automated collection of cyclist data using computer …docs.trb.org/prp/13-0745.pdf · 1 automated...

Automated Collection of Cyclist Data Using Computer Vision Techniques 1

Mohamed H. Zaki, PhD, (corresponding author) 2

Post-Doctoral Fellow, Department of Civil Engineering 3

University of British Columbia 4

6250 Applied Science Lane 5

Vancouver, BC, Canada V6T 1Z4 6

[email protected] 7

8

9

Tarek Sayed, PhD, P.Eng. 10

Distinguished University Scholar 11

Professor, Department of Civil Engineering 12




[email protected] 16

17 18

Andrew Cheung, 19

Research Assistant, Department of Civil Engineering 20




[email protected] 24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

Word count: 4768 Text + 5 Tables + 7 Figures = 7768 words 41

42

TRB 2013 Annual Meeting Paper revised from original submittal.

Abstract 1

2

One of the main challenges in conducting detailed analysis of cyclist behavior is the lack of reliable 3

data. Collecting data through manual methods is a labour-intensive and time consuming process. 4

Two of the important areas of cyclist data collection are volume counts and average speed 5

measurement. A volume count is important as it provides the basis for necessary exposure measures 6

and conveys essential information of traffic patterns. It can also serve as a performance measure of 7

the facility. Cyclist speed data is used for traffic control and safety studies. Video sensors, when 8

complemented with computer vision can offer a promising approach for the automated collection of 9

traffic data. The approach is characterized by the wealth of data they can capture, store and analyze. 10

Through the application of computer vision techniques, it is possible to obtain precise spatial and 11

temporal measurements of the road-users in a resource-efficient way. This paper demonstrates the use 12

of a set of computer vision techniques for the automated collection of cyclist data. The cyclist tracks 13

obtained from video analysis are used to perform screen line counting as well as cyclist speed 14

measurements. The applications are demonstrated using a real-world data set from a roundabout in 15

Vancouver, British Columbia. Further analysis was conducted on the mean speed of cyclists with 16

regards to several factors such as the travel path, helmet usage, and group size. The motivation of 17

this research is to improve the understanding of cyclists’ behavior and how it varies under different 18

conditions. Several conclusions can be drawn from the analysis of cyclist speed behaviour. Group 19

size, travel path, lane position and helmet usage were found affect the cyclist mean speed. Single 20

cyclists had a slightly, but significantly higher mean cycling speed compared to group cyclists. The 21

mean cycling speed was highest for the cyclists using the road rather than the sidewalk. The mean 22

cycling speed decreases for non-helmet users. 23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49


Motivation 1

2

As modern cities are changing to become greener and to provide a more liveable environment, there 3

is a pressing demand to shift towards active modes of transportation such as cycling. These active 4

modes of transportation can play a very important role in achieving efficient transportation systems 5

as they offer many benefits. Cycling is considered as one of the most environmentally sustainable 6

forms of transportation [1]. Its benefits include improving public health by increasing the amount of 7

physical activity [2], reduced congestion, energy conservation, and pollution reduction. However, 8

several challenges can inhibit a wider endorsement of cycling as a mainstay commuting mode. 9

These challenges relate to the mobility, accepted levels of safety and efficient resources [3]. The 10

fear of injury on the road, the weather affecting the road conditions and congestion of road traffic 11

are some of the common concerns for daily commuting cyclists. For instance, Winters et al. [4] 12

conducted a comprehensive study in Vancouver that investigated a total of 73 potential motivators 13

and deterrents which influence cyclists’ decisions. Overall, cycling was shown to provide more 14

benefits in the long term that outweigh its negative perceptions. Therefore, it is of great interest to 15

understand and study in greater details the elements that constitutes cyclist trip behavior in order to 16

develop effective policies and infrastructure. 17

18

19

One of the main challenges in conducting detailed analysis on cyclists’ behavior is the lack of 20

reliable data. Collecting reliable data is often labour-intensive and time consuming as it is usually 21

collected by manual counts or measurements [5] [6]. This lack of reliable data can have a 22

significant impact on several transportation engineering and planning aspects. Two of the important 23

areas of cyclist data collection are volume count and average speed measurement. Volume count is 24

important as it provides a basis for the necessary exposure measures [7] and conveys essential 25

information of the traffic patterns along the road segments of interest. It can also serve as a 26

performance measure of the facility [8] [9]. Cyclist speed data is used for traffic control and safety 27

studies. In recent years, some degree of automation has been implemented for cyclist data collection 28

( [10] [5] [11]). However, these approaches have geometric restrictions such as the need for 29

physical choke points or the feasibility of implementing overhead structures at cyclist walkways 30

[6]. As a result, most cyclist counts in open space such as unconstrained street networks are still 31

performed manually. Inconsistent and scarce cyclist counts make it difficult to justify capital for 32

implementing cyclist facilities. Given the aforementioned issues, there is a significant need to 33

explore methods of increasing the availability of cyclist data. One of these methods is to develop 34

reliable automated techniques for data collection such as the use of computer vision (CV) 35

techniques. Computer vision techniques are not new to the transportation field. For example, in 36

recent years, computer vision has been used to track vehicle and cyclist trajectories to study traffic 37

conflicts and their implications on traffic safety [12] [13]. 38

39

The paper will demonstrate the use of a set of computer vision techniques for the automated 40

collection of cyclist data. The cyclist tracks obtained are used to perform screen line based cyclist 41

counting as well as cyclist speed measurements. The applications are demonstrated using a real-42

world data set from a roundabout in Vancouver, British Columbia. Further analysis was conducted 43

on the mean cycling speed of cyclists with regards to several factors such as the travel path, helmet 44

usage, and group size. The goal of this research is to improve the understanding of cyclists’ 45

behavior in order to improve the riding condition and provide an efficient and safe commuting 46

environment. Examining the different attributes affecting the cycling performance provides valuable 47

information that aid in accommodating a wide range of potential users. Several benefits can be 48

withdrawn for transportation applications. Those include applications in cyclists’ simulation 49

models, planning and management of cyclists’ trips, evaluating facility level-of-service and cyclists 50

signals design [14] [9]. 51

52


Previous Studies 1

Cyclists Data Collection 2

3

Cyclists’ data collection and behavior analysis have been an active research topic in the last decade. 4

The focus has been on studying cyclists speed in different route configurations from signalized 5

intersection in mixed and dedicated bikes lanes to urban trails [15] [16] [17] [18]. Factors that 6

influence the speed such as the grade and the cyclist attributes have been identified [19] [20]. Table 7

1 summarizes the results of some recent studies on the subject. It is important to note that the 8

majority of the surveyed studies are based on manual speed measurements by field or video 9

observations. Attempts to use computer vision techniques are still primitive. The manual methods 10

currently used in practice for the collection of cyclists’ data lack the ability to capture microscopic 11

changes in position and speed. Automated video analysis is becoming more popular as it overcomes 12

the shortcomings present in manual methods that are commonly used. Table 1 covers only relevant 13

results in the past two decades, pointer to earlier studies can be found in [8] as well as the AASHTO 14

Guide [21] of development of bicycles facilities, which reported recommended speed design based 15

on cyclists expertise population. According to a study by the FHWA, a design speed of 32 km/h is 16

sufficient for most cyclists to travel at their desired speeds [22] [8]. 17

18

Cyclists Detection and Tracking 19

20

For automatic tracking of cyclists, moving road users must be detected and tracked frame-by-frame 21

and classified into cyclists and non-cyclists [23] [24]. Common problems in this challenging task 22

include global illumination variations, shadow handling, and multiple object tracking [25]. Road-23

user classification is undertaken during the video analysis, in order to aid the tracking of the moving 24

objects. CV based classification has found applications in traffic monitoring and the activity 25

recognition [25] [26]. Typical sequence of analysis steps start with object detection, hypothesis 26

generation, classification, and finally tracking. Classification relies in general on descriptive or 27

discriminative features identified in the analysis. In several instances, the classification relies on 28

library information from previously learned models. Bicycles don't have a precise shape and their 29

classification inherits many of the challenges. A blend of methods based on shape detection has 30

been proposed in a series of papers ( [24], [27]). The authors in [28] described a vision-based 31

bicycle classification. The algorithm relies on features detection around the wheels regions. 32

Relying on particular features as wheel detection can be challenging especially when the camera 33

field of view or traffic occlusion limits access to such features. The authors in [29] examined the 34

performance of different types of texture-based and motion features for the discrimination between 35

bicycles and pedestrians. In [29], the study used shape based and background subtraction to identify 36

and classify pedestrians and cyclists. They proposed using the distributions of typical velocity along 37

geometric information (width to height ratio values). In summary, velocity values are hence more 38

suitable to be used in conjunction with the appearance based classifier. Reported count accuracy 39

was around 81 percent [29]. 40


Table 1: Relevant Studies on the Cyclists’ data collection Study Locations Reported 15th

Percentile Cycling

Speed

Reported 85th

Percentile

Cycling Speed

Mean speed (SD) Number of

Subjects

Method1 Significant

Attributes2

Insignificant

Attributes2

[16] Signalized

Intersection

8.19 Km/Hr. 20.41 Km/Hr. 11.44 Km/Hr. (3.01) 561 1.2 2.3, 2.4 2.1

[30] 6 Signalized

Intersections

- - 8.4 to 14 Km/Hr.

(crossing) and 14.3 to

26.6 Km/Hr. (rolling

starts)

550 1.1 - -

[18] 16 Signalized

Intersections

10.8 Km/Hr. - 12.7 Km/Hr. 442 1.2 2.5, 2.7 -

[31] Citywide - - 16 Km/Hr. 8 1.5 2.1 2.6 -

[32] 2 Signalized

Intersections

10.62 Km/Hr. - 11.92 Km/Hr. 122 1.2 2.1, 2.9 -

[33] Road Segment - - 14.8 Km/Hr. (4.18) 152 1.6 - 2.1, 2.7

[34] Citywide - 22 Km/Hr. 21.6 Km/Hr. (6.01) 16 1.5 2.9 2.1

[17] 2 Trails - - 21.42 Km/Hr. (7.56) 140 1.1 - -

[35] [36] 7 Different

Sites

- - 16.5 Km/Hr. 1557 1.3 2.2,2.7,2.1 -

[20] 10 Signalized

Intersections

9.77 Km/Hr. - 14.76 Km/Hr. (5.29) 2097 1.2 2.5 -

[19] Signalized

Intersection

- - 15.84 Km/Hr. >150 1.4 2.9, 2.8, 2.6 -

[37] Trail 11 Km/Hr. 22 Km/Hr. 17 Km/Hr. (6) 367 2 - -

1Method: 1) Field Observation, 2) Manual Video Analysis, 3) Semi-automated Video Analysis, 4) Automated Analysis 5) GPS 6) Radar

2Attributes: 1) Gender 2) Helmet Usage 3)Turn\Through 4)Group 5)Standing/Roll Crossing 6)Trip environment 7) Age 8) Light condition 9) Grade


Research Procedure 1

The Computer Vision System 2

3

The automated cyclist data collection is performed using a computer vision system developed at the 4

University of British Columbia [12] (See Figure 1). The backbone of the system is the detection and 5

tracking of moving features based on the standard Kanade-Lucas-Tomasi (KLT) features tracking 6

algorithm [38] . Follow up to the tracking step is the grouping of features sharing spatial proximity 7

and dynamical similarities. The grouping suggests the trajectory of a moving object (e.g., a road-8

user). Objects are subsequently classified based on the analysis of their trajectories and speed 9

profiles [39]. In such way, only cyclist trajectories are identified and kept for further data collection 10

analysis. Further details of the computer vision system can be found in [12]. Initialization of the CV 11

system consists of setting tuning parameters to ensure proper tracking under different environment 12

conditions. 13

14

An important process in the generation of the road-users trajectories is to create a mapping from 15

world coordinates to image plane coordinates using a homography matrix (camera calibration). The 16

positional analysis of road users requires accurate estimation of the camera parameters. During 17

video recording, the three-dimensional real-world is captured on a two-dimensional image space. 18

This mapping enables the recovery of real-world coordinates of points that appear in the video. In 19

such way, the trajectory information such as positions and velocities are recorded in real-world 20

coordinates rather than pixel based coordinate. The camera calibration procedure consists primarily 21

of the estimation of certain camera parameters. In [40], this is set as an error minimization problem 22

between selected projected geometric features (e.g., points, lines) onto world plane space and the 23

actual measurements of these entities in projection (video image) space. The parameters are 24

obtained by minimizing the difference between the projection of geometric entities, e.g. 25

Corresponding Points, Distances, Angles, Global Up Directions points and lines, onto world or 26

image plane spaces and the real-world measurements of these entities. Camera calibration is 27

performed at the beginning of the data collection procedure [40]. Details of the adopted mixed-28

feature camera calibration approach are presented in [40]. 29

30

31

32 Figure 1: Computer Vision System Architecture 33

34

35


Data collection consists of providing accurate speed measurements as well as counts for the 1

cyclists. A screen line count procedure will identify the cyclist tracks crossing a predefined screen 2

lines at a selected location. The sum of the counts represents the final output of the process, which 3

are the automatic count of cyclists crossing the screen line. The object trajectory tracks from the 4

tracking algorithm are used directly as an input for mean speed measurements. Two parallel screen 5

lines with a known distance apart are specified on the screen, and the time elapsed for each track to 6

pass through the two lines is recorded. Dividing the known distance by the elapsed time generates 7

the mean speed of that particular track. 8

9

The performance of the computer vision tracking system can be assessed by its ability to accurately 10

detect cyclists. This can be measured by the validation of counts made with screen lines. The results 11

of the automatic counting process are validated with manual counts made for the same screen line 12

and same footage. Another important measure is the ability to track cyclists accurately in space and 13

time. This can be achieved by the validation of average speed measurements taken from the 14

extracted trajectories. To validate the results of the speed measurement, a specific length on the 15

image with known physical distance is used. This section of road space is selected such that there is 16

a high degree of overlap with the section defined by the screen lines in the automatic speed 17

calculations. This ensures that the results of the speed measurements are comparable. A number of 18

cyclist tracks are selected, and the time it takes for them to transverse the specific distance is 19

obtained. By dividing the length of the known distance by the time elapsed, the average cycling 20

speed is obtained for the selected cyclists. The manually calculated speed is compared to the 21

automatically measured speed. Given an n cyclists, xman is a vector of their manually calculated 22

mean speeds and xaut is a vector of the automatically calculated speeds. The Root Mean Square 23

Error (RMSE) is defined as follows: 𝑅𝑀𝑆𝐸(𝑥𝑚𝑎𝑛, 𝑥𝑎𝑢𝑡) = √(𝛴𝑖=1

𝑛 (𝑥𝑚𝑎𝑛𝑖 −𝑥𝑎𝑢𝑡

𝑖 )2

𝑛. After the validation, 24

the mean speed calculated from the automatic tracking can be used for further analysis. Cyclists are 25

classified based on their attributes, and statistical tests are used to test for significant differences 26

between the classes. 27

Case Study 28

Site Characteristics 29

30

Following a surge in roundabout designs in North America, several initiatives to build modern 31

roundabout have been undertaken in greater Vancouver [41]. The case study in this paper is a 32

recently built roundabout in the University of British Columbia main campus. The roundabout of 33

interest is situated in the south eastern entrance to the campus at the intersection of Wesbrook Mall 34

and West 16th

Avenue. This is a mixed traffic, yield controlled roundabout comprised of four 35

approaches as defined by Sakshaug et al [42]. Prior to entering the carriageway, the bicycle lane 36

merges with the sidewalk. Therefore, it can be confusing for cyclists when entering into the ring. 37

Cyclists can either merge with the motor vehicle traffic in the carriage entry bay or continue onto 38

the sidewalk to merge with the pedestrian traffic. Because cyclists can be grouped with either the 39

motor-vehicle or the pedestrian traffic, they have two alternatives to get to their destination. Data 40

collection on this location is applied as a predecessor for a traffic safety study conducted at the 41

same location. The goal is to provide information necessary to tackle concerns related to cyclists 42

behavior in general and at roundabouts in particular [43], [44]. During the study, the weather 43

conditions varied over the selected days. Out of these days, March 16, 2011, and March 17, 2011, 44

were selected to be analyzed. The first day consisted of cloudy and overcast conditions with 45

occasional sunny breaks throughout the morning. It was observed that it was mostly sunny around 46

noon and consistently cloudy for the remainder of the day. The second day had similar cloudy and 47

overcast conditions in the morning with a surprise rain shower for a very short period between 48


11:15 AM to 12:00PM. The day continued to have long periods of overcast conditions with sunny 1

breaks in between. 2

3

Data Collection 4

5

A video camera was used to record the activities that occurred over a time period of two days. It 6

was placed in a position to capture the southern and eastern approaches of the roundabout. The 7

length of each recording was 12 hours per day, beginning at 9:00 AM and ending at 9:00PM. For 8

ground truth data, the videos were manually observed to log the behavioural characteristics of 9

cyclists’ activities, assign attributes to gather information of riders, identify possible conflict events, 10

and manually count the number of cyclists. The assigned classification and attribution was 11

completed manually with excellent accuracy due to the quality of the recording. 12

13

14

15 Figure 2: Bicycles Trajectories 16

Counting Validation 17

18

For counting purposes, screen lines were strategically placed1 to capture the movements of the 19

cyclists entering and exiting the roundabout as shown in Figure 3. A screen line was installed to 20

capture cyclists using the eastern sidewalk facility. A second screen line was placed spanning 21

midway across the southern approach to count the number of cyclists entering the roundabout from 22

this approach. A third screen line captures the two lanes within the roundabout. A final screen is 23

installed to capture cyclists leaving the roundabout (leaving the university campus) eastward to 16th

24

Avenue. The counting validation was performed by comparing the manual counts with the 25

automated counting performed by analyzing the trajectories coordinates. The results of the 26

comparison indicated that the automated counts provided accuracy above 84 percent for the four 27

selected screens as detailed in Table 2. The maximum resulting classification error of the counts of 28

around 15% is considered fairly acceptable. The error is generally related to errors involving the 29

object tracks. These errors can include over-segmentation, over-grouping and missed detection. 30

Over-segmentation occurs when more than one trajectory is attributed to a single cyclist. In terms of 31

counting, the effect of this is an inflation of counts in the case that both the tracks will pass the 32

1 Installation and placement referred to the designation of the coordinates of a virtual screen so it can be used in the counting algorithm and

superimposed on the video for the manual count.


screen line. A main cause for this is when a cyclist along with its shadow are detected and 1

considered as road-users rather than a single one. Over-grouping occurs when a trajectory is 2

attributed to several cyclists. This happens when group of cyclists are traveling at nearly the same 3

speed and fairly at closed proximity with each other. The effect of this is a deflation of actual 4

counts. Some tracks may not be complete and may have not passed the screen lines to be accounted 5

for. The main factor that contributed to such inaccuracy is occlusion caused by other road-users. 6

Another reason may be due to the selected location of the screen line. It should be noted that the 7

higher accuracy for screen S4 is likely due to the clear camera angle that limited the effect of partial 8

occlusions. This is in contrast to other screens positions at which trajectories were susceptible to 9

interruption. For example, at screen S3, cyclists in the inner lane suffered greatly from occlusion 10

due to vehicles traveling at the outer lane. At screen S2, the number of missed detection of cyclists 11

is less than at the other screens. There is a low volume of traffic going through the roundabout exit 12

(near the camera), therefore occlusion from that traffic is very low reducing the chance of 13

misdetection. On the other hand, miss-detection at screen S1 is due to the distance of that screen 14

from the camera. 15

16

17

18 Figure 3: Counting Screens 19

20

Table 2: Counting Validation 21

22

s1 s2 s3 s4 Average Performance

Manual Count 211 82 315 296

Automated Count 182 89 276 286

Count Difference 29 -7 39 10

Count Accuracy 0.862559 0.914634 0.87619 0.966216 0.90490002

Undercount(UC) 39 3 69 34

Overcount (OC) 10 10 30 24

Actual Auto. Count 172 79 246 262

UC Error 0.184834 0.036585 0.219048 0.114865 0.13883299

OC Error 0.047393 0.121951 0.095238 0.081081 0.08641594

23

24

25


Speed Validation 1

2

The average speed validation was performed in a similar manner as count validation. To capture the 3

mean speed of cyclists, two adjacent screen lines, with a known physical distance between them, 4

were placed in the center of a relativity straight sidewalk segment as shown in Figure 4. If the 5

screen lines were placed within a curved section, the calculated speed would not be representative 6

because of the physical wavering nature of cyclist’s direction of travel. Therefore, the sidewalk was 7

selected since the direction of travel is more predictable and the object tracks would be 8

approximately linear. The time it took a track to cross both lines is recorded and divided by the 9

distance in order to measure average speed. 10

11

Seventy cyclists with good tracking (continuous throughout) were selected. The length between the 12

two screens lines was determined to be 14 m. The time it took for the 70 cyclists to travel across 13

this distance was timed manually through the video footage. The manual calculated average speed 14

of a sample of 70 cyclists over the two days was 14.1 km/hr. The automated system yielded an 15

average speed of 13.41 km/hr. The results of the validation showed an acceptable agreement 16

between the manual and automated calculations in Figure 5.a (RMSE = 0.374 m/s, R2

= 0.8647). 17

Cumulative distributions of the speeds for both manual and automated techniques are shown in 18

Figure 5.b. Sources of residual errors include the assumption that the cyclists follow the shortest 19

path between two check lines, potential camera calibration inaccuracy, and cyclist tracks noise. 20

21

22

23 Figure 4: Speed Screens for Speed Validation 24

25


(a) Automated Speed Evaluation

(b) Cumulative Speed Comparison

Figure 5: Speed Validation 1

Cyclist Speed Analysis 2

3

It is reasonable to assume that the average speed of cyclists will vary according to their location in 4

the facility as well as their interaction with other road-users. Figure 6 shows a partial speed 5

distribution of the cyclists as they traverse the roundabout. From the heat map variation, it is shown 6

that the speed is smaller before entering the roundabout from the southern approach. Also, the 7

average speed is low when cyclists choose to use the sidewalk. On the other hand, the average 8

speed tends to be higher when leaving the roundabout as well as within the main facility. It is also 9

interesting to note that the speed of the cyclists is reduced as they move northward and come 10

merging with vehicles entering the roundabout from the eastern side. 11

12

13


1 Figure 6: Average Speed Distribution (Km/h) throughout the Roundabout 2

3

Comparison between Travel Speed among different Cyclist Attributes 4

5

Further analysis was carried out on the data to characterize the behaviour of cyclists as they entered 6

and exited the roundabout. 753 cyclists were selected for the attributes analysis. Several attributes 7

were identified to gain further understanding of the cyclist behaviour. Factors such as direction of 8

travel, group size, helmet usage as well as travel routes can affect the behaviour of cyclist and may 9

provide some insight to their decision making. The attributes that have been assigned to each 10

cyclist were considered accurate due to the high quality of the video segments. For example, the 11

trajectory of each cyclist within the intersection is manually observed and a labeling decision is 12

made based on the cyclists’ position in a lane. A maximum of 4 categories for each variable was 13

specified, either because of finite number of options, or because this was the maximum number that 14

was considered possible for a non-intrusive classification. A summary of the classification is shown 15

in Table 3. 16

17

Table 3: Chosen Attributes and their Corresponding Categories 18

Variables Categories

Group Size 1 2+

Lane Position Left Middle Right

Helmet Use With Without

Travel Path Street Sidewalk

19

20

21

22

23

24

25

26

27


Table 4: Classification Groups Summary 1

Group Size Categories 1 2+

Count

Percent

681

92.78

53

7.22

Lane Position Categories Left Middle Right

Count

Percent

42

5.7

372

50.54

322

43.75

Helmet Use Categories With Without Can’t Tell

Count

Percent

681

90.43

53

7.03

19

2.52

Travel Path Categories Street Sidewalk Both

Count

Percent

395

53.81

303

41.28

36

4.9

2

As the location is a mixed traffic roundabout, cyclists have the option to merge with the vehicle 3

traffic on the road or the pedestrian traffic on the sidewalk. Regardless of the cyclist’s decision to 4

merge with a particular group, there are a set of rules that cyclists are advised to follow [45] . Table 5

4 shows that the majority of cyclists do stay on the road or the sidewalk. However, 5% of the 6

cyclists observed using both the road and the sidewalk. The lane position was originally considered 7

in the analysis of the cyclists’ routing decision. However, it was observed that cyclists entering and 8

exiting the roundabout do not stay in a particular lane region, such as the right, middle, and left 9

lanes. In the Bike Sense Manual published by the British Columbia Cycling Coalition, recommends 10

cyclists to keep right and to shoulder check if lane changes are required [45]. Because of low traffic 11

volume, cyclists have a tendency to freely choose which side of the lane they would prefer. As a 12

result, their lane position would change constantly along the direction of travel. It is noted that the 13

total number of cyclists varies among categories as cyclists who may be classified in more than one 14

category (e.g. an alternating position within lanes) were excluded. 15

16

17

Table 5: Significance Analysis Summary 18

19

20

Table 5 shows that the travel path, helmet use, travel lane position and group size significantly 21

affect the cyclist speed. Traveling rightmost has a slightly, but significantly higher mean speed 22

compared to cycling in the middle of the leftmost part of the lane. Mean cycling speed was higher 23

for cyclists travelling alone than in groups. However, this result may be valid for small group size 24

only as other studies have shown that cycling in a group (known as peloton in group road bike 25

Attribute Categories Statistics ANOVA/t-stat (P-Value)

Count Mean Speed (SD)

Group Size 1 681 4.66 (3.87) 0.011

2+ 53 3.96 (3.5)

Lane Position Left 42 4.06 (2.89) 0.026

Middle 372 4.5 (4.67)

Right 322 4.8 (2.98)

Helmet Use With 681 4.73 (3.79) <0.0001

Without 53 3.06 (2.29)

Travel Path Street 395 5.91 (2.45) <0.0001

Sidewalk 303 3.12 (1.24)


racing) may increase the speed of individual cyclists due to a decrease in wind resistance [46]. One 1

interesting finding was that helmet usage affects the travel speed as cyclists with helmets tend to 2

travel at higher speeds. This result is consistent with a recently published data [35], which attributed 3

this difference in the false perception of increased safety for those utilizing helmets. Figures 7(a) 4

and 7(b) show the cumulative speed distribution for different groups (sidewalk Vs Street as well as 5

Helmet Usage Vs non Helmet). 6

7

Road safety is a major concern for cyclists due to their vulnerability on the road. Safety programs 8

and measures have been developed to advise cyclists to take precautions when travelling. Safety 9

precautions that were manually observed in this study include the use of a helmet, hand signals, and 10

dismounting while crossing at a crosswalk. It was observed that 90% of the cyclists had a helmet, 11

1% of cyclists on the road used hand signals to merge lanes or indicate exit, and 1.7% of cyclists 12

who used the crosswalk dismounted and walked through. These observations illustrate cyclists’ 13

behaviour and provide insight to the number of cyclists who abide by the road rules and use the 14

proper safety precautions on the road. It is not surprising to see that the majority of cyclists are 15

wearing a helmet since there have been many studies which show that wearing a helmet can reduce 16

the number of serious head injuries [45]. It is noted that the benefit of enforcing helmet laws is still 17

a debatable subject [47]. Missing from the abovementioned data collection is the gender and bicycle 18

type. This information could not be properly extracted due to the camera angle limitation. 19

20

(a) Sidewalk Vs Street Travel Path


(b) Helmet Vs Non Helmet Usage

Figure 7.a. Percentage of Cyclists Crossings below a Given Speed 1 2

Summary 3

4

The result of the automatic counting and tracking were generally acceptable in terms of their 5

accuracy (average percentage of error). This demonstrates that automated accurate cyclist counts 6

and tracking can be performed using computer vision techniques. This can expand the possibilities 7

for cyclist data collection significantly. The automation of cyclist data collection can widen the 8

range of feasible data both geographically (different locations) and temporally (for longer periods of 9

time). Out of the 753 cyclists analyzed, the majority of cyclists exhibited predictable behaviour as 10

some travel on the road and some on the sidewalk travelling at an average speed of around 14 11

Km/Hr. Another significant observation is that only a small percentage of cyclists did not use 12

helmets when cycling (90% of cyclists were observed wearing a helmet). Several conclusions can 13

be drawn from the analysis of cyclist speed behaviour. Group size, travel path, lane position and 14

helmet usage were found to significantly affect the cyclist mean speed. Single cyclists had a 15

slightly, but significantly higher mean cycling speed compared to group cycling. The mean cycling 16

speed was highest for the cyclists using the road rather than the sidewalk. The mean cycling speed 17

decreases for non-helmet users. The study did not consider attributes such as the cyclist gender, age 18

and bike type classification. This limitation is related to the accuracy of the manual classification. 19

Because this study is non-intrusive, no validation can be performed for the classification. One area 20

of potential research is to assess a relationship between the accuracy of tracking and object density. 21

Visual inspection reveals that some cyclists were over-segmented (multiple tracks for a single 22

cyclist) or over-grouped and occluded, particularly when cyclist density was high. 23

24

Discussion 25

26

The results in this paper are primary findings that should be validated and refined by expending to 27

additional data sets. Nevertheless, this study fills an important gap on the data collection for cyclists 28


at roundabouts. Further data collection studies may cover speed behavior with different intersection 1

configurations as well as free flow facilities (road segments, designated bike lanes and trails). 2

Capacity and density estimation procedures would be developed based on the current computer 3

vision capabilities in terms of classification, counting and data collection. Allen et al. highlight the 4

need for research to quantify the effects of bicycles on the capacity of signalized intersections [48], 5

a development that would enable the creation of a composite level of service for a mixed flow of 6

intersections and eliminate the use of motor vehicle equivalents for bicyclists. 7

8

In addition to transportation applications, cyclists’ data collection can provide useful information 9

for some health applications. Cyclists data collection involving cadence (pedaling effort per unit of 10

time) is directly related to the effort [2]. Each cyclist’s pedal rotation is observed to introduce a 11

periodic fluctuation in the speed profile [49], and therefore, the cadence parameter can be computed 12

by analyzing the speed profile signal. Identifying the cadence frequency corresponds to detecting 13

the dominant periodicity in the noisy signal of the speed profile. This is reduced to the problem of 14

evaluating the power spectral density of the speed profile. It would be interesting to explore the 15

significance of cyclists’ characteristics on the cadence parameters. Primarily, results showed that 16

there is significant difference between the cadence means of cyclists using the sidewalk and those 17

using the street. Close manual inspection of the population revealed that the majority of the street 18

commuters are “experienced “with proper gearing and high level of fitness. On the other hand, the 19

general population of those using the side walk facility tend to cycle in a more casual way. A 20

thorough examination of the cadence parameters is planned for future research. 21

22

Bibliography 23

24

[1] D. Boyd, "Vancouver 2020: A Bright Green Future," City of Vancouver, 2010.

[2] D. Hartog and J. Johan, "Do the Health Benefits of Cycling Outweigh the Risks?,"

Environmental health perspectives, vol. 118, no. 8, pp. 1109-1116, 2010.

[3] J. Pucher, R. Buehler and M. Seinen, "Bicycling renaissance in North America? An update and

re-appraisal of cycling trends and policies," Transportation Research Part A: Policy and

Practice, vol. 45, no. 6, pp. 451-475, 2011.

[4] M. Winters, G. Davidson, D. Kao and K. Teschke, "Motivators and deterrents of bicycling:

comparing influences on decisions to ride," Transportation, vol. 38, no. 1, pp. 153-168, 2011.

[5] SRF Consulting Group, "Bicycle and Pedestrian Detection," Prepared for Federal Highway

Administration and Minnesota DOT, Minneapolis, MN, 2003.

[6] R. Schneider, R. Patten and J. Toole, "Case Study Analysis of Pedestrian and Bicycle Data

Collection in U.S. Communities," Transportation Research Record: Journal of the

Transportation Research Board, vol. 1939, pp. 77-90, 2005.

[7] J. Molino, J. Kennedy, P. Johnson, P. Beuse, A. Emo and A. Do, "Pedestrian and Bicyclist

Exposure to Risk: Methodology for Estimation in an Urban Environment," Transportation

Research Record: Journal of the Transportation Research Board, vol. 2140, pp. 145-156, 2009.

[8] D. P. Allen, N. Rouphail, J. E. Hummer and J. S. Milazzo, "Operational Analysis of

Uninterrupted Bicycle Facilities," Transportation Research Record: Journal of the


[9] L. Dixon, "Bicycle and pedestrian level-of-service performance measures and standards for

congestion management systems," Transportation research record: Journal of the


[10] Eco-counter, "http://www.eco-compteur.com," [Online].

[11] E. Minge, S. Petersen and J. Kotzenmacher, "Evaluation of Nonintrusive Technologies for


Traffic Detection, Phase 3," Transportation Research Record: Journal of the Transportation

Research Board, vol. 2256, no. , pp. 95-103, 2011.

[12] N. Saunier and T. Sayed, "A feature-based tracking algorithm for vehicles in intersections,"

IEEE Canadian Conference on Computer and Robot Vision, 2006.

[13] A. Laureshyn, H. Ardo, A. Svensson and T. Jonsson, "Application of automated video analysis

for behavioural studies: concept and experience," IET Intelligent Transport Systems, vol. 3, no.

3, pp. 345 -357, 2009.

[14] L. F. Miranda-Moreno, J. Strauss and P. Morency, "Exposure Measures and Injury Frequency

Models for Analysis of Cyclist Safety at Signalized Intersections," in TRB Annual Meeting

Compendium of Papers, Washington DC, 2011.

[15] D. B. Taylor and H. S. Mahmassani, "Coordinating traffic signals for bicycle progression,"

Transportation Research Record, vol. 1705, pp. 85-92, 2000.

[16] H. Ling and J. Wu, "A study on cyclist behavior at signalized intersections," IEEE

Transactions on Intelligent Transportation Systems, vol. 5, no. 4, pp. 293-299, 2004.

[17] M. R. Virkler, "Flow Characteristics On Shared Hiking/biking/jogging Trails," Transportation

Research Record: Journal of the Transportation Research Board, vol. 1636, pp. 43-46, 1998.

[18] W. Pein, "Bicyclist performance on a multiuse trail," Transportation Research Record: Journal

of the Transportation Research Board, vol. 1578, pp. 127-131, 1997.

[19] A. Kassim, L. Pascoe, K. Ismail and A. E. Halim, "Vision-Based Analysis of Cyclists' Speed,"

in TRB 91st Annual Meeting Compendium of Papers, Washington DC, 2012.

[20] D. I. Rubins and S. L. Handy, "Times of Bicycle Crossings: Case Study of Davis, California.,"

Transportation Research Record, vol. 1939, pp. 22-27, 2005.

[21] "Guide for the Development of Bicycle Facilities.," AASHTO, Washington,, 1999.

[22] "Safety and Locational Criteria for Bicycle Facilities, User Manual Volume," FHWA, U.S.

Department of Transportation, 1976.

[23] N. Buch, J. Orwell and S. Velastin, "Urban road user detection and classification using 3D wire

frame models," Computer Vision, vol. 4, no. 2, pp. 105-116, 2010.

[24] H. Cho, P. Rybski and W. Zhang, "Vision-based bicyclist detection and tracking for intelligent

vehicles," IEEE Intelligent Vehicles Symposium, pp. 454-461, 2010.

[25] B. Morris and M. Trivedi, "A Survey of Vision-Based Trajectory Learning and Analysis for

Surveillance," Circuits and Systems for Video Technology, IEEE Transactions on, vol. 18, no.

8, pp. 1114-1127, 2008.

[26] Y. Wang and G. Mori, "Human Action Recognition by Semi-Latent Topic Models," IEEE

Transactions on Pattern Analysis and Machine Intelligence , vol. 31, no. 10, pp. 1762-1774,

2009.

[27] H. Cho, P. Rybski and W. Zhang, "Vision-based Bicycle Detection and Tracking using a

Deformable Part Model and an EKF Algorithm," IEEE Conference on Robotics and

Automation, 2010.

[28] S. Messelodi, C. M. Modena and G. Cattoni, "Vision-based bicycle/motorcycle classification,"

Pattern Recognition Letters, vol. 28, pp. 1719-1726, 2007.

[29] G. Somasundaram, V. Morellas and N. Papanikolopoulos, "Counting pedestrians and bicycles

in traffic scenes," International IEEE Conference on Intelligent Transportation Systems, pp. 1-

6, October 2009.

[30] A. Wachtel, J. Forester and D. Pelz, "Signal clearance timing for bicyclists," ITE Journal, vol.

65(3), pp. 38-45, 1995.

[31] A. M. El-Geneidy, K. J. Krizek and M. Iacono, "Predicting bicycle travel speeds along

different facilities using gps data: a proof-of-concept model," in TRB 86th Annual Meeting

Compendium of Papers, Washington DC, 2007.


[32] N. Wheeler, R. Conrad and M. A. Figliozzi, "A Statistical Analysis of Bicycle Rider

Performance: The impact of gender on riders’ performance at signalized intersections," in

Proceedings 89th Annual Meeting Transportation Research Board, Washington DC, 2010.

[33] R. S. Thompson and F. P. Rivara, "Bike speed measurements in a recreational population:

validity of self reported speed.," Injury Prevention, vol. 3(1), p. 43–45, 1997.

[34] J. Parkin and J. Rotheram, "Design speeds and acceleration characteristics of bicycle traffic for

use in planning, design and appraisal," Transport Policy, vol. 17, no. 5, pp. 335-341, 2010.

[35] A. Messiah, A. Constant, B. Contrand, M.-L. Felonneau and E. Lagarde, "Risk Compensation:

A Male Phenomenon? Results From a Controlled Intervention Trial Promoting Helmet Use

Among Cyclists," American Journal of Public Health, vol. 102, no. S2, pp. S204-S206, 2012.

[36] A. Constant, A. Messiah, M.-L. Felonneau and E. Lagarde, "Investigating Helmet Promotion

for Cyclists: Results from a Randomised Study with Observation of Behaviour, Using a Semi-

Automatic Video System," PLoS ONE, vol. 7, no. 2, p. e31651, 2012.

[37] B. W. Landis, T. A. Petritsch and H. F. Huang, "Characteristics of Emerging Road and Trail

Users and Their Safety," Federal Highway Administration, 2004.

[38] J. Shi and C. Tomasi, "Good Features to Track," IEEE Conference on Computer Vision and

Pattern Recognition, , 1994.

[39] M. H. Zaki and T. Sayed, "Automated Classification of Road-User Movement Trajectories," in

CSCE International Transportation Specialty Conference, Edmonton, AB, 2012.

[40] K. Ismail, S. Sayed and N. Saunier, "Camera Calibration for Urban Traffic Scenes: Practical

Issues and a Robust Approach," Transportation Research Board Annual Meeting Compendium

of Papers, 2010.

[41] G. Furtado, "Accommodating vulnerable road users in roundabout design," in Transportation

Association Of Canada (TAC), , 2004.

[42] L. Sakshaug, A. Laureshyn, A. Svensson and C. Hyden, "Cyclists in roundabouts—Different

design solutions," Accident Analysis and Prevention, vol. 42, no. 4, pp. 1338-1351, 2010.

[43] M. Moller and T. Hels, "Cyclists' perception of risk in roundabouts," Accident Analysis and

Prevention, vol. 40, no. 3, pp. 1055-1062, 2008.

[44] S. Daniels, E. Nuyts and G. Wets, "The effects of roundabouts on traffic safety for bicyclists:

An observational study," Accident Analysis and Prevention, vol. 40, no. 2, pp. 518-526, 2008.

[45] "Bike Sense: The British Columbia Bicycle Operator’s Manual.," Greater Victoria Cycling

Coalition, Victoria, 2004.

[46] A. Iniguez and J. Iniguez, "Aerodynamics of a cycling team in a time trial: does the cyclist at

the front benefit?," European Journal of Physics, vol. 30, no. 6, p. 1365–1369, 2009.

[47] D. Robinson, "Bicycle helmet legislation: Can we reach a consensus?," Accident Analysis and

Prevention, vol. 39, no. 1, pp. 86-93, 2007.

[48] D. P. Allen, J. E. Hummer, N. M. Rouphail and J. S. Milazzo, "Effect of Bicycles on Capacity

of Signalized Intersections," Transportation Research Record: Journal of the Transportation

Research Board, vol. 1646, pp. 87-95, 1998.

[49] J. Moore, J. Kooijman and A. Schwab, "Rider motion identification during normal bicycling by

means of principal component analysis," Multibody System Dynamics, vol. 25, pp. 225-244,

2011.

1

2


automated collection of cyclist data using computer …docs.trb.org/prp/13-0745.pdf · 1 automated...

Documents