchmdashPart CEmerging Technologies vol 35 pp 34ndash45 2013

[7] X Zhang E Onieva A Perallos E Osaba and V C SLee ldquoHierarchical fuzzy rule-based system optimized withgenetic algorithms for short term traffic congestion predictionrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 127ndash142 2014

[8] S Thajchayapong E S Garcia-Trevino and J A Barria ldquoDis-tributed classification of traffic anomalies using microscopictraffic variablesrdquo IEEETransactions on Intelligent TransportationSystems vol 14 no 1 pp 448ndash458 2013

[9] M Lippi M Bertini and P Frasconi ldquoShort-term traffic flowforecasting an experimental comparison of time-series anal-ysis and supervised learningrdquo IEEE Transactions on IntelligentTransportation Systems vol 14 no 2 pp 871ndash882 2013

[10] F Zheng andH Van Zuylen ldquoUrban link travel time estimationbased on sparse probe vehicle datardquo Transportation ResearchmdashPart C Emerging Technologies vol 31 pp 145ndash157 2013

[11] F Dion and H Rakha ldquoEstimating dynamic roadway traveltimes using automatic vehicle identification data for low sam-pling ratesrdquoTransportation Research Part BMethodological vol40 no 9 pp 745ndash766 2006

[12] D Paterson and G Rose ldquoDynamic travel time estima-tion on instrumented freewaysrdquo in Proceedings of the 6thWorld Congress on Intelligent Transportation Systems TorontoCanada 1999

[13] J Kwon B Coifman and P Bickel ldquoDay-to-day travel-timetrends and travel-time prediction from loop-detector datardquoTransportation Research Record no 1717 pp 120ndash129 2000

[14] J Rice and E Van Zwet ldquoA simple and effective method forpredicting travel times on freewaysrdquo IEEE Transactions onIntelligent Transportation Systems vol 5 no 3 pp 200ndash2072004

[15] C D R Lindveld RThijs P H L Bovy andN J van der ZijppldquoEvaluation of online travel time estimators and predictorsrdquoTransportation Research Record vol 1719 pp 45ndash53 2000

[16] H Yang S Chen YWang BWu andZWang ldquoComparison ofdelay estimation models for signalised intersections using fieldobservations in Shanghairdquo IET Intelligent Transport Systems vol10 no 3 pp 165ndash174 2016

[17] C L Mak and H S L Fan ldquoDevelopment of dual-stationautomated expressway incident detection algorithmsrdquo IEEETransactions on Intelligent Transportation Systems vol 8 no 3pp 480ndash490 2007

[18] K Hi-Ri-O-Tappa C Likitkhajorn A Poolsawat and S Tha-jchayapong ldquoTraffic incident detection system using series ofpoint detectorsrdquo in Proceedings of the 15th International IEEEConference on Intelligent Transportation Systems (ITSC rsquo12) pp182ndash187 Anchorage Alaska USA September 2012

[19] V Alexiadis J Colyar J Halkias R Hranac and G McHaleldquoThe next generation simulation programrdquo ITE Journal vol 74no 8 pp 22ndash26 2004

[20] F G Habtemichael M Cetin and K A Anuar ldquoIncident-induced delays on freeways quantificationmethod by groupingsimilar traffic patternsrdquo Transportation Research Record vol2484 pp 60ndash69 2015

[21] J Li C-J Lan and X Gu ldquoEstimation of incident delay andits uncertainty on freeway networksrdquo Transportation ResearchRecord vol 1959 pp 37ndash45 2006

[22] Y Chung ldquoAssessment of non-recurrent congestion causedby precipitation using archived weather and traffic flow datardquoTransport Policy vol 19 no 1 pp 167ndash173 2012

[23] J A Barria and SThajchayapong ldquoDetection and classificationof traffic anomalies using microscopic traffic variablesrdquo IEEETransactions on Intelligent Transportation Systems vol 12 no3 pp 695ndash704 2011

[24] K HiriOtappa and S Thajchayapong ldquoGeneralizability andtransferability of incident detection algorithm using dynamictime warpingrdquo in Proceedings of the in 19th World Congress onIntelligent Transportation Systems Vienna Austria 2012

Journal of Advanced Transportation 13

[25] S Thajchayapong and J A Barria ldquoSpatial inference of traffictransition using microndashmacro traffic variablesrdquo IEEE Transac-tions on Intelligent Transportation Systems vol 16 no 2 pp854ndash864 2015

[26] C A Ratanamahatana and E Keogh ldquoEverything you knowabout dynamic time warping is wrongrdquo in Proceedings of the3rd Workshop on Mining Temporal and Sequential Data in Con-junction with the 10th ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining (KDD rsquo04) CiteseerSeattle Wash USA August 2004

[27] Transportation Research Thesauras Transportation ResearchBoard National Academy httptrttrborgtrtaspNN=Bthd

[28] N Chiabaut C Buisson and L Leclercq ldquoFundamental dia-gram estimation through passing rate measurements in conges-tionrdquo IEEE Transactions on Intelligent Transportation Systemsvol 10 no 2 pp 355ndash359 2009

[29] J J Fernandez-Lozano M Martın-Guzman J Martın-Avilaand A Garcıa-Cerezo ldquoA wireless sensor network for urbantraffic characterization and trend monitoringrdquo Sensors vol 15no 10 pp 26143ndash26169 2015

[30] J-S Oh C Oh S G Ritchie and M Chang ldquoReal-timeestimation of accident likelihood for safety enhancementrdquoJournal of Transportation Engineering vol 131 no 5 pp 358ndash363 2005

[31] A P Chassiakos and Y J Stephanedes ldquoSmoothing algorithmsfor incident detectionrdquo Transportation Research Record vol1394 pp 8ndash16 1993

[32] V Alarcon-Aquino and J A Barria ldquoAnomaly detection incommunication networks using waveletsrdquo IEE ProceedingsmdashCommunications vol 148 no 6 pp 355ndash362 2001

[33] D Berndt and J Clifford ldquoUsing dynamic time warping tofind patterns in time seriesrdquo in Proceedings of the Workshop onKnowledge Discovery in Databases (KDD rsquo94) vol 10 pp 359ndash370 Seattle Wash USA August 1994

[34] K Kiratiratanapruk and S Siddhichai ldquoVehicle detection andtracking for traffic monitoring systemrdquo in Proceedings of theIEEE Region 10 Conference (TENCON rsquo06) pp 1ndash4 IEEE HongKong November 2006

[35] A Hegyi B De Schutter and J Hellendoorn ldquoOptimal coordi-nation of variable speed limits to suppress shock wavesrdquo IEEETransactions on Intelligent Transportation Systems vol 6 no 1pp 102ndash112 2005

[36] Y Lv Y Duan W Kang Z Li and F-Y Wang ldquoTraffic flowprediction with big data a deep learning approachrdquo IEEETransactions on Intelligent Transportation Systems vol 16 no2 pp 865ndash873 2015

[37] P Lopez-Garcia E Onieva E Osaba A D Masegosa and APerallos ldquoA hybrid method for short-term traffic congestionforecasting using genetic algorithms and cross entropyrdquo IEEETransactions on Intelligent Transportation Systems vol 17 no 2pp 557ndash569 2016

Journal of Advanced Transportation 9

Upstream sensorDownstream sensor

805 820 835750Time (hoursminutes)

01020304050607080

Spee

d (k

mh

)

(a)

One time step aheadTwo time steps ahead

805 820 835750Time (hoursminutes)

minus3minus2minus1

0123

Spee

d di

ffere

nces

(km

h)

(b)

Figure 7 Detecting delay occurrences associated with shockwaves in NGSIM data using (a) the proposed algorithm and (b) benchmarkalgorithm [17] Dashed vertical lines denote the time points where the algorithms detect delay occurrences Red squares denote the durationwhere the occurrences of shockwaves are observed

with both abrupt and gradual speed changes as shown inFigure 7(a)

Once the occurrences of shockwaves are detected theproposed algorithm proceeds to estimate delay as describedin Section 44 Using the spatial extent estimation describedin Section 44 out of four sensors three sensors are usuallyfound to detect speed drops associated with shockwaves at atime Therefore the most downstream sensor and the mostupstream sensor are selected as the start and end sensorsrespectively and 119897119899119875 is found to be 048 km

Table 1 shows the values of V119899119875 V119899119873 120591119899119875 and 120591119899119873 in (1)for each of the four shockwave occurrences For comparisonpurposes delay estimation is also applied for the second andthird occurrences that are detected by benchmark algorithm(see Figure 7) Note that since benchmark algorithm cannotdetect the first and fourth occurrences estimation operationis not activated during these times and normal speed andtravel times are used instead We further note that all fouroccurrences and their associated travel times and delays canbe calculated consecutively online by the proposed algorithm

As shown in Table 1 the four occurrences of shockwavesdetected in Figure 7 are associated with travel times (120591119899119875 in(1)) of 5682 s 5554 s 6098 s and 5379 s respectively Theproposed algorithm is able to detect all four occurrences andthese estimated travel times would all be included in the delaycalculation On the other hand if the benchmark algorithmwas used instead only the second and third occurrences andtheir associated travel times would be reported resulting inless accurate delay calculationWe note that even though thedifference between the delays obtained from the proposedalgorithm and benchmark algorithm in this example maybe small significant improvement can be seen when higherdegree of disruptions are assessed

The results shown in Table 1 have many applicationsin advanced traveler information systems and advancedtraffic management systems A direct application would beto estimate travel time of the freeway segment and report itto motorists The ability to detect occurrences of shockwavescan also be used in other applications such as calculatingshockwave propagation time [25] It can also be used to warn

motorists to adjust intervehicle spacing to avoid rear-endcollisions

62 Assessment Using Real-World Data Collected in ThailandThis section assesses the capability of the proposed algorithmand the benchmark algorithm [17] to detect the sixteenrecorded traffic anomalies in the real-world data set describedin Section 51 As inputs both algorithms use speed flowand occupancy collected by roadside sensorsThe assessmentfollows a 119896-fold cross-validation technique The real-worlddata set consists of 16 anomaly cases 1198631 1198632 11986316 soevaluation is performed 16 times At an iteration 119894 119894 =1 2 16 the anomaly case119863119894 is reserved as a test set whilethe other anomaly cases are collectively used to calibratethe proposed algorithm and benchmark algorithm That isin the first iteration 1198632 11986316 collectively serve as thetraining set which is then tested with1198631Then in the seconditeration 1198631 1198633 11986316 are used for training while 1198632 isused for testing and so on

Table 2 shows that the proposed algorithm achieveshigher detection rate and lower false alarm rate than thebenchmark algorithm Particularly compared to benchmarkalgorithm the proposed algorithm improves the detectionrate and false positive rate by approximately 10 and 7respectively Figures 8 and 9 show examples of anomaliesdetected by the proposed algorithm and found to causetraffic delay Based on the records these anomalies areassociated with disabled vehicles parking over the shoulderarea

It is found that for the benchmark algorithm miss-detection usually occurs under low traffic flow This algo-rithm always needs data from at least two sensors and themagnitude of traffic flow measured by the roadside sensorsis not large enough to trigger the benchmark algorithmWhile medium to heavy traffic flow triggers the benchmarkalgorithm it also increases false alarms This algorithm isdesigned to directly assess the difference between upstreamand downstream sensors of the roadside sensors whichsubsequently increases its sensitivity and triggers many falsealarms

10 Journal of Advanced Transportation

Table 1 Estimation results onNGSIMdata using the proposed algorithm and benchmark algorithm 119897119899119875 =048 kmNote that if the shockwaveis not detected the estimated travel time 120591119899119875 would be equal to normal travel time 120591119899119873

1st 2nd 3rd 4thProposed algorithmShockwaves detected Yes Yes Yes YesV119899119875 (see (1)) 3041 kmh 3112 kmh 2834 kmh 3213 kmh120591119899119875 (see (1)) 5682 s 5554 s 6098 s 5379 s119889119899 = 120591119899119875 minus 120591119899119873 1841 s 1893 s 2399 s 541 sBenchmark algorithm [17]Shockwaves detected No Yes Yes NoV119899119875 (see (1)) 4499 kmh 3112 kmh 2834 kmh 3572 kmh120591119899119875 (see (1)) 3841 s 5554 s 6098 s 4838 s119889119899 = 120591119899119875 minus 120591119899119873 000 s 1893 s 2399 s 000 sIf all 4 shockwaves are not detectedV119899119873 (see (1)) 4499 kmh 4721 kmh 4671 kmh 3572 kmh120591119899119873 (see (1)) 3841 s 3660 s 3700 s 4838 s

Table 2 Results on anomaly detection using data set collected inThailand

Performanceparameters

Proposedalgorithm

Benchmarkalgorithm [17]

DR () 94 83FPR () 5 12MTTD (seconds) 340 140

Figure 8 An anomaly detected at the 4th camera sensor

Unlike the benchmark algorithm the proposed algorithmuses DTW to assess traffic data from each sensor individuallyThis enables the proposed algorithm to recognize patternsassociated with anomalies more effectively including underlow traffic flow The use of DTW also reduces a numberof false alarms as it ensures that unrecognized patternswill not trigger detection at individual sensors In respectto mean time to detection (MTTD) it takes longer forthe proposed algorithm to detect anomalies compared tobenchmark algorithm as it needs to wait till enough temporalsamples of traffic data have been collected in each sensorbefore the detection operation can be triggered Howeverthe mean time to detection is still less than 6 minutes which

Figure 9 Another anomaly detected at the 4th camera sensor

should give enough time for traffic operators to response [8]To reduce mean time to detection of the proposed algorithmshorter time samples of traffic data can be used ReducingMTTD while maintaining acceptable levels of DR and FPRis worth further investigation

To deploy the proposed algorithm computation timeof the algorithm could be an added delay to the MTTDEven though the computation time was not systematicallymeasured for the performance evaluation the computationtime was in order of seconds For the computerrsquos specificationthat we use which was a personal laptop with 25 GHz CPUand 4GB of RAM the computation time was approximately[1 3) seconds As the computing time (a few seconds) wasmuch smaller than the mean time to detection (up to 6minutes) it is reasonable to consider the computing timenegligible

This section further discusses the results when the pro-posed algorithm was applied to estimate 60 delay casesdescribed in Section 51 using different input time windowsizes (119871) and prediction horizons (119875) For practical purposesit is important to select suitable input time window sizeand prediction horizon to enable the proposed method to

Journal of Advanced Transportation 11

L = 3 L = 10L = 5

P = 5

P = 10

P = 15

0

005

01

015

02

025

03

035

04M

APE

Figure 10Mean absolute percentage error (MAPE) fromestimating60 delay cases using different input time window sizes 119871 (minutes)and prediction horizons 119875 (minutes)

perform effectively Even though the 15-minute interval isoften recommended [9] to avoid strong fluctuation usinglower input timewindow sizes can providemore fine-grainedinformation for practitioners [25] Therefore in this paperwe use 119871 = 3 5 and 10 minutes for input time windowsize For prediction horizons we use 119875 = 5 10 and 15minutes as this paper is focused on short-term estimation(less than 15 minutes) Also changes in traffic condition onthis experimental site are normally found to occur in less than15 minutes

The proposed algorithm is assessed on its capability toestimate delay using MAPE Figure 10 shows MAPE fromestimating 60 delay cases in the real-world data set withdifferent input timewindow sizes (119871) and prediction horizons(119875) These results show that the estimation part of theproposed algorithm can achieve MAPE lower than 01 forthis set of real-world data This accuracy is achieved becausethe estimation part consists of both spatial and temporalestimations which enable the proposed algorithm to firstestimate the area disrupted by anomalies and then estimatethe delay itself Furthermore it can be seen that increasinginput time window size (119871) reduces MAPE as using moreinput samples enhances estimation accuracy For the impactof prediction horizons (119875) MAPE tends to increase withlarger prediction horizons as it is more difficult to estimatedelays farther ahead into the future However as shownin Figure 10 when 119875 is increased MAPE in fact increasesat a much smaller rate when larger values of 119871 are usedTherefore it is recommended for practitioners that if theobjective is to estimate delays far ahead into the future (usinglarge prediction horizons) larger input time window size isneeded

Based on the findings in this section it is importantthat practitioners select a suitable input time window sizeaccording to their practical objectives If the objective is toprovide prompt delay information small input time windowsize should be used On the other hand if the objective is toprovide accurately estimated delays for traffic managementpurposes larger input time window sizes should be chosenFurthermore based on the proposed algorithm a morecomplex and flexible system can also be further developedFor example a stepwise approach can be implemented on topof the proposed algorithm where a small input time windowsize is used as a trigger to promptly discover the occurrenceof traffic delay Then in the next step the input time windowsize can be increased to collect more traffic data to achievemore accurate detection and estimation of traffic delay

For anomalies with large scale effect beyond the 11 kmexperimental site we anticipate that the proposed algo-rithm should be able to detect those anomalies if theyoccur downstream and cause changes in traffic patterns topropagate upstream into the 11 km segment To accuratelyestimate delay basis patterns associated with such large scaleanomalies need to be incorporated to distinguish themThenan extrapolation technique needs to be incorporated to enablethe proposed algorithm to estimate delay beyond the 11 kmexpressway segment where local traffic information cannotbe directly observed [25] This is worth further investigation

7 Conclusions and Future Work

This paper proposes an algorithm to detect anomalies andestimate delay using traffic data from roadside sensors Theproposed algorithm uses traffic data from individual sensorsto detect anomalies and then uses data from consecutivedownstream and upstream sensors together to spatially andtemporally estimate delay Performance evaluations usingreal-world data show that the proposed algorithm achieveshigher detection rate and lower false positive rate when usedto detect anomalies compared to an existing benchmarkalgorithmAn investigationwithNGSIMdata also shows thatthe proposed algorithm is also able to detect the presence ofshockwaves and subsequently estimate delaymore efficientlythan benchmark algorithm Further results show that theestimation error increases at a much smaller rate whensuitable input time window size and prediction horizon areselected

There are interesting aspects that are worth furtherinvestigation One interesting aspect is to reduce the pro-cessing time while maintaining low detection and estima-tion errors A possible approach is to introduce a featureextraction method that can extract hidden patterns whensmaller input time window sizes are used [8] Alternativeapproaches include studying a set of hidden representativetraffic flow features in advance by pretraining a deep-learningmodel [36] and populating traffic features into a geneticalgorithm learning method to reduce dimensionality [37]Another interesting aspect is to improve inference capabilityby incorporating a spatial inferencemethod to estimate trafficdata itself at the locations where they cannot be directlymeasured [25] Also it would be practically interesting to

12 Journal of Advanced Transportation

assess the proposed algorithm with different road types andgeometries

In addition to improvements of the algorithm an interest-ing aspect to investigate is the applications of the algorithmThe solution provided in this paper is designed to be astreaming algorithm which can be run online thereforeit can be naturally extended to use in various real-timeapplications For example the detected anomaly and esti-mated additional travel time would be useful for both trafficoperators and travelers if the information is delivered ina timely and convenient manner Active communicationbetween roadside sensors and vehicle-to-infrastructure (V2I)could be used to promptly transmit the information from oursolution to nearby roadside message boards to display theinformation to the drivers on the roads or vehicles with V2Icommunication to adapt routing and navigation

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this article

Acknowledgments

The authors would like to thankDr Supakorn Siddhichai andhis IMG team at NECTEC for their assistance in providingreal-world data for the analysis

References

[1] N Parrado and Y Donoso ldquoCongestion based mechanism forroute discovery in a V2I-V2V system applying smart devicesand IoTrdquo Sensors vol 15 no 4 pp 7768ndash7806 2015

[2] K Nellore and G P Hancke ldquoA survey on urban trafficmanagement system using wireless sensor networksrdquo Sensorsvol 16 no 2 article 157 2016

[3] E I Vlahogianni M G Karlaftis and J C Golias ldquoShort-term traffic forecasting where we are and where wersquore goingrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 3ndash19 2014

[4] R Li and G Rose ldquoIncorporating uncertainty into short-term travel time predictionsrdquo Transportation Research Part CEmerging Technologies vol 19 no 6 pp 1006ndash1018 2011

[5] Y Chung ldquoAmethodological approach for estimating temporaland spatial extent of delays caused by freeway accidentsrdquo IEEETransactions on Intelligent Transportation Systems vol 13 no 3pp 1454ndash1461 2012

[6] J Weng and Q Meng ldquoEstimating capacity and traffic delayin work zones an overviewrdquo Transportation ResearchmdashPart CEmerging Technologies vol 35 pp 34ndash45 2013

[7] X Zhang E Onieva A Perallos E Osaba and V C SLee ldquoHierarchical fuzzy rule-based system optimized withgenetic algorithms for short term traffic congestion predictionrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 127ndash142 2014

[8] S Thajchayapong E S Garcia-Trevino and J A Barria ldquoDis-tributed classification of traffic anomalies using microscopictraffic variablesrdquo IEEETransactions on Intelligent TransportationSystems vol 14 no 1 pp 448ndash458 2013

[9] M Lippi M Bertini and P Frasconi ldquoShort-term traffic flowforecasting an experimental comparison of time-series anal-ysis and supervised learningrdquo IEEE Transactions on IntelligentTransportation Systems vol 14 no 2 pp 871ndash882 2013

[10] F Zheng andH Van Zuylen ldquoUrban link travel time estimationbased on sparse probe vehicle datardquo Transportation ResearchmdashPart C Emerging Technologies vol 31 pp 145ndash157 2013

[11] F Dion and H Rakha ldquoEstimating dynamic roadway traveltimes using automatic vehicle identification data for low sam-pling ratesrdquoTransportation Research Part BMethodological vol40 no 9 pp 745ndash766 2006

[12] D Paterson and G Rose ldquoDynamic travel time estima-tion on instrumented freewaysrdquo in Proceedings of the 6thWorld Congress on Intelligent Transportation Systems TorontoCanada 1999

[13] J Kwon B Coifman and P Bickel ldquoDay-to-day travel-timetrends and travel-time prediction from loop-detector datardquoTransportation Research Record no 1717 pp 120ndash129 2000

[14] J Rice and E Van Zwet ldquoA simple and effective method forpredicting travel times on freewaysrdquo IEEE Transactions onIntelligent Transportation Systems vol 5 no 3 pp 200ndash2072004

[15] C D R Lindveld RThijs P H L Bovy andN J van der ZijppldquoEvaluation of online travel time estimators and predictorsrdquoTransportation Research Record vol 1719 pp 45ndash53 2000

[16] H Yang S Chen YWang BWu andZWang ldquoComparison ofdelay estimation models for signalised intersections using fieldobservations in Shanghairdquo IET Intelligent Transport Systems vol10 no 3 pp 165ndash174 2016

[17] C L Mak and H S L Fan ldquoDevelopment of dual-stationautomated expressway incident detection algorithmsrdquo IEEETransactions on Intelligent Transportation Systems vol 8 no 3pp 480ndash490 2007

[18] K Hi-Ri-O-Tappa C Likitkhajorn A Poolsawat and S Tha-jchayapong ldquoTraffic incident detection system using series ofpoint detectorsrdquo in Proceedings of the 15th International IEEEConference on Intelligent Transportation Systems (ITSC rsquo12) pp182ndash187 Anchorage Alaska USA September 2012

[19] V Alexiadis J Colyar J Halkias R Hranac and G McHaleldquoThe next generation simulation programrdquo ITE Journal vol 74no 8 pp 22ndash26 2004

[20] F G Habtemichael M Cetin and K A Anuar ldquoIncident-induced delays on freeways quantificationmethod by groupingsimilar traffic patternsrdquo Transportation Research Record vol2484 pp 60ndash69 2015

[21] J Li C-J Lan and X Gu ldquoEstimation of incident delay andits uncertainty on freeway networksrdquo Transportation ResearchRecord vol 1959 pp 37ndash45 2006

[22] Y Chung ldquoAssessment of non-recurrent congestion causedby precipitation using archived weather and traffic flow datardquoTransport Policy vol 19 no 1 pp 167ndash173 2012

[23] J A Barria and SThajchayapong ldquoDetection and classificationof traffic anomalies using microscopic traffic variablesrdquo IEEETransactions on Intelligent Transportation Systems vol 12 no3 pp 695ndash704 2011

[24] K HiriOtappa and S Thajchayapong ldquoGeneralizability andtransferability of incident detection algorithm using dynamictime warpingrdquo in Proceedings of the in 19th World Congress onIntelligent Transportation Systems Vienna Austria 2012

Journal of Advanced Transportation 13

[25] S Thajchayapong and J A Barria ldquoSpatial inference of traffictransition using microndashmacro traffic variablesrdquo IEEE Transac-tions on Intelligent Transportation Systems vol 16 no 2 pp854ndash864 2015

[26] C A Ratanamahatana and E Keogh ldquoEverything you knowabout dynamic time warping is wrongrdquo in Proceedings of the3rd Workshop on Mining Temporal and Sequential Data in Con-junction with the 10th ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining (KDD rsquo04) CiteseerSeattle Wash USA August 2004

[27] Transportation Research Thesauras Transportation ResearchBoard National Academy httptrttrborgtrtaspNN=Bthd

[28] N Chiabaut C Buisson and L Leclercq ldquoFundamental dia-gram estimation through passing rate measurements in conges-tionrdquo IEEE Transactions on Intelligent Transportation Systemsvol 10 no 2 pp 355ndash359 2009

[29] J J Fernandez-Lozano M Martın-Guzman J Martın-Avilaand A Garcıa-Cerezo ldquoA wireless sensor network for urbantraffic characterization and trend monitoringrdquo Sensors vol 15no 10 pp 26143ndash26169 2015

[30] J-S Oh C Oh S G Ritchie and M Chang ldquoReal-timeestimation of accident likelihood for safety enhancementrdquoJournal of Transportation Engineering vol 131 no 5 pp 358ndash363 2005

[31] A P Chassiakos and Y J Stephanedes ldquoSmoothing algorithmsfor incident detectionrdquo Transportation Research Record vol1394 pp 8ndash16 1993

[32] V Alarcon-Aquino and J A Barria ldquoAnomaly detection incommunication networks using waveletsrdquo IEE ProceedingsmdashCommunications vol 148 no 6 pp 355ndash362 2001

[33] D Berndt and J Clifford ldquoUsing dynamic time warping tofind patterns in time seriesrdquo in Proceedings of the Workshop onKnowledge Discovery in Databases (KDD rsquo94) vol 10 pp 359ndash370 Seattle Wash USA August 1994

[34] K Kiratiratanapruk and S Siddhichai ldquoVehicle detection andtracking for traffic monitoring systemrdquo in Proceedings of theIEEE Region 10 Conference (TENCON rsquo06) pp 1ndash4 IEEE HongKong November 2006

[35] A Hegyi B De Schutter and J Hellendoorn ldquoOptimal coordi-nation of variable speed limits to suppress shock wavesrdquo IEEETransactions on Intelligent Transportation Systems vol 6 no 1pp 102ndash112 2005

[36] Y Lv Y Duan W Kang Z Li and F-Y Wang ldquoTraffic flowprediction with big data a deep learning approachrdquo IEEETransactions on Intelligent Transportation Systems vol 16 no2 pp 865ndash873 2015

[37] P Lopez-Garcia E Onieva E Osaba A D Masegosa and APerallos ldquoA hybrid method for short-term traffic congestionforecasting using genetic algorithms and cross entropyrdquo IEEETransactions on Intelligent Transportation Systems vol 17 no 2pp 557ndash569 2016

10 Journal of Advanced Transportation

Table 1 Estimation results onNGSIMdata using the proposed algorithm and benchmark algorithm 119897119899119875 =048 kmNote that if the shockwaveis not detected the estimated travel time 120591119899119875 would be equal to normal travel time 120591119899119873

1st 2nd 3rd 4thProposed algorithmShockwaves detected Yes Yes Yes YesV119899119875 (see (1)) 3041 kmh 3112 kmh 2834 kmh 3213 kmh120591119899119875 (see (1)) 5682 s 5554 s 6098 s 5379 s119889119899 = 120591119899119875 minus 120591119899119873 1841 s 1893 s 2399 s 541 sBenchmark algorithm [17]Shockwaves detected No Yes Yes NoV119899119875 (see (1)) 4499 kmh 3112 kmh 2834 kmh 3572 kmh120591119899119875 (see (1)) 3841 s 5554 s 6098 s 4838 s119889119899 = 120591119899119875 minus 120591119899119873 000 s 1893 s 2399 s 000 sIf all 4 shockwaves are not detectedV119899119873 (see (1)) 4499 kmh 4721 kmh 4671 kmh 3572 kmh120591119899119873 (see (1)) 3841 s 3660 s 3700 s 4838 s

Table 2 Results on anomaly detection using data set collected inThailand

Performanceparameters

Proposedalgorithm

Benchmarkalgorithm [17]

DR () 94 83FPR () 5 12MTTD (seconds) 340 140

Figure 8 An anomaly detected at the 4th camera sensor

Unlike the benchmark algorithm the proposed algorithmuses DTW to assess traffic data from each sensor individuallyThis enables the proposed algorithm to recognize patternsassociated with anomalies more effectively including underlow traffic flow The use of DTW also reduces a numberof false alarms as it ensures that unrecognized patternswill not trigger detection at individual sensors In respectto mean time to detection (MTTD) it takes longer forthe proposed algorithm to detect anomalies compared tobenchmark algorithm as it needs to wait till enough temporalsamples of traffic data have been collected in each sensorbefore the detection operation can be triggered Howeverthe mean time to detection is still less than 6 minutes which

Figure 9 Another anomaly detected at the 4th camera sensor

should give enough time for traffic operators to response [8]To reduce mean time to detection of the proposed algorithmshorter time samples of traffic data can be used ReducingMTTD while maintaining acceptable levels of DR and FPRis worth further investigation

To deploy the proposed algorithm computation timeof the algorithm could be an added delay to the MTTDEven though the computation time was not systematicallymeasured for the performance evaluation the computationtime was in order of seconds For the computerrsquos specificationthat we use which was a personal laptop with 25 GHz CPUand 4GB of RAM the computation time was approximately[1 3) seconds As the computing time (a few seconds) wasmuch smaller than the mean time to detection (up to 6minutes) it is reasonable to consider the computing timenegligible

This section further discusses the results when the pro-posed algorithm was applied to estimate 60 delay casesdescribed in Section 51 using different input time windowsizes (119871) and prediction horizons (119875) For practical purposesit is important to select suitable input time window sizeand prediction horizon to enable the proposed method to

Journal of Advanced Transportation 11

L = 3 L = 10L = 5

P = 5

P = 10

P = 15

0

005

01

015

02

025

03

035

04M

APE

Figure 10Mean absolute percentage error (MAPE) fromestimating60 delay cases using different input time window sizes 119871 (minutes)and prediction horizons 119875 (minutes)

perform effectively Even though the 15-minute interval isoften recommended [9] to avoid strong fluctuation usinglower input timewindow sizes can providemore fine-grainedinformation for practitioners [25] Therefore in this paperwe use 119871 = 3 5 and 10 minutes for input time windowsize For prediction horizons we use 119875 = 5 10 and 15minutes as this paper is focused on short-term estimation(less than 15 minutes) Also changes in traffic condition onthis experimental site are normally found to occur in less than15 minutes

The proposed algorithm is assessed on its capability toestimate delay using MAPE Figure 10 shows MAPE fromestimating 60 delay cases in the real-world data set withdifferent input timewindow sizes (119871) and prediction horizons(119875) These results show that the estimation part of theproposed algorithm can achieve MAPE lower than 01 forthis set of real-world data This accuracy is achieved becausethe estimation part consists of both spatial and temporalestimations which enable the proposed algorithm to firstestimate the area disrupted by anomalies and then estimatethe delay itself Furthermore it can be seen that increasinginput time window size (119871) reduces MAPE as using moreinput samples enhances estimation accuracy For the impactof prediction horizons (119875) MAPE tends to increase withlarger prediction horizons as it is more difficult to estimatedelays farther ahead into the future However as shownin Figure 10 when 119875 is increased MAPE in fact increasesat a much smaller rate when larger values of 119871 are usedTherefore it is recommended for practitioners that if theobjective is to estimate delays far ahead into the future (usinglarge prediction horizons) larger input time window size isneeded

Based on the findings in this section it is importantthat practitioners select a suitable input time window sizeaccording to their practical objectives If the objective is toprovide prompt delay information small input time windowsize should be used On the other hand if the objective is toprovide accurately estimated delays for traffic managementpurposes larger input time window sizes should be chosenFurthermore based on the proposed algorithm a morecomplex and flexible system can also be further developedFor example a stepwise approach can be implemented on topof the proposed algorithm where a small input time windowsize is used as a trigger to promptly discover the occurrenceof traffic delay Then in the next step the input time windowsize can be increased to collect more traffic data to achievemore accurate detection and estimation of traffic delay

For anomalies with large scale effect beyond the 11 kmexperimental site we anticipate that the proposed algo-rithm should be able to detect those anomalies if theyoccur downstream and cause changes in traffic patterns topropagate upstream into the 11 km segment To accuratelyestimate delay basis patterns associated with such large scaleanomalies need to be incorporated to distinguish themThenan extrapolation technique needs to be incorporated to enablethe proposed algorithm to estimate delay beyond the 11 kmexpressway segment where local traffic information cannotbe directly observed [25] This is worth further investigation

7 Conclusions and Future Work

This paper proposes an algorithm to detect anomalies andestimate delay using traffic data from roadside sensors Theproposed algorithm uses traffic data from individual sensorsto detect anomalies and then uses data from consecutivedownstream and upstream sensors together to spatially andtemporally estimate delay Performance evaluations usingreal-world data show that the proposed algorithm achieveshigher detection rate and lower false positive rate when usedto detect anomalies compared to an existing benchmarkalgorithmAn investigationwithNGSIMdata also shows thatthe proposed algorithm is also able to detect the presence ofshockwaves and subsequently estimate delaymore efficientlythan benchmark algorithm Further results show that theestimation error increases at a much smaller rate whensuitable input time window size and prediction horizon areselected

There are interesting aspects that are worth furtherinvestigation One interesting aspect is to reduce the pro-cessing time while maintaining low detection and estima-tion errors A possible approach is to introduce a featureextraction method that can extract hidden patterns whensmaller input time window sizes are used [8] Alternativeapproaches include studying a set of hidden representativetraffic flow features in advance by pretraining a deep-learningmodel [36] and populating traffic features into a geneticalgorithm learning method to reduce dimensionality [37]Another interesting aspect is to improve inference capabilityby incorporating a spatial inferencemethod to estimate trafficdata itself at the locations where they cannot be directlymeasured [25] Also it would be practically interesting to

12 Journal of Advanced Transportation

assess the proposed algorithm with different road types andgeometries

In addition to improvements of the algorithm an interest-ing aspect to investigate is the applications of the algorithmThe solution provided in this paper is designed to be astreaming algorithm which can be run online thereforeit can be naturally extended to use in various real-timeapplications For example the detected anomaly and esti-mated additional travel time would be useful for both trafficoperators and travelers if the information is delivered ina timely and convenient manner Active communicationbetween roadside sensors and vehicle-to-infrastructure (V2I)could be used to promptly transmit the information from oursolution to nearby roadside message boards to display theinformation to the drivers on the roads or vehicles with V2Icommunication to adapt routing and navigation

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this article

Acknowledgments

The authors would like to thankDr Supakorn Siddhichai andhis IMG team at NECTEC for their assistance in providingreal-world data for the analysis

References

[1] N Parrado and Y Donoso ldquoCongestion based mechanism forroute discovery in a V2I-V2V system applying smart devicesand IoTrdquo Sensors vol 15 no 4 pp 7768ndash7806 2015

[2] K Nellore and G P Hancke ldquoA survey on urban trafficmanagement system using wireless sensor networksrdquo Sensorsvol 16 no 2 article 157 2016

[3] E I Vlahogianni M G Karlaftis and J C Golias ldquoShort-term traffic forecasting where we are and where wersquore goingrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 3ndash19 2014

[4] R Li and G Rose ldquoIncorporating uncertainty into short-term travel time predictionsrdquo Transportation Research Part CEmerging Technologies vol 19 no 6 pp 1006ndash1018 2011

[5] Y Chung ldquoAmethodological approach for estimating temporaland spatial extent of delays caused by freeway accidentsrdquo IEEETransactions on Intelligent Transportation Systems vol 13 no 3pp 1454ndash1461 2012

[6] J Weng and Q Meng ldquoEstimating capacity and traffic delayin work zones an overviewrdquo Transportation ResearchmdashPart CEmerging Technologies vol 35 pp 34ndash45 2013

[7] X Zhang E Onieva A Perallos E Osaba and V C SLee ldquoHierarchical fuzzy rule-based system optimized withgenetic algorithms for short term traffic congestion predictionrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 127ndash142 2014

[8] S Thajchayapong E S Garcia-Trevino and J A Barria ldquoDis-tributed classification of traffic anomalies using microscopictraffic variablesrdquo IEEETransactions on Intelligent TransportationSystems vol 14 no 1 pp 448ndash458 2013

[9] M Lippi M Bertini and P Frasconi ldquoShort-term traffic flowforecasting an experimental comparison of time-series anal-ysis and supervised learningrdquo IEEE Transactions on IntelligentTransportation Systems vol 14 no 2 pp 871ndash882 2013

[10] F Zheng andH Van Zuylen ldquoUrban link travel time estimationbased on sparse probe vehicle datardquo Transportation ResearchmdashPart C Emerging Technologies vol 31 pp 145ndash157 2013

[11] F Dion and H Rakha ldquoEstimating dynamic roadway traveltimes using automatic vehicle identification data for low sam-pling ratesrdquoTransportation Research Part BMethodological vol40 no 9 pp 745ndash766 2006

[12] D Paterson and G Rose ldquoDynamic travel time estima-tion on instrumented freewaysrdquo in Proceedings of the 6thWorld Congress on Intelligent Transportation Systems TorontoCanada 1999

[13] J Kwon B Coifman and P Bickel ldquoDay-to-day travel-timetrends and travel-time prediction from loop-detector datardquoTransportation Research Record no 1717 pp 120ndash129 2000

[14] J Rice and E Van Zwet ldquoA simple and effective method forpredicting travel times on freewaysrdquo IEEE Transactions onIntelligent Transportation Systems vol 5 no 3 pp 200ndash2072004

[15] C D R Lindveld RThijs P H L Bovy andN J van der ZijppldquoEvaluation of online travel time estimators and predictorsrdquoTransportation Research Record vol 1719 pp 45ndash53 2000

[16] H Yang S Chen YWang BWu andZWang ldquoComparison ofdelay estimation models for signalised intersections using fieldobservations in Shanghairdquo IET Intelligent Transport Systems vol10 no 3 pp 165ndash174 2016

[17] C L Mak and H S L Fan ldquoDevelopment of dual-stationautomated expressway incident detection algorithmsrdquo IEEETransactions on Intelligent Transportation Systems vol 8 no 3pp 480ndash490 2007

[18] K Hi-Ri-O-Tappa C Likitkhajorn A Poolsawat and S Tha-jchayapong ldquoTraffic incident detection system using series ofpoint detectorsrdquo in Proceedings of the 15th International IEEEConference on Intelligent Transportation Systems (ITSC rsquo12) pp182ndash187 Anchorage Alaska USA September 2012

[19] V Alexiadis J Colyar J Halkias R Hranac and G McHaleldquoThe next generation simulation programrdquo ITE Journal vol 74no 8 pp 22ndash26 2004

[20] F G Habtemichael M Cetin and K A Anuar ldquoIncident-induced delays on freeways quantificationmethod by groupingsimilar traffic patternsrdquo Transportation Research Record vol2484 pp 60ndash69 2015

[21] J Li C-J Lan and X Gu ldquoEstimation of incident delay andits uncertainty on freeway networksrdquo Transportation ResearchRecord vol 1959 pp 37ndash45 2006

[22] Y Chung ldquoAssessment of non-recurrent congestion causedby precipitation using archived weather and traffic flow datardquoTransport Policy vol 19 no 1 pp 167ndash173 2012

[23] J A Barria and SThajchayapong ldquoDetection and classificationof traffic anomalies using microscopic traffic variablesrdquo IEEETransactions on Intelligent Transportation Systems vol 12 no3 pp 695ndash704 2011

[24] K HiriOtappa and S Thajchayapong ldquoGeneralizability andtransferability of incident detection algorithm using dynamictime warpingrdquo in Proceedings of the in 19th World Congress onIntelligent Transportation Systems Vienna Austria 2012

Journal of Advanced Transportation 13

[25] S Thajchayapong and J A Barria ldquoSpatial inference of traffictransition using microndashmacro traffic variablesrdquo IEEE Transac-tions on Intelligent Transportation Systems vol 16 no 2 pp854ndash864 2015

[26] C A Ratanamahatana and E Keogh ldquoEverything you knowabout dynamic time warping is wrongrdquo in Proceedings of the3rd Workshop on Mining Temporal and Sequential Data in Con-junction with the 10th ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining (KDD rsquo04) CiteseerSeattle Wash USA August 2004

[27] Transportation Research Thesauras Transportation ResearchBoard National Academy httptrttrborgtrtaspNN=Bthd

[28] N Chiabaut C Buisson and L Leclercq ldquoFundamental dia-gram estimation through passing rate measurements in conges-tionrdquo IEEE Transactions on Intelligent Transportation Systemsvol 10 no 2 pp 355ndash359 2009

[29] J J Fernandez-Lozano M Martın-Guzman J Martın-Avilaand A Garcıa-Cerezo ldquoA wireless sensor network for urbantraffic characterization and trend monitoringrdquo Sensors vol 15no 10 pp 26143ndash26169 2015

[30] J-S Oh C Oh S G Ritchie and M Chang ldquoReal-timeestimation of accident likelihood for safety enhancementrdquoJournal of Transportation Engineering vol 131 no 5 pp 358ndash363 2005

[31] A P Chassiakos and Y J Stephanedes ldquoSmoothing algorithmsfor incident detectionrdquo Transportation Research Record vol1394 pp 8ndash16 1993

[32] V Alarcon-Aquino and J A Barria ldquoAnomaly detection incommunication networks using waveletsrdquo IEE ProceedingsmdashCommunications vol 148 no 6 pp 355ndash362 2001

[33] D Berndt and J Clifford ldquoUsing dynamic time warping tofind patterns in time seriesrdquo in Proceedings of the Workshop onKnowledge Discovery in Databases (KDD rsquo94) vol 10 pp 359ndash370 Seattle Wash USA August 1994

[34] K Kiratiratanapruk and S Siddhichai ldquoVehicle detection andtracking for traffic monitoring systemrdquo in Proceedings of theIEEE Region 10 Conference (TENCON rsquo06) pp 1ndash4 IEEE HongKong November 2006

[35] A Hegyi B De Schutter and J Hellendoorn ldquoOptimal coordi-nation of variable speed limits to suppress shock wavesrdquo IEEETransactions on Intelligent Transportation Systems vol 6 no 1pp 102ndash112 2005

[36] Y Lv Y Duan W Kang Z Li and F-Y Wang ldquoTraffic flowprediction with big data a deep learning approachrdquo IEEETransactions on Intelligent Transportation Systems vol 16 no2 pp 865ndash873 2015

[37] P Lopez-Garcia E Onieva E Osaba A D Masegosa and APerallos ldquoA hybrid method for short-term traffic congestionforecasting using genetic algorithms and cross entropyrdquo IEEETransactions on Intelligent Transportation Systems vol 17 no 2pp 557ndash569 2016

Journal of Advanced Transportation 11

L = 3 L = 10L = 5

P = 5

P = 10

P = 15

0

005

01

015

02

025

03

035

04M

APE

Figure 10Mean absolute percentage error (MAPE) fromestimating60 delay cases using different input time window sizes 119871 (minutes)and prediction horizons 119875 (minutes)

perform effectively Even though the 15-minute interval isoften recommended [9] to avoid strong fluctuation usinglower input timewindow sizes can providemore fine-grainedinformation for practitioners [25] Therefore in this paperwe use 119871 = 3 5 and 10 minutes for input time windowsize For prediction horizons we use 119875 = 5 10 and 15minutes as this paper is focused on short-term estimation(less than 15 minutes) Also changes in traffic condition onthis experimental site are normally found to occur in less than15 minutes

The proposed algorithm is assessed on its capability toestimate delay using MAPE Figure 10 shows MAPE fromestimating 60 delay cases in the real-world data set withdifferent input timewindow sizes (119871) and prediction horizons(119875) These results show that the estimation part of theproposed algorithm can achieve MAPE lower than 01 forthis set of real-world data This accuracy is achieved becausethe estimation part consists of both spatial and temporalestimations which enable the proposed algorithm to firstestimate the area disrupted by anomalies and then estimatethe delay itself Furthermore it can be seen that increasinginput time window size (119871) reduces MAPE as using moreinput samples enhances estimation accuracy For the impactof prediction horizons (119875) MAPE tends to increase withlarger prediction horizons as it is more difficult to estimatedelays farther ahead into the future However as shownin Figure 10 when 119875 is increased MAPE in fact increasesat a much smaller rate when larger values of 119871 are usedTherefore it is recommended for practitioners that if theobjective is to estimate delays far ahead into the future (usinglarge prediction horizons) larger input time window size isneeded

Based on the findings in this section it is importantthat practitioners select a suitable input time window sizeaccording to their practical objectives If the objective is toprovide prompt delay information small input time windowsize should be used On the other hand if the objective is toprovide accurately estimated delays for traffic managementpurposes larger input time window sizes should be chosenFurthermore based on the proposed algorithm a morecomplex and flexible system can also be further developedFor example a stepwise approach can be implemented on topof the proposed algorithm where a small input time windowsize is used as a trigger to promptly discover the occurrenceof traffic delay Then in the next step the input time windowsize can be increased to collect more traffic data to achievemore accurate detection and estimation of traffic delay

For anomalies with large scale effect beyond the 11 kmexperimental site we anticipate that the proposed algo-rithm should be able to detect those anomalies if theyoccur downstream and cause changes in traffic patterns topropagate upstream into the 11 km segment To accuratelyestimate delay basis patterns associated with such large scaleanomalies need to be incorporated to distinguish themThenan extrapolation technique needs to be incorporated to enablethe proposed algorithm to estimate delay beyond the 11 kmexpressway segment where local traffic information cannotbe directly observed [25] This is worth further investigation

7 Conclusions and Future Work

This paper proposes an algorithm to detect anomalies andestimate delay using traffic data from roadside sensors Theproposed algorithm uses traffic data from individual sensorsto detect anomalies and then uses data from consecutivedownstream and upstream sensors together to spatially andtemporally estimate delay Performance evaluations usingreal-world data show that the proposed algorithm achieveshigher detection rate and lower false positive rate when usedto detect anomalies compared to an existing benchmarkalgorithmAn investigationwithNGSIMdata also shows thatthe proposed algorithm is also able to detect the presence ofshockwaves and subsequently estimate delaymore efficientlythan benchmark algorithm Further results show that theestimation error increases at a much smaller rate whensuitable input time window size and prediction horizon areselected

There are interesting aspects that are worth furtherinvestigation One interesting aspect is to reduce the pro-cessing time while maintaining low detection and estima-tion errors A possible approach is to introduce a featureextraction method that can extract hidden patterns whensmaller input time window sizes are used [8] Alternativeapproaches include studying a set of hidden representativetraffic flow features in advance by pretraining a deep-learningmodel [36] and populating traffic features into a geneticalgorithm learning method to reduce dimensionality [37]Another interesting aspect is to improve inference capabilityby incorporating a spatial inferencemethod to estimate trafficdata itself at the locations where they cannot be directlymeasured [25] Also it would be practically interesting to

12 Journal of Advanced Transportation

assess the proposed algorithm with different road types andgeometries

In addition to improvements of the algorithm an interest-ing aspect to investigate is the applications of the algorithmThe solution provided in this paper is designed to be astreaming algorithm which can be run online thereforeit can be naturally extended to use in various real-timeapplications For example the detected anomaly and esti-mated additional travel time would be useful for both trafficoperators and travelers if the information is delivered ina timely and convenient manner Active communicationbetween roadside sensors and vehicle-to-infrastructure (V2I)could be used to promptly transmit the information from oursolution to nearby roadside message boards to display theinformation to the drivers on the roads or vehicles with V2Icommunication to adapt routing and navigation

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this article

Acknowledgments

The authors would like to thankDr Supakorn Siddhichai andhis IMG team at NECTEC for their assistance in providingreal-world data for the analysis

References

[1] N Parrado and Y Donoso ldquoCongestion based mechanism forroute discovery in a V2I-V2V system applying smart devicesand IoTrdquo Sensors vol 15 no 4 pp 7768ndash7806 2015

[2] K Nellore and G P Hancke ldquoA survey on urban trafficmanagement system using wireless sensor networksrdquo Sensorsvol 16 no 2 article 157 2016

[3] E I Vlahogianni M G Karlaftis and J C Golias ldquoShort-term traffic forecasting where we are and where wersquore goingrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 3ndash19 2014

[4] R Li and G Rose ldquoIncorporating uncertainty into short-term travel time predictionsrdquo Transportation Research Part CEmerging Technologies vol 19 no 6 pp 1006ndash1018 2011

[5] Y Chung ldquoAmethodological approach for estimating temporaland spatial extent of delays caused by freeway accidentsrdquo IEEETransactions on Intelligent Transportation Systems vol 13 no 3pp 1454ndash1461 2012

[6] J Weng and Q Meng ldquoEstimating capacity and traffic delayin work zones an overviewrdquo Transportation ResearchmdashPart CEmerging Technologies vol 35 pp 34ndash45 2013

[7] X Zhang E Onieva A Perallos E Osaba and V C SLee ldquoHierarchical fuzzy rule-based system optimized withgenetic algorithms for short term traffic congestion predictionrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 127ndash142 2014

[8] S Thajchayapong E S Garcia-Trevino and J A Barria ldquoDis-tributed classification of traffic anomalies using microscopictraffic variablesrdquo IEEETransactions on Intelligent TransportationSystems vol 14 no 1 pp 448ndash458 2013

[9] M Lippi M Bertini and P Frasconi ldquoShort-term traffic flowforecasting an experimental comparison of time-series anal-ysis and supervised learningrdquo IEEE Transactions on IntelligentTransportation Systems vol 14 no 2 pp 871ndash882 2013

[10] F Zheng andH Van Zuylen ldquoUrban link travel time estimationbased on sparse probe vehicle datardquo Transportation ResearchmdashPart C Emerging Technologies vol 31 pp 145ndash157 2013

[11] F Dion and H Rakha ldquoEstimating dynamic roadway traveltimes using automatic vehicle identification data for low sam-pling ratesrdquoTransportation Research Part BMethodological vol40 no 9 pp 745ndash766 2006

[12] D Paterson and G Rose ldquoDynamic travel time estima-tion on instrumented freewaysrdquo in Proceedings of the 6thWorld Congress on Intelligent Transportation Systems TorontoCanada 1999

[13] J Kwon B Coifman and P Bickel ldquoDay-to-day travel-timetrends and travel-time prediction from loop-detector datardquoTransportation Research Record no 1717 pp 120ndash129 2000

[14] J Rice and E Van Zwet ldquoA simple and effective method forpredicting travel times on freewaysrdquo IEEE Transactions onIntelligent Transportation Systems vol 5 no 3 pp 200ndash2072004

[15] C D R Lindveld RThijs P H L Bovy andN J van der ZijppldquoEvaluation of online travel time estimators and predictorsrdquoTransportation Research Record vol 1719 pp 45ndash53 2000

[16] H Yang S Chen YWang BWu andZWang ldquoComparison ofdelay estimation models for signalised intersections using fieldobservations in Shanghairdquo IET Intelligent Transport Systems vol10 no 3 pp 165ndash174 2016

[17] C L Mak and H S L Fan ldquoDevelopment of dual-stationautomated expressway incident detection algorithmsrdquo IEEETransactions on Intelligent Transportation Systems vol 8 no 3pp 480ndash490 2007

[18] K Hi-Ri-O-Tappa C Likitkhajorn A Poolsawat and S Tha-jchayapong ldquoTraffic incident detection system using series ofpoint detectorsrdquo in Proceedings of the 15th International IEEEConference on Intelligent Transportation Systems (ITSC rsquo12) pp182ndash187 Anchorage Alaska USA September 2012

[19] V Alexiadis J Colyar J Halkias R Hranac and G McHaleldquoThe next generation simulation programrdquo ITE Journal vol 74no 8 pp 22ndash26 2004

[20] F G Habtemichael M Cetin and K A Anuar ldquoIncident-induced delays on freeways quantificationmethod by groupingsimilar traffic patternsrdquo Transportation Research Record vol2484 pp 60ndash69 2015

[21] J Li C-J Lan and X Gu ldquoEstimation of incident delay andits uncertainty on freeway networksrdquo Transportation ResearchRecord vol 1959 pp 37ndash45 2006

[22] Y Chung ldquoAssessment of non-recurrent congestion causedby precipitation using archived weather and traffic flow datardquoTransport Policy vol 19 no 1 pp 167ndash173 2012

[23] J A Barria and SThajchayapong ldquoDetection and classificationof traffic anomalies using microscopic traffic variablesrdquo IEEETransactions on Intelligent Transportation Systems vol 12 no3 pp 695ndash704 2011

[24] K HiriOtappa and S Thajchayapong ldquoGeneralizability andtransferability of incident detection algorithm using dynamictime warpingrdquo in Proceedings of the in 19th World Congress onIntelligent Transportation Systems Vienna Austria 2012

Journal of Advanced Transportation 13

[25] S Thajchayapong and J A Barria ldquoSpatial inference of traffictransition using microndashmacro traffic variablesrdquo IEEE Transac-tions on Intelligent Transportation Systems vol 16 no 2 pp854ndash864 2015

[26] C A Ratanamahatana and E Keogh ldquoEverything you knowabout dynamic time warping is wrongrdquo in Proceedings of the3rd Workshop on Mining Temporal and Sequential Data in Con-junction with the 10th ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining (KDD rsquo04) CiteseerSeattle Wash USA August 2004

[27] Transportation Research Thesauras Transportation ResearchBoard National Academy httptrttrborgtrtaspNN=Bthd

[28] N Chiabaut C Buisson and L Leclercq ldquoFundamental dia-gram estimation through passing rate measurements in conges-tionrdquo IEEE Transactions on Intelligent Transportation Systemsvol 10 no 2 pp 355ndash359 2009

[29] J J Fernandez-Lozano M Martın-Guzman J Martın-Avilaand A Garcıa-Cerezo ldquoA wireless sensor network for urbantraffic characterization and trend monitoringrdquo Sensors vol 15no 10 pp 26143ndash26169 2015

[30] J-S Oh C Oh S G Ritchie and M Chang ldquoReal-timeestimation of accident likelihood for safety enhancementrdquoJournal of Transportation Engineering vol 131 no 5 pp 358ndash363 2005

[31] A P Chassiakos and Y J Stephanedes ldquoSmoothing algorithmsfor incident detectionrdquo Transportation Research Record vol1394 pp 8ndash16 1993

[32] V Alarcon-Aquino and J A Barria ldquoAnomaly detection incommunication networks using waveletsrdquo IEE ProceedingsmdashCommunications vol 148 no 6 pp 355ndash362 2001

[33] D Berndt and J Clifford ldquoUsing dynamic time warping tofind patterns in time seriesrdquo in Proceedings of the Workshop onKnowledge Discovery in Databases (KDD rsquo94) vol 10 pp 359ndash370 Seattle Wash USA August 1994

[34] K Kiratiratanapruk and S Siddhichai ldquoVehicle detection andtracking for traffic monitoring systemrdquo in Proceedings of theIEEE Region 10 Conference (TENCON rsquo06) pp 1ndash4 IEEE HongKong November 2006

[35] A Hegyi B De Schutter and J Hellendoorn ldquoOptimal coordi-nation of variable speed limits to suppress shock wavesrdquo IEEETransactions on Intelligent Transportation Systems vol 6 no 1pp 102ndash112 2005

[36] Y Lv Y Duan W Kang Z Li and F-Y Wang ldquoTraffic flowprediction with big data a deep learning approachrdquo IEEETransactions on Intelligent Transportation Systems vol 16 no2 pp 865ndash873 2015

[37] P Lopez-Garcia E Onieva E Osaba A D Masegosa and APerallos ldquoA hybrid method for short-term traffic congestionforecasting using genetic algorithms and cross entropyrdquo IEEETransactions on Intelligent Transportation Systems vol 17 no 2pp 557ndash569 2016

12 Journal of Advanced Transportation

assess the proposed algorithm with different road types andgeometries

In addition to improvements of the algorithm an interest-ing aspect to investigate is the applications of the algorithmThe solution provided in this paper is designed to be astreaming algorithm which can be run online thereforeit can be naturally extended to use in various real-timeapplications For example the detected anomaly and esti-mated additional travel time would be useful for both trafficoperators and travelers if the information is delivered ina timely and convenient manner Active communicationbetween roadside sensors and vehicle-to-infrastructure (V2I)could be used to promptly transmit the information from oursolution to nearby roadside message boards to display theinformation to the drivers on the roads or vehicles with V2Icommunication to adapt routing and navigation

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this article

Acknowledgments

The authors would like to thankDr Supakorn Siddhichai andhis IMG team at NECTEC for their assistance in providingreal-world data for the analysis

References

[1] N Parrado and Y Donoso ldquoCongestion based mechanism forroute discovery in a V2I-V2V system applying smart devicesand IoTrdquo Sensors vol 15 no 4 pp 7768ndash7806 2015

[2] K Nellore and G P Hancke ldquoA survey on urban trafficmanagement system using wireless sensor networksrdquo Sensorsvol 16 no 2 article 157 2016

[3] E I Vlahogianni M G Karlaftis and J C Golias ldquoShort-term traffic forecasting where we are and where wersquore goingrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 3ndash19 2014

[4] R Li and G Rose ldquoIncorporating uncertainty into short-term travel time predictionsrdquo Transportation Research Part CEmerging Technologies vol 19 no 6 pp 1006ndash1018 2011

[5] Y Chung ldquoAmethodological approach for estimating temporaland spatial extent of delays caused by freeway accidentsrdquo IEEETransactions on Intelligent Transportation Systems vol 13 no 3pp 1454ndash1461 2012

[6] J Weng and Q Meng ldquoEstimating capacity and traffic delayin work zones an overviewrdquo Transportation ResearchmdashPart CEmerging Technologies vol 35 pp 34ndash45 2013

[7] X Zhang E Onieva A Perallos E Osaba and V C SLee ldquoHierarchical fuzzy rule-based system optimized withgenetic algorithms for short term traffic congestion predictionrdquoTransportation ResearchmdashPart C Emerging Technologies vol43 pp 127ndash142 2014

[8] S Thajchayapong E S Garcia-Trevino and J A Barria ldquoDis-tributed classification of traffic anomalies using microscopictraffic variablesrdquo IEEETransactions on Intelligent TransportationSystems vol 14 no 1 pp 448ndash458 2013

[9] M Lippi M Bertini and P Frasconi ldquoShort-term traffic flowforecasting an experimental comparison of time-series anal-ysis and supervised learningrdquo IEEE Transactions on IntelligentTransportation Systems vol 14 no 2 pp 871ndash882 2013

[10] F Zheng andH Van Zuylen ldquoUrban link travel time estimationbased on sparse probe vehicle datardquo Transportation ResearchmdashPart C Emerging Technologies vol 31 pp 145ndash157 2013

[11] F Dion and H Rakha ldquoEstimating dynamic roadway traveltimes using automatic vehicle identification data for low sam-pling ratesrdquoTransportation Research Part BMethodological vol40 no 9 pp 745ndash766 2006

[12] D Paterson and G Rose ldquoDynamic travel time estima-tion on instrumented freewaysrdquo in Proceedings of the 6thWorld Congress on Intelligent Transportation Systems TorontoCanada 1999

[13] J Kwon B Coifman and P Bickel ldquoDay-to-day travel-timetrends and travel-time prediction from loop-detector datardquoTransportation Research Record no 1717 pp 120ndash129 2000

[14] J Rice and E Van Zwet ldquoA simple and effective method forpredicting travel times on freewaysrdquo IEEE Transactions onIntelligent Transportation Systems vol 5 no 3 pp 200ndash2072004

[15] C D R Lindveld RThijs P H L Bovy andN J van der ZijppldquoEvaluation of online travel time estimators and predictorsrdquoTransportation Research Record vol 1719 pp 45ndash53 2000

[16] H Yang S Chen YWang BWu andZWang ldquoComparison ofdelay estimation models for signalised intersections using fieldobservations in Shanghairdquo IET Intelligent Transport Systems vol10 no 3 pp 165ndash174 2016

[17] C L Mak and H S L Fan ldquoDevelopment of dual-stationautomated expressway incident detection algorithmsrdquo IEEETransactions on Intelligent Transportation Systems vol 8 no 3pp 480ndash490 2007

[18] K Hi-Ri-O-Tappa C Likitkhajorn A Poolsawat and S Tha-jchayapong ldquoTraffic incident detection system using series ofpoint detectorsrdquo in Proceedings of the 15th International IEEEConference on Intelligent Transportation Systems (ITSC rsquo12) pp182ndash187 Anchorage Alaska USA September 2012

[19] V Alexiadis J Colyar J Halkias R Hranac and G McHaleldquoThe next generation simulation programrdquo ITE Journal vol 74no 8 pp 22ndash26 2004

[20] F G Habtemichael M Cetin and K A Anuar ldquoIncident-induced delays on freeways quantificationmethod by groupingsimilar traffic patternsrdquo Transportation Research Record vol2484 pp 60ndash69 2015

[21] J Li C-J Lan and X Gu ldquoEstimation of incident delay andits uncertainty on freeway networksrdquo Transportation ResearchRecord vol 1959 pp 37ndash45 2006

[22] Y Chung ldquoAssessment of non-recurrent congestion causedby precipitation using archived weather and traffic flow datardquoTransport Policy vol 19 no 1 pp 167ndash173 2012

[23] J A Barria and SThajchayapong ldquoDetection and classificationof traffic anomalies using microscopic traffic variablesrdquo IEEETransactions on Intelligent Transportation Systems vol 12 no3 pp 695ndash704 2011

[24] K HiriOtappa and S Thajchayapong ldquoGeneralizability andtransferability of incident detection algorithm using dynamictime warpingrdquo in Proceedings of the in 19th World Congress onIntelligent Transportation Systems Vienna Austria 2012

Journal of Advanced Transportation 13

[25] S Thajchayapong and J A Barria ldquoSpatial inference of traffictransition using microndashmacro traffic variablesrdquo IEEE Transac-tions on Intelligent Transportation Systems vol 16 no 2 pp854ndash864 2015

[26] C A Ratanamahatana and E Keogh ldquoEverything you knowabout dynamic time warping is wrongrdquo in Proceedings of the3rd Workshop on Mining Temporal and Sequential Data in Con-junction with the 10th ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining (KDD rsquo04) CiteseerSeattle Wash USA August 2004

[27] Transportation Research Thesauras Transportation ResearchBoard National Academy httptrttrborgtrtaspNN=Bthd

[28] N Chiabaut C Buisson and L Leclercq ldquoFundamental dia-gram estimation through passing rate measurements in conges-tionrdquo IEEE Transactions on Intelligent Transportation Systemsvol 10 no 2 pp 355ndash359 2009

[29] J J Fernandez-Lozano M Martın-Guzman J Martın-Avilaand A Garcıa-Cerezo ldquoA wireless sensor network for urbantraffic characterization and trend monitoringrdquo Sensors vol 15no 10 pp 26143ndash26169 2015

[30] J-S Oh C Oh S G Ritchie and M Chang ldquoReal-timeestimation of accident likelihood for safety enhancementrdquoJournal of Transportation Engineering vol 131 no 5 pp 358ndash363 2005

[31] A P Chassiakos and Y J Stephanedes ldquoSmoothing algorithmsfor incident detectionrdquo Transportation Research Record vol1394 pp 8ndash16 1993

[32] V Alarcon-Aquino and J A Barria ldquoAnomaly detection incommunication networks using waveletsrdquo IEE ProceedingsmdashCommunications vol 148 no 6 pp 355ndash362 2001

[33] D Berndt and J Clifford ldquoUsing dynamic time warping tofind patterns in time seriesrdquo in Proceedings of the Workshop onKnowledge Discovery in Databases (KDD rsquo94) vol 10 pp 359ndash370 Seattle Wash USA August 1994

[34] K Kiratiratanapruk and S Siddhichai ldquoVehicle detection andtracking for traffic monitoring systemrdquo in Proceedings of theIEEE Region 10 Conference (TENCON rsquo06) pp 1ndash4 IEEE HongKong November 2006

[35] A Hegyi B De Schutter and J Hellendoorn ldquoOptimal coordi-nation of variable speed limits to suppress shock wavesrdquo IEEETransactions on Intelligent Transportation Systems vol 6 no 1pp 102ndash112 2005

[36] Y Lv Y Duan W Kang Z Li and F-Y Wang ldquoTraffic flowprediction with big data a deep learning approachrdquo IEEETransactions on Intelligent Transportation Systems vol 16 no2 pp 865ndash873 2015

[37] P Lopez-Garcia E Onieva E Osaba A D Masegosa and APerallos ldquoA hybrid method for short-term traffic congestionforecasting using genetic algorithms and cross entropyrdquo IEEETransactions on Intelligent Transportation Systems vol 17 no 2pp 557ndash569 2016

Journal of Advanced Transportation 13

[25] S Thajchayapong and J A Barria ldquoSpatial inference of traffictransition using microndashmacro traffic variablesrdquo IEEE Transac-tions on Intelligent Transportation Systems vol 16 no 2 pp854ndash864 2015

[26] C A Ratanamahatana and E Keogh ldquoEverything you knowabout dynamic time warping is wrongrdquo in Proceedings of the3rd Workshop on Mining Temporal and Sequential Data in Con-junction with the 10th ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining (KDD rsquo04) CiteseerSeattle Wash USA August 2004

[27] Transportation Research Thesauras Transportation ResearchBoard National Academy httptrttrborgtrtaspNN=Bthd

[28] N Chiabaut C Buisson and L Leclercq ldquoFundamental dia-gram estimation through passing rate measurements in conges-tionrdquo IEEE Transactions on Intelligent Transportation Systemsvol 10 no 2 pp 355ndash359 2009

[29] J J Fernandez-Lozano M Martın-Guzman J Martın-Avilaand A Garcıa-Cerezo ldquoA wireless sensor network for urbantraffic characterization and trend monitoringrdquo Sensors vol 15no 10 pp 26143ndash26169 2015

[30] J-S Oh C Oh S G Ritchie and M Chang ldquoReal-timeestimation of accident likelihood for safety enhancementrdquoJournal of Transportation Engineering vol 131 no 5 pp 358ndash363 2005

[31] A P Chassiakos and Y J Stephanedes ldquoSmoothing algorithmsfor incident detectionrdquo Transportation Research Record vol1394 pp 8ndash16 1993

[32] V Alarcon-Aquino and J A Barria ldquoAnomaly detection incommunication networks using waveletsrdquo IEE ProceedingsmdashCommunications vol 148 no 6 pp 355ndash362 2001

[33] D Berndt and J Clifford ldquoUsing dynamic time warping tofind patterns in time seriesrdquo in Proceedings of the Workshop onKnowledge Discovery in Databases (KDD rsquo94) vol 10 pp 359ndash370 Seattle Wash USA August 1994

[34] K Kiratiratanapruk and S Siddhichai ldquoVehicle detection andtracking for traffic monitoring systemrdquo in Proceedings of theIEEE Region 10 Conference (TENCON rsquo06) pp 1ndash4 IEEE HongKong November 2006

[35] A Hegyi B De Schutter and J Hellendoorn ldquoOptimal coordi-nation of variable speed limits to suppress shock wavesrdquo IEEETransactions on Intelligent Transportation Systems vol 6 no 1pp 102ndash112 2005

[36] Y Lv Y Duan W Kang Z Li and F-Y Wang ldquoTraffic flowprediction with big data a deep learning approachrdquo IEEETransactions on Intelligent Transportation Systems vol 16 no2 pp 865ndash873 2015

[37] P Lopez-Garcia E Onieva E Osaba A D Masegosa and APerallos ldquoA hybrid method for short-term traffic congestionforecasting using genetic algorithms and cross entropyrdquo IEEETransactions on Intelligent Transportation Systems vol 17 no 2pp 557ndash569 2016

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of