Transportation Research Part C 56 (2015) 161176

Contents lists available at ScienceDirect

Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc

Automated classication based on video data at intersections

with heavy pedestrian and bicycle trafc: Methodology and
application
Sohail Zangenehpour a, Luis F. Miranda-Moreno b,, Nicolas Saunier c,1
a
Department of Civil Engineering and Applied Mechanics, McGill University, Room 165, Macdonald Engineering Building, 817 Sherbrooke Street West,
Montral, Qubec H3A 2K6, Canada
b
Department of Civil Engineering and Applied Mechanics, McGill University, Room 268, Macdonald Engineering Building, 817 Sherbrooke Street West, Montral,
Qubec H3A 2K6, Canada
c
Department of Civil, Geological and Mining Engineering, Polytechnique Montral, C.P. 6079, succ. Centre-Ville, Montral, Qubec H3C 3A7, Canada

a r t i c l e

i n f o

Article history:
Received 8 April 2014
Received in revised form 2 April 2015
Accepted 2 April 2015
Available online 17 April 2015
Keywords:
Trafc data collection
Video analysis
Object classication
Object tracking
Road safety

a b s t r a c t
Pedestrians and cyclists are amongst the most vulnerable road users. Pedestrian and cyclist
collisions involving motor-vehicles result in high injury and fatality rates for these two
modes. Data for pedestrian and cyclist activity at intersections such as volumes, speeds,
and spacetime trajectories are essential in the eld of transportation in general, and road
safety in particular. However, automated data collection for these two road user types
remains a challenge. Due to the constant change of orientation and appearance of pedestrians and cyclists, detecting and tracking them using video sensors is a difcult task.
This is perhaps one of the main reasons why automated data collection methods are more
advanced for motorized trafc. This paper presents a method based on Histogram of
Oriented Gradients to extract features of an image box containing the tracked object and
Support Vector Machine to classify moving objects in crowded trafc scenes. Moving
objects are classied into three categories: pedestrians, cyclists, and motor vehicles. The
proposed methodology is composed of three steps: (i) detecting and tracking each moving
object in video data, (ii) classifying each object according to its appearance in each frame,
and (iii) computing the probability of belonging to each class based on both object appearance and speed. For the last step, Bayes rule is used to fuse appearance and speed in order
to predict the object class. Using video datasets collected in different intersections, the
methodology was built and tested. The developed methodology achieved an overall classication accuracy of greater than 88%. However, the classication accuracy varies across
modes and is highest for vehicles and lower for pedestrians and cyclists. The applicability
of the proposed methodology is illustrated using a simple case study to analyze cyclist
vehicle conicts at intersections with and without bicycle facilities.
2015 Elsevier Ltd. All rights reserved.

Corresponding author. Tel.: +1 (514) 398 6589; fax: +1 (514) 398 7361.
E-mail addresses: [email protected] (S. Zangenehpour), [email protected] (L.F. Miranda-Moreno), nicolas.saunier@
polymtl.ca (N. Saunier).
1
Tel.: +1 (514) 340 4711x4962.
http://dx.doi.org/10.1016/j.trc.2015.04.003
0968-090X/ 2015 Elsevier Ltd. All rights reserved.

162

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

1. Introduction
With the increase in computing power and capacity of sensors coupled with their decreasing economical cost, the eld of
Intelligent Transportation Systems (ITS) has seen considerable improvements in automated trafc monitoring systems. The
aim is not only to collect trafc data, e.g. ow, density and average speed at specic locations in the road network, but also
detailed microscopic information about each road user (position and speed) continuously and over large areas of the network. A great amount of the workload of trafc monitoring will thus shift from human operators to emerging automated
systems with improved performance and the possibility to perform new tasks such as road safety monitoring (Anon,
2001), leading to the development of more accurate and practical methods for data collection, safety diagnostics and
evaluation of trafc engineering countermeasures at critical road facilities such as intersections.
Intersections are critical elements of the road network for safety, given that a high concentration of conicts, crashes and
injuries occurs at these locations. In cities like Montreal, 60% of pedestrian and cyclist injuries occur at intersections (Strauss
et al., 2013). Given the importance of this topic, in research and practice, several recent studies have looked at different
safety issues at intersections using traditional approaches based on historical crash data (Miranda-Moreno et al., 2011)
and surrogate approaches such as conict analysis (Ismail et al., 2009b). Independent of the method for road safety diagnosis,
obtaining macroscopic and microscopic trafc data is fundamental. In the traditional safety approach, exposure measures are
often developed based on trafc counts of each user type (e.g., vehicle, pedestrian and bicycle volumes over a given time
period). In the surrogate approach, road user trajectories are necessary to compute measures such as Time To Collision
(TTC), Post Encroachment Time (PET), and gap time (Saunier et al., 2010).
Road users can be detected and classied using a variety of sensors like inductive-loops, magnetic sensors, microwave and
laser radars, infrared and ultrasonic sensors (Klein et al., 2006). However, it seems that the most convenient way to obtain
spatial data such as road user trajectories over a certain area, if not the only, is through video sensors. These sensors have several advantages, in particular the ability to capture naturalistic movements of road users with a small risk of catching their
attention, the relative ease of installation, the richness of extracted data and the relatively low cost (Saunier et al., 2011).
The main challenge with video sensors is developing an automated process to obtain trajectories by user type (e.g. for
pedestrians, cyclists and vehicles) in order to avoid manual processing that is costly and time-consuming. Automated video
processing is even more complex at urban signalized intersections which have a high mix of trafc conditions where all three
main road user types (pedestrians, cyclists and motorized vehicles) are present. This issue has attracted some attention in
research. The need for classication algorithms has been highlighted and addressed in Zaki and Sayed (2013) and Ismail
et al. (2010). Tracking and collecting observational data for cyclists and pedestrians is more difcult than for vehicles
because of their non-rigidity, more varied appearance, and less organized movements. In addition, they often move in groups
close to each other which make them even harder to detect and track.
Accordingly, this research aims to develop an automated road user classication methodology combining a tracking algorithm and a HOG approach to obtain object appearance and speeds. Different classiers are proposed to determine object
class. Moving objects are classied into three categories: pedestrians, cyclists, and motor vehicles. The proposed methodology includes different tasks: detection and tracking of each moving object and classication according to object appearance
in each frame and its speed throughout the video. Although this methodology may not be entirely novel in the eld of computer science, this work is the rst to combine and use this method in the eld of transportation. An existing open-source
tracking tool called Trafc Intelligence is used (Saunier, n.d.). As part of this research, the accuracy of different classiers is
evaluated to show the advantages of classiers using two sources of information for classication: the appearance and speed
of an object. Finally, the proposed method is demonstrated through an example which investigates the safety effectiveness of
a specic treatment at an intersection, in this case a bicycle facility. This case study aims to illustrate one of the potential
applications of the developed methodology. This methodology, however, is not limited to analysing the effectiveness of different safety treatments.
The following section provides a literature review on studies looking at object classication in trafc video, limitations
and applications.
2. Background
The literature on automated video data analysis for trafc operations and safety is very extensive. In this paper, the literature review is concentrated on studies looking at object tracking and classication as well as their applications in road safety.
Some of the gaps in this eld are also discussed.
The rst step in extracting microscopic data from a video is tracking objects (nding the position of each object in time).
In Yilmaz et al. (2006), the authors provide a general survey on object tracking. As reported in Ismail et al. (2009a) and
according to Forsyth et al. (2005), the different approaches for the detection and tracking of road users are categorized into
tracking by detection, using ow, or with probability.
Tracking by detection: object detection is done using background detection and subtraction with the current image
(Antonini et al., 2006) or a model of image appearance using color distribution or edge characteristics (Gavrila and
Munder, 2006). This approach works well in situations where the objects are well separated.

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

163

Tracking using ow: objects will be detected by matching their pixels in successive images. This approach is also called
feature-based tracking and has been used in many trafc monitoring systems such as (Saunier and Sayed, 2006).
Tracking with probability: probabilistic Bayesian tracking framework is used for object detection. In a simple case,
independent Kalman lters can be used to predict the future state of the object and track them, but this approach
may fail in situations where the objects interact and occlude each other.
Due to unavailability of public and standard benchmarks, tracking performance is still difcult to report and compare
(Ismail et al., 2009a). Similar to object detection and tracking, signicant progress has been made in object classication
for images in recent years, but generic multi-class object classication is still a challenging task. Most of the research focuses
on the design and extraction of the best features or variables to describe the objects in the images. There are two main
classes of description variables:
(1) Variables describing the objects appearance, i.e. the pixels. New features have successfully been developed which are
invariant to various image transformations like translation, rotation and scaling. Among these are the Histogram of
Oriented Gradients features (HOG) (Dalal and Triggs, 2005), Scale-Invariant Feature Transform features (SIFT)
(Lowe, 2004), Speeded Up Robust Features (SURF) (Bay et al., 2008), DAISY (Tola et al., 2010), Local Binary Patterns
(LBP) (Ojala et al., 2002) and Fast Retina Keypoint (FREAK) (Alahi et al., 2012).
(2) Variables describing the objects shape or contour. A good overview of the use of this description variable can be found
in Bose (2005). The simplest descriptions are the area and aspect ratio of the bounding box of the object.
Once object instances are turned into numerical vectors (using description variables), this becomes a traditional classication problem that can be addressed using machine learning or other techniques to learn generative or discriminative
models. A popular state of the art technique is Support Vector Machines (SVM), readers are referred to Dalal and Triggs
(2005) for an example. There is also a renewed interest in nearest-neighbour techniques for object classication (Morris
and Trivedi, 2008; Hasegawa and Kanade, 2005).
Road user classication is a useful addition to trafc monitoring systems and efforts have already been done in this area.
One of the rst methods was proposed in a simple system (Lipton et al., 1998) to classify and then track vehicles and
pedestrians. The classication was done using a Mahalanobis-based distance and a classication accuracy of 86.8% and
82.8% was achieved for vehicles and pedestrians, respectively.
Fitting a 3D model is another way to classify objects in trafc monitoring. Complex 3D models are used in Messelodi et al.
(2005) to classify vehicles into seven classes. The object description includes other visual features such as brightness and
color histograms. A SVM classier can also be used to differentiate between sub-classes, such as between bicycles and
motorcycles or between buses and trucks. A global detection rate as high as 92.5% has been reported, however this value
varies for different classes. In Kanhere and Bircheld (2008), in simple highway settings, using feature-based tracking as well
as the number of features making up the objects height, over 90% of road users were correctly classied. The work presented
in Morris and Trivedi (2008) extracts the standard description of blobs by simple morphological measurements and targets
real-time trafc monitoring on highways. Its performance is not clear as it reports results for different condence levels.
Although the work presented in Zhang et al. (2007) is called unsupervised by its authors, using k-means, it implicitly relies
on prior knowledge of the road users in the scene. The description variables are the velocity of the object area, the
compactness, dened as the ratio of the object area over the square of the object perimeter, the time derivative of the area
and the angle between the motion direction and the direction of the major axis of the shape. It should be noted that none of
these studies focused on busy locations like intersections with heavy cyclist and pedestrian trafc.
The method to count and classify composite objects presented in Somasundaram et al. (2013) relies on various descriptors combined in a Nave Bayes framework or simply concatenated as inputs of a SVM classier. The reported classication
accuracy is 92% and the counting accuracy is 95%. In Zaki and Sayed (2013), after tracking each moving object in video, the
type is classied based on speed prole information, like maximum speed and stride frequency. In this work, a classication
accuracy of 94.8% and 88.6% are reported respectively, for binary classication of motorized versus non-motorized road users
and for the classication of three main types of road users.
Finally, the use of multiple detections provided by a tracking system in each frame is one of the commonalities in this
literature. By integrating the instantaneous classication, the system achieves more robust performance, see Hsieh et al.
(2006) for some quantitative results that illustrate this point.
Very few methods in the current literature have shown the ability to collect microscopic data separately for different road
users, specically for pedestrians and cyclists. Most past studies focus mainly on obtaining data and trajectories of motor
vehicles. Due to the constant change of orientation and appearance of pedestrians and cyclists, detecting and tracking them
in video data is a difcult task. This is perhaps one of the main reasons why automated data collection methods have mainly
been developed to detect and track motorized trafc. Moreover, among the existing methods with the ability to classify road
users in trafc videos, most of them use simple classication methods based on speed or appearance and none have tried to
combine both using different criteria to improve the accuracy of the classication. Due to the lack of standardized benchmarks and public implementation, it is difcult to replicate previous work since some details remain unpublished, causing
the accuracy of these methods to be questioned. Also there are very few studies that have applied their methods to the safety

164

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

of non-motorized modes and therefore have not considered all modes and all stages: data collection, trajectory extraction,
classication and surrogate analysis.
In recent years, non-motorized safety issues have attracted a lot of attention. One particular subject of interest is
investigating the safety effectiveness of engineering treatments such as, bicycle boxes at intersections, the presence of
bicycle facilities and curb extensions. Typically two approaches have been used for evaluating safety and treatment effectiveness: traditional crash-based studies and surrogate safety methods. Despite the popularity of traditional crashed-based
studies, a general shortcoming is the need to wait for crashes to occur over several years before and after the treatments
installation and at control group sites. Implementation of traditional beforeafter studies can take a long time and demands
considerable resources, in particular for active transportation with low crash frequency. Also, the effectiveness over time can
change due to road user adaptation. Longitudinal crash-based safety studies require even more years of data, which makes
them infeasible. Because of the lack of data such as cyclistvehicle crash history as well as cyclist volumes before and after
the installation of treatments, surrogate safety analysis seems to be more suitable. Surrogate measures do not need to wait
for accidents to occur and allow for quicker evaluation of treatments and adjustments if its performance is not satisfactory.
An example of this type of analysis has been shown in St-Aubin et al. (2013).
3. Methodology
This research presents a methodology based on histograms of oriented gradients (HOG) to extract features of an image
box containing the tracked object and on a Support Vector Machine as a classier, to classify moving objects in trafc scenes
with motorized and non-motorized modes. Our method classies moving objects into three main types of road users: pedestrians, cyclists, and motor vehicles. The methodology consists of several steps: (i) tracking each moving object in the video,
(ii) classifying its appearance in each frame, and (iii) computing the probability of belonging to each class based on its
appearance and speed over time. For this purpose, several classiers are used to fuse appearance and speed to predict the
class of each object. For the rst step, classiers have to be calibrated (trained) before they can be applied to classify road
users.
All the different steps in our methodology are shown in Fig. 1.
The only prior knowledge used for designing the classiers are the speed distributions per type of road user and the dataset containing images for each type of road user, gathered by automated tracking and labeled manually. Using the same
parameters and methods described in this section, it is expected that one can replicate the same results. Additional details
of each component in the methodology are presented as follow:
3.1. Tracker
The proposed approach classies the output of a generic feature-based moving object tracker (Saunier and Sayed, 2006).
This algorithm can be summarized in two steps:
(1) Individual pixels are detected and tracked from frame to frame and recorded as trajectories using the Kanade-LucasTomasi feature tracking algorithm (Shi and Tomasi, 1994).

Video

Traffic
Intelligence
(Tracker)

Trajectory &
Image Boxes
of all Moving
Objects

Manual
Classification

Dataset of
Training
Images

Resizing

HOG Feature
Descriptors

Train the
SVM Model

(a) Training the classifier

Video

Traffic
Intelligence
(Tracker)

Trajectory &
Image Boxes
of all Moving
Objects

Resizing

HOG Feature
Descriptors

SVM Model

Integrate
Speed

Predict the
Class

(b) Predicting the object class

Fig. 1. Steps involved in (a) training the classier and (b) predicting the class of each object.

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

165

(2) A moving object is composed of many features which must therefore be grouped. Feature trajectories are grouped
based on consistent common motion.
The parameters of this algorithm are tuned through trial and error, leading to a trade-off between over-segmentation (one
object being tracked as many) and over-grouping (many objects tracked as one). Readers are referred to Saunier and Sayed
(2006) for more details. Some application of this tracker are shown in Ismail et al. (2009c), Saunier et al. (2011) and St-Aubin
et al. (2013).
3.2. Sample dataset for appearance classication
A dataset containing images for each type of road user to classify pedestrians, cyclists and vehicles, is used as prior knowledge to train and test the appearance-based classiers. Using the object trajectories provided by the tracker, the square
bounding boxes of the features of each moving object are automatically computed. The region of interest within the bounding box is saved and then manually classied into three groups: pedestrian, cyclist, and motor vehicle (1500 bounding box
for each user type). It is worth mentioning that:
 The videos used for extracting training data are different from the videos used to test the algorithms performance.
 For the training dataset, two different cameras with different resolutions and view angles were used in locations different
from where the testing videos were recorded, as can be seen in Fig. 2a for a sample of training video and Fig. 2b for a
sample of test video. This implies that the algorithm does not have a high sensitivity to camera resolution or angle as well
as to the site under study.
 The tracker does not necessarily track the entire object. It is possible that parts of the pedestrian, cyclist or vehicle are not
within the extracted image box. In this case, only part of a pedestrians body or a wheel or bumper of a vehicle is being
tracked. Since this situation will occur also during prediction, these object portions are added to the training dataset as
well (Fig. 2c and d).
3.3. Feature descriptor
The rst element to select in an appearance-based classier is the description feature or descriptor best suited to discriminate between road user classes. Among the many image descriptors documented in the literature, HOG is used since
it has been applied with success to object classication, in particular pedestrian detection in static images (Dalal and
Triggs, 2005) and vehicle detection (Kembhavi et al., 2011). HOG features concentrate on the contrast of silhouette contours
against the background. It works by dividing each image into cells in which histograms of gradient directions or edge orientations are computed. The cells making up the image can have different illumination and contrast which can be corrected
by normalization. This is achieved by grouping together adjacent cells into larger connected blocks and calculating a measure
of intensity for these new blocks. The individual cells within the block can then be normalized based on the larger block. The
implementation of the HOG algorithm used in this work is part of an open source computer vision and machine learning
library for the Python programming language (available at http://scikit-image.org/).
3.4. Feature classication
The next step is to classify the chosen descriptors into the different road user classes to obtain the base appearance-based
classier. Supervised learning methods are used for classication tasks where the classes are known and labeled instances
can be obtained (Aha et al., 1991). In this work, the instances are the HOG features computed over an image sub-region, and
their labels correspond to the road user type (pedestrian, cyclist, and vehicle). A training algorithm builds a model of the
labeled data that can then be applied to new, unlabeled input data, to predict their class. Articial neural networks
(White, 1989), K-Nearest Neighbors (KNN) (Hastie and Tibshirani, 1996) and Support Vector Machine (SVM) (Burges,
1998) are well-known supervised classiers. Among the many methods and models developed in the eld of machine learning, SVMs are one of the most commonly used classiers as they have good generalization capabilities (Burges, 1998). The
method presented in this paper relies on a SVM with the HOG features of an image sub-region as the inputs and one of the
three road user types as the output.
A SVM is by nature a binary classier. For multi-class problems, several strategies exist in the literature, such as one versus rest where a classier is trained for each class, or one versus one where a classier is trained for each pair of classes.
The SVM algorithm used in this work is the open source implementation LibSVM (Chang and Lin, 2011) available in the
OpenCV library which uses the one versus one strategy: the nal class is decided by majority vote of the underlying binary
SVMs. This appearance-based classier is called HOG-SVM.
3.5. Speed information
Aside from the appearance of an object in the video, another criterion that can help predict the type of object is its speed
(Ismail et al., 2009a). Instantaneous speed measurements can be aggregated over time and compared to a threshold to

166

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

(a) Training video sample, resolution of 800x600

(b) Testing video sample, resolution of 1280x960

(c) Sample of complete objects

(d) Sample of objects which do not include the entire road user
Fig. 2. Sample of extracted road user images used for training and testing.

eliminate a possible object type. For example, it is nearly impossible for a pedestrian to walk at a speed of 15 km/h. Speed can
also be combined with other information, such as appearance, using probability principles to increase the classication
accuracy. In this study, alternative methods to combine criteria are used to design and test different classiers.
To use speed as a criterion, one rst needs to dene a discriminative aggregated indicator of the instantaneous speed
measurements made in each frame. The usual aggregation functions are: maximum, mean, median or percentiles of the
speed measurements (e.g., 85th). Since the speed given by the tracker may be noisy and the maximum and mean are sensitive to noise, the median speed is used. From this point forward, the speed Si of object i refers to the median of each road
users instantaneous speed. The speed distributions of each of the three road user classes are considered as prior knowledge
and are gathered through automated tracking and manual object classication in the sample videos (Fig. 3). Note that other
distribution types and parameters were tested to nd the best representative types of distributions for the speeds.
3.6. Classier design
Based on the two criteria, the median of the speed measurements and the classication of the HOG-SVM in each frame,
the following classiers are derived:
Classier I: this is the simplest and relies on two speed thresholds to predict the type of each object. These two speed
thresholds are extracted from empirical speed distributions for the different types of road users. For this study, based on
Fig. 3d, the pedestriancyclist speed threshold Thpc is 6.5 km/h and the cyclistvehicle speed threshold Thcv is 14.5 km/h
(intersection points of pedestrian/cyclist and cyclist/vehicle speed distributions, respectively). Classication is performed
following:

8
classify as pedestrian
>
< If 0 6 Si 6 Thpc ;
Else if Thpc < Si 6 Thcv ; classify as cyclist
>
:
Else;
classify as v ehicle
Classier II: this classier only uses the appearance of each object through the video to predict its type (with HOG-SVM). A
method is needed to decide based on the multiple predictions made for each frame in which the object is tracked. The proportion of frames in which the object is classied can be considered as the class probability, denoted as: Pr(Ci|Ai), where Ci

167

25

100

25

20

80

20
15
10

Frequency

30

Frequency

Frequency

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

15
10

0
2

10

40
20

60

10

20

30

Median of speed (Km/h)

Median of speed (Km/h)

(a) Distribution of pedestrian speed

(b) Distribution of cyclist speed

0
-10

40

10

20

30

40

50

Median of speed (Km/h)

(c) Distribution of vehicle speed

Pr (S i | C i)

0.5

Pedestrian
Cyclist
Vehicle

0.4
0.3
0.2
0.1
0

10

15

20

25

30

35

40

Median of speed (km/h)

(d) Probability function of Pr (Si | Ci)

Fig. 3. Speed distribution and probability functions of each object type used for classiers design.

and Ai stand for class (pedestrian, cyclist or vehicle) and appearance of object i (constituted by all the image boxes extracted
during its tracking), respectively:

PrC i jAi

Number of frames that the object image box was classified as C i

Number of frames that the object appears in the video

Finally, the most likely class (the class with the highest number of detections) is the predicted type for the object.
Classier III: this classier combines both appearance-based and speed-based classiers based on a simple algorithm illustrated below to switch between the following three possible situations. Speed thresholds are chosen as the 99th percentile so
that very few road users (less than 1%) of a certain type may have a median speed above the selected threshold. The thresh99

99

olds are 7.5 km/h for Thpc and 30 km/h for Thcv and the algorithm is:

8
99
>
apply three class HOG-SVM pedestrian; cyclist;
>
< If 0 6 Si 6 Thpc ;
99
99
> Else if Thpc < Si 6 Thcv ; apply two class HOG-SVM cyclist; v ehicle
>
:
Else;
classify as v ehicle

v ehicle

(1) In the rst case, the speed of the tracked object is lower than the pedestriancyclist speed threshold: the object can
either be a pedestrian, a cyclist, or a vehicle. In this situation a HOG-SVM classier trained for the three classes is used.
(2) In the second case, the speed of the tracked object is lower than the cyclistvehicle speed threshold but higher than
the pedestriancyclist speed threshold. It is very unlikely that the object is a pedestrian: it can either be a cyclist or a
vehicle. In this situation a binary HOG-SVM classier trained for the two classes, cyclist and vehicle, is used. In this
situation it is expected that a binary classier outperforms a multi-class classier.
(3) The speed of the tracked object is higher than the cyclistvehicles speed threshold. In this situation the object can only
be a vehicle and no classier is needed.
Classier IV: this classier combines the probability of each class given the speed and appearance information using Bayes
rule and the nave assumption of independence of these two pieces of information used for classication. Although this
assumption is probably not true, the resulting classier shows good performance empirically, similarly to the nave
Bayes classiers that can still be optimal as (Domingos and Pazzani, 1997) suggested, even when the assumption of
independence is violated. To obtain this classier, consider the typical Bayesian classier given by the posterior distribution
(likelihood  prior). This is formulated as:

PrC i jSi ; Ai

PrC i
PrSi ; Ai jC i
PrSi ; Ai

168

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

where Ci, Si and Ai stand for class, speed and appearance of object i, respectively. Then, by the assumption of independence of
speed and appearance:

PrC i jSi ; Ai

PrC i
PrSi jC i PrAi jC i
PrSi PAi

Also, using conditional probability, one can write:

PrAi jC i PrC i PrC i jAi PrAi

Replacing (1) into (2), gives:

PrC i jSi ; Ai

PrC i jAi
PrSi jC i
PrSi

Finally, given that P(Si) is independent of the classes, it can be said that:

PrC i jSi ; Ai / PrC i jAi PrSi jC i

Pr(Ci|Ai) is the probability of each class obtained from classier III. Pr(Si|Ci) is estimated through distributions tted to the
empirical speed distributions of the three road user classes, gathered through manual object classication in sample videos
and shown for this study in Fig. 3ac. The speed distributions of pedestrians and vehicles are tted to normal distributions
and the speed distribution of cyclists is tted to a lognormal distribution.
The parameters of these speed distributions (Fig. 3) are the following:
(1) Pedestrian speed distribution: normal distribution with mean of Sp 4:91 km=h and standard deviation of
sp = 0.88 km/h.
(2) Cyclist speed distribution: log-normal distribution with location parameter of lc 2:31 (mean of Sc 11:00 km=h)
and scale parameter of 1c 0:42 (standard deviation of sc = 4.83 km/h).
(3) Vehicle speed distribution: normal distribution with mean of Sv 18:45 km=h and standard deviation of sv = 7.6 km/h.
Finally, the class of the object is selected as the one with the highest Pr(Ci|Si, Ai). For more information about nave Bayes
classiers and the nave assumption of independence of information please refer to Domingos and Pazzani (1997) and
Friedman et al. (1997). It is expected that using speed for classifying objects increases the classication accuracy under normal conditions (pedestrians have lower speeds than cyclists and cyclists have lower speeds that vehicles) but decreases the
accuracy under abnormal conditions like running pedestrians with high speeds, very fast cyclists or very slow vehicles.
The video data used in this paper are available upon request. The implementations of the classiers, as well as the training
and testing functions are available under an open source license on the Trafc Intelligence project website: https://bitbucket.
org/Nicolas/trafcintelligence/wiki/Road%20User%20Classication.
4. Results
In the rst part of this section, the accuracy of the classication method is evaluated. In the second part, the applicability
of our methodology is illustrated showing the entire process (data collection and processing, classication and analysis).
4.1. Accuracy
For training the HOG-SVM classier, a dataset containing 1500 square bounding boxes for each road user type (pedestrian, cyclist, and vehicle) is used. Videos from intersections of Rue Milton/Rue University and Rue Saint-Urbain/Rue
Villeneuve, in Montreal, were selected to constitute the training dataset (for more detail please refer to Section 3.2). To test
the accuracy of the designed classiers, two videos taken at two different sites from the training phase were used. The rst
video was recorded at the signalized intersection of Avenue des Pins/Rue Saint-Urbain in Montreal during peak hours for
232 min and a total of 4756 objects were manually classied to create the ground truth. To show the independence of
the proposed method to the tested condition and viewpoint, a second video was recorded from another intersection (signalized intersection of Avenue du Mont-Royal/Rue Saint-Urbain in Montreal, Fig. 7b) during peak hours (for a total of 154 min).
In this video, a total of 2909 objects were manually classied to create the ground truth. The predicted class (obtained from
each automated classier) and the ground truth (observed, manually labeled) were then compared to compute the accuracy
of each classier.
For classication problems, it is crucial to report performance measures for each class and not only the global accuracy.
The components of the confusion matrix cij are the number of objects of true class i predicted in class j. The performance
measures are thus dened globally and for each class k:

ckk
Recallk P
j c kj

ckk
Precisionk P
i c ik

P
k ckk
Accuracy P P
i
j c ij

169

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

Table 1
Confusion matrices showing each classiers performance for the intersection of Avenue des Pins/Rue Saint-Urbain.
Ground truth

Accuracy (%)

Pedestrian

Cyclist

Vehicle

Total

Precision (%)

Pedestrian
Cyclist
Vehicle
Total
Recall

946
77
0
1023
92.5%

86
324
78
488
66.4%

277
793
2175
3245
67.0%

1309
1194
2253
4756

72.3
27.1
96.5

72.4

Classier II

Pedestrian
Cyclist
Vehicle
Total
Recall

742
121
160
1023
72.5%

191
244
53
488
50.0%

584
37
2624
3245
80.9%

1517
402
2837
4756

48.9
60.7
92.5

75.9

Classier III

Pedestrian
Cyclist
Vehicle
Total
Recall

726
131
166
1023
71.0%

43
373
72
488
76.4%

64
177
3004
3245
92.6%

833
681
3242
4756

87.2
54.8
92.7

86.3

Classier IV

Pedestrian
Cyclist
Vehicle
Total
Recall

969
42
12
1023
94.7%

53
371
64
488
76.0%

180
198
2867
3245
88.4%

1202
611
2943
4756

80.6
60.7
97.4

88.5

Classier IV (balanced observation)

Pedestrian
Cyclist
Vehicle
Total
Recall

374
20
6
400
93.5%

40
308
52
400
77.0%

23
22
355
400
88.8%

437
350
413
1200

85.6
88.0
86.0

86.4

Predicted
Classier I

The results of classication for the video recorded from intersection of Avenue des Pins/Rue Saint-Urbain are shown in
Table 1. Classier IV has the best recall for pedestrians and the best precision for cyclists and vehicles, while classier III
has the best recall rate for vehicles and cyclists and the best precision for pedestrians. Overall classier IV has the best accuracy among the tested classiers. In the rst test video the majority of the trafc was motorized vehicles (around 68%) with
fewer pedestrians (around 22%) and cyclists (around 10%). In order to estimate the performance of the best designed classiers if the trafc had the same number of road users in each class, the performance for a balanced number of observations of
each user type (400 observations for each user type) is also shown in Table 1. This illustrates that the accuracy changes when
the class distribution changes, and also illustrates that the precision for cyclists is low in part because of relatively few
cyclists in the video.
The classication results for the second video, recorded from intersection of Avenue du Mont-Royal/Rue Saint-Urbain
(with use of classier IV) are shown in Table 2. From this table it can be seen that with this dataset, the accuracy of classier
IV is greater than 93%. Note that due to a higher percentage of vehicle trafc in the second video compared to the rst video
(75% versus 68% of road users respectively for the second and the rst video), the accuracy of classication for the second
video is higher than the one for the rst video. However the accuracy for the balanced observation (200 observations for each
user type) remains the same, around 86%.
It is worth mentioning that misclassication occurred in cases where multiple objects were tracked as a single object
(caused by over-grouping in the tracker output) or when only a portion of an object was tracked (caused by
over-segmentation in the tracker output). A sample of these situations is shown in Fig. 4.
Despite the satisfactory performance of the classiers, the classication of cyclists is the most challenging and has the
lowest precision and recall. This clearly shows the challenge of cyclist classication, since the cyclist can look like a
pedestrian and move at vehicular speeds.
The performance of all classiers relies on that of the tracker and therefore any tracking error may lead to error in the
classication process. If the tracker fails to track certain moving objects, these objects are lost from the dataset and never
get classied, affecting the ability to obtain precise ow counts and trajectories. In most cases, even when the tracker only
identied part of a pedestrian, cyclist or vehicle, the classiers were still able to classify the objects correctly.
Another way to visualize the results of the proposed classier is through heat-maps (number of positions in discretized
two-dimensional space bins) for the three road user classes (Fig. 5). The heat-maps show the good performance of classier
IV since the trajectories of the different road user types are overall in the expected locations: pedestrians are on the
sidewalks and crosswalks, cyclists are mostly in the bicycle facility, and vehicles are on the road, in the lanes. From the
heat-maps it is also easy to identify where the classier makes errors. For example a few cyclists in the bicycle facility
have been classied as vehicles or there are some vehicles at the top of the camera view which are classied as pedestrians
or cyclists.

170

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

Table 2
Confusion matrices showing the performance of classier IV for the intersection of Avenue du Mont-Royal/Rue Saint-Urbain.
Ground truth

Predicted
Classier IV

Classier IV (balanced observation)

Accuracy (%)

Pedestrian

Cyclist

Vehicle

Total

Precision (%)

Pedestrian
Cyclist
Vehicle
Total
Recall

424
44
28
496
85.5%

23
160
30
213
75.1%

41
30
2132
2203
96.8%

488
234
2190
2912

86.9
68.4
97.4

93.3

Pedestrian
Cyclist
Vehicle
Total
Recall

172
17
11
200
86.0%

20
151
29
200
75.5%

5
2
193
200
96.5%

197
170
233
600

87.3
88.8
82.8

86.0

Fig. 4. Example of tracked objects that are hard to classify.

(a) Snapshot of the video

(c) Cyclist trajectory heat-map

(b) Vehicle trajectory heat-map

(d) Pedestrian trajectory heat-map

(e) Scale used for trajectorie heat-maps (log-scale)

Fig. 5. Snapshot of video and position heat-maps for the three road user types (classied by Classier IV). The most and least used map locations are
respectively red and blue (heat-map colors range from blue to red, passing through cyan, yellow, and orange) (the resolution of each heat-map cell is
3  3 pixels in image space). (For interpretation of the references to color in this gure legend, the reader is referred to the web version of this article.)

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

171

(a) ROC curves for four different classifiers and three classes

(b) ROC curves for three classes and four different classifiers
Fig. 6. ROC curves showing true positive rate versus false positive rate for different parameter settings, for all classiers and road user types.

Due to the number of parameters for each classier, performance comparison between different classiers is not straightforward and can be biased by poor choices of parameters. The Receiver Operating Characteristic (ROC) (Fawcett, 2006) curve
is a tool to compare different methods over several parameter settings. Although ROC was originally designed for binary classication, it can be modied for three-class classication. A ROC curve is a graphical plot of true positive rate (true positives
out of all the positives, i.e. the recall rate) versus false positive rate (false positives out of all the negatives) for different
parameter settings. If, again, the components of the confusion matrix cij are the number of objects of true class i predicted
in class j, true positive rate and false positive rate for each class k are dened as:

ckk
True Positiv e Ratek Recallk P
j ckj
P

False Positiv e Ratek P

i;ik c ik

i;ik

j c ij

One point in a ROC space is better than another if it is closer to the top left corner (higher true positive rate with lower
false positive rate). Fig. 6a shows ROC curves separately for the four different classiers while Fig. 6b shows the ROC curves
separately for the three classes. The parameter ranges used for each classier to produce the ROC curves are presented in
Table 3. Fig. 6 shows that classifying cyclists is the hardest among the three classes and classier IV has the best performance
of all four classiers. Note that only the convex hull of all the points (false positive rate, true positive rate) is plotted for
clarity.
Since in Fig. 6a the deviation from the top left corner of the ROC curve for classier IV is less than the others, intuitively, it
can be concluded that classier IV is the least sensitive to its parameters compared to the other proposed classiers.
4.2. Application case study on cyclist safety
As an example of the applicability of the entire process, a simple case study is presented here showing all the steps from
data collection to surrogate safety analysis. The aim is to carry out a comparative analysis between two intersections: an
intersection with a bicycle facility and another without a bicycle facility, but with similar trafc and geometric characteristics. The objective of this application is to investigate the safety effect of a bicycle facility on cyclists at intersections based on
interactions between cyclists and right-turning vehicles (an interaction is constituted by each pair of cyclist in the bicycle
facility or lane and right-turning vehicle existing simultaneously in the scene) according to two surrogate conict measures.

172

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

Table 3
Parameter range for each classier to produce ROC curves.

Classier I

Number of tested parameters

Parameters

273

Thpc range from 0 to 12 km/h

Thcv range from 0.5 to 29 km/h

Classier II

66

Rescaled image size range from 25  25 to 100  100 pixels

Cells per block in {1, 2, 4}
Pixels per cell in {4, 8, 12}

Classier III

109

Thpc range from 6 to 15 km/h

x
Thcv range from 25 to 35 km/h
Rescaled image size range from 40  40 to 80  80 pixels
Cells per block in {1, 2, 4}
Pixels per cell in {4, 8}

Classier IV

184

Thpc range from 6 to 9 km/h

x
Thcv range from 25 to 30 km/h
Cells per block in {2, 4}
Pixels per cell in {4, 8}
Pedestrian speed standard deviation sp range from 0.7 to 1.05
Cyclist speed scale parameter range from 0.34 to 0.5
Vehicle speed standard deviation sv range from 6 to 9.1

For the intersection with a bicycle facility we chose the intersection of Rue Saint-Urbain/Avenue des Pins and for the intersection without a bicycle facility we chose Rue Saint-Urbain/Avenue du Mont-Royal (Fig. 7a and b). Specic trajectories of
objects can be isolated based on their origins and destinations and then classied to obtain trajectory heat-maps for cyclists
passing through and for right-turning vehicles (Fig. 7c and d). These gures show the strengths of the isolation and classication methods. These heat-maps can also be used to identify conict areas at intersections as well as the potential collision
angle between cyclists and right-turning vehicles.
The most common denition of a trafc conict is an observable situation in which two or more road users approach
each other in time and space to such an extent that there is a possibility of collision if their movements remain unchanged
(Hyden and Amundsen, 1977). As surrogate measures of safety, the classical time-to-collision (TTC) measure as well as the
post-encroachment-time (PET) are used for comparison purposes. TTC is dened at each instant as the time until two objects

(a) Intersection with a bicycle facility

(b) Intersection without a bicycle facility

(c) Heat-map of cyclists and right-turning vehicles (d) Heat-map of cyclists and right-turning vehicles
at an intersection with a bicycle facility
at an intersection without a bicycle facility
Fig. 7. Chosen intersections for studying the effectiveness of bicycle facilities (a and b) and heat-maps of the trajectories of cyclists and right-turning
vehicles at these intersections (c and d).

173

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

Table 4
Surrogate measures for the intersections with and without a bicycle facility.

Measurements

Without bicycle facility

With bicycle facility

Hours of video
Cyclists
Right-turning vehicles
Average cyclist speed
Average vehicle speed

2.57
119
263
11.8
12.3

3.88
438
622
15.2
13.7

TTC15 < 5 s
TTC15 < 1.5 s
PET < 5 s
PET < 1.5 s
TTC conict rate*
TTC dangerous conict rate*
PET conict rate*
PET dangerous conict rate*

4
2
37
2
328
164
3038
164

13
4
161
10
185
57
2293
142

Conicts per million potential conicts.

would collide if their movements remain unchanged: this depends on predicting the road users future positions, which is
generally done at constant velocity. For a detailed discussion and comparison of motion prediction methods, the readers are
referred to Mohamed and Saunier (2013). PET is dened as the observed time between the departure of the encroaching
cyclist from the conict point and the arrival of the rst vehicle to the conict point at the intersection or vice versa
(Gettman and Head, 2003). TTC and PET are computed for all interactions. TTC is then aggregated over time to produce a
single indicator summarizing the interaction severity: a percentile of the distribution (the 15th percentile, TTC15) is chosen
over the minimum TTC to avoid sensitivity to noise. Based on these two measures, dangerous conicts are dened as an
interaction with TTC15 or PET below 1.5 s. Conict rates (for conict with TTC15 or PET below 5 s) and dangerous conict
rates are dened as follows:

Interactions with TTC 15 less than 5 seconds; per Hour  106

Tracked Cyclists per Hour  Tracked Vehicles per Hour

TTC Dangerous Conflict Rate

PET Conflict Rate

Interactions with TTC 15 less than1:5 seconds; per Hour  106

Tracked Cyclists per Hour  Tracked Vehicles per Hour

Interactions with PETlessthan 5 seconds; per Hour  106

Tracked Cyclists per Hour  Tracked Vehicles per Hour

Conflict Rate

Without Bicycle Facility

Without Bicycle Facility (Cumulative)

With Bicycle Facility

With Bicycle Facility (Cumulative)

90
80
70
60
50
40
30
20
10
0

350
300
250
200
150
100
50
0
0

0.5

1.5

2.5

3.5

4.5

Cumulative Conflict Rate

TTC Conflict Rate

Without Bicycle Facility

Without Bicycle Facility (Cumulative)

With Bicycle Facility

With Bicycle Facility (Cumulative)

Conflict Rate

600

3500
3000
2500
2000
1500
1000
500
0

500
400
300
200
100
0
0

0.5

1.5

2.5

3.5

4.5

Cumulative Conflict Rate

(a) TTC15 (seconds)

(b) PET (seconds)

Fig. 8. TTC15 (a) and PET (b) distributions for the intersections with and without a bicycle facility.

174

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

PET Dangerous Conflict Rate

Interactions with PET lessthan1:5 seconds; per Hour  106

Tracked Cyclists per Hour  Tracked Vehicles per Hour

The units are in conicts per million potential conicts (the number of potential conicts is dened as the total number of
cyclists in the bicycle facility/lane per hour multiplied by the total number of right-turning vehicles per hour).
Unlike TTC based on motion prediction at constant velocity, PET is not sensitive to speed noise and each conict only has
one specic value. The number of tracked cyclists and right-turning vehicles as well as TTC15 and PET for each pair of cyclists
and right-turning vehicles were derived automatically from the classied road user trajectories. Results are reported in
Table 4.
The distributions of TTC15, PET and their cumulative distributions (the ones less than 5 s) for both intersections are shown
in Fig. 8.
Both TTC15 and PET results suggest that the intersection with a bicycle facility is safer than the one without a bicycle facility (Fig. 8). Due to the effect of speed noise on the computation of TTC15 based on motion prediction at constant velocity and
the availability of more data for PET (by denition of right turning interactions, it is possible that one conict does not have
TTC15 but still has PET, this can be seen in TTC15 and PET distributions in Fig. 8), it seems logical to consider PET as a better
surrogate safety measure in this study and rely more on its results.
This case study shows one of the applications of our methodology but does not intend to draw conclusions about the
effectiveness or safety of bicycle facilities at intersections. To obtain conclusive results, data for a representative sample
of intersections both with and without bicycle facilities are needed.

5. Final discussion
Since the tested classiers have different precision and recall rates, the choice of the best classier depends on the application and preference for missed detections or false alarms for one class or another. For example, if it is important to detect
as many pedestrians as possible at the expense of other road users being classied as pedestrian, classier IV is the best
(recall rate of 94.7% for pedestrians). On the other hand, if it is important that no other road user other than pedestrian is
classied as pedestrian then classier III is the best (precision of 87.2% for pedestrians). Overall, classier IV (accuracy of
88.5%) has the best performance among the tested classiers, showing the advantage of methods using both sources of information for classication: objects appearance and speed. There are several ways to improve the accuracy of the designed
classiers:
(1) Using video data from different viewpoints to train the classier. Using this approach, the classier is generalized for
different camera angles. One question is whether the performance will break down if the viewpoints become too
different. The question of using more consistent viewpoints with the same angle is also raised as it may improve
appearance-based classication by reducing the variability of object appearance.
(2) As discussed previously, the classier accuracy relies on the performance of the tracker algorithm: a way to improve
classication accuracy is therefore to improve the tracking algorithm. These are some ideas that can help improve the
tracking performance:
 Increase the camera angle to see objects separate from each other in crowded scenes since one of the major issues
of the tracker is over-grouping in dense trafc.
 Compensate the sheye effect of the camera lens. A camera with a sheye lens was used to cover as much of the
intersection as possible. However, sheye lenses produce strong visual distortion at the corners of the video frame
(Fig. 2b). This effect reduces the accuracy of the tracker to map the position of objects in real world coordinates and
speed estimation. By correcting for the sheye effect of the camera, the usage of position and speed of an object will
be more reliable for classication.
(3) In this paper HOG and SVM with a radial basis function were used as feature descriptor and classier. Other feature
descriptors and classiers should be tested to see if better accuracy can be achieved.
(4) Background subtraction is another possible way to increase the performance of the classiers, especially to obtain
more precise images of each object (more precisely around its contour), although this will not help if road users
overlap.
(5) Since the appearance of each object in the center of video is richer than its appearance in the edge of the video (mostly
due to viewpoint and sh eye effect of the camera), assigning more value and weight to the classied frames of each
object in the center of the video can possibly improve the overall accuracy of classication.
6. Conclusion
The need of microscopic data (trajectories, speeds, counts) classied by user type is more and more recognized in the
transportation literature in general and in trafc safety in particular. This research presents a novel methodology to design,
integrate and combine classiers capable of classifying moving objects in crowded trafc video scenes (for example at
intersections) into three main road user types: pedestrians, cyclists, and motor vehicles. Given the limitations of single

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

175

classication methods based on speed or appearance, this research combines these methods through several classiers in
order to improve the classication performance.
Among the four tested classiers, the one that combines the probability of both the objects appearance and speed
achieved systematically better performance (classier IV) than the other tested classiers. Overall the accuracy of the best
classier (classier IV) is greater than 88%. Due to the similarity in appearance between pedestrians and cyclists (a cyclist
consists of a bicycle and a human who rides a bicycle) and of the large range of cyclist speed, cyclists are the most difcult
road user to classify. False positive rates for the best classier are 19.4% for pedestrians, 39.3% for cyclists, and 2.6% for
vehicles, while the rates for false negative are 5.3%, 24.0%, and 11.6%, respectively. However due to the lack of available
benchmarks and accessibility to other methods, comparison with other classication methods is not possible. To address this
issue, a software implementation of the methods presented in this work is available as open source software.
As part of the contributions, the entire process is illustrated, from video data collection to surrogate safety analysis, using
cyclistvehicle interactions at intersections as an application environment. Our methodology integrates a set of open-source
tracking codes that have been used in previous work and extended for the current application. The applicability of our methodology is illustrated for automated surrogate analysis in situations with crowded and mixed trafc.
As part of the future work, different bicycle/pedestrian treatments will be evaluated such as bicycle boxes and curb extensions using surrogate measures. The methodology can be also improved according to the points highlighted in Section 5.
Alternative video sensors can also be tested such as thermal cameras to deal with some of the limitations of the regular video
cameras in low light conditions, shade, adverse weather, and occlusion in high density conditions.
Acknowledgments
This work is part of a research project on vulnerable road users safety funded by the Qubec Fund for Research on Nature
and Technology (FRQNT), the Qubec Ministry of Transportation (MTQ), and the Qubec Fund for Health Research (FRQS) as
part of their research program on road safety. The authors would like to thank Taras Romancyshyn, research assistant, for
helping with video recording and the manual classication analysis as well as Ting Fu for helping with video recording
and Jillian Strauss for proofreading this paper.
References
Aha, D.W., Kibler, D., Albert, M.K., 1991. Instance-based learning algorithms. Mach. Learn. 6 (1), 3766, <http://link.springer.com/10.1007/BF00153759>
(accessed 23.07.13).
Alahi, A., Ortiz, R., Vandergheynst, P., 2012. FREAK: fast retina keypoint. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition, pp. 510517.
<http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=6247715> (accessed 26.05.13).
Anon, 2001. Trafc Monitoring Guide. U.S. Department of Transportation, Federal Highway Administration, Washington, DC.
Antonini, G. et al, 2006. Behavioral priors for detection and tracking of pedestrians in video sequences. Int. J. Comput. Vis. 69 (2), 159180, <http://link.
springer.com/10.1007/s11263-005-4797-0> (accessed 24.06.13).
Bay, H. et al, 2008. Speeded-up robust features (SURF). Comput. Vis. Image Understand. 110 (3), 346359, <http://linkinghub.elsevier.com/retrieve/pii/
S1077314207001555> (accessed 22.05.13).
Bose, B., 2005. Classifying Tracked Moving Objects in Outdoor Urban Scenes. EECS Department, MIT, Research Qualifying Examination Report.
Burges, C., 1998. A tutorial on support vector machines for pattern recognition. Data Mining Knowl. Discov. 167, 121167, <http://link.springer.com/article/
10.1023/A:1009715923555> (accessed 16.07.13).
Chang, C.-C., Lin, C.-J., 2011. LIBSVM: a library for support vector machines. ACM Trans. Intell. Syst. Technol. 2 (3), 127, <http://dl.acm.org/citation.
cfm?doid=1961189.1961199> (accessed 22.05.13).
Dalal, N., Triggs, B., 2005. Histograms of oriented gradients for human detection. In: 2005 IEEE Computer Society Conference on Computer Vision and
Pattern Recognition (CVPR05), vol. 1, pp. 886893. <http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1467360> (accessed 23.05.13).
Domingos, P., Pazzani, M., 1997. On the optimality of the simple Bayesian classier under zero-one loss. Mach. Learn. 130, 103130, <http://link.
springer.com/article/10.1023/A:1007413511361> (accessed 08.12.14).
Fawcett, T., 2006. An introduction to ROC analysis. Pattern Recogn. Lett. 27 (8), 861874, <http://linkinghub.elsevier.com/retrieve/pii/
S016786550500303X> (accessed 06.08.13).
Forsyth, D.A. et al, 2005. Computational studies of human motion: Part 1, tracking and motion synthesis. Found. Trends Comput. Graph. Vis. 1 (2/3), 77254,
<http://www.nowpublishers.com/product.aspx?product=CGV&doi=0600000005> (accessed 29.05.13).
Friedman, N., Geiger, D., Goldszmidt, M., 1997. Bayesian network classiers. Mach. Learn. 29 (23), 131163, <http://link.springer.com/article/10.1023/A:
1007465528199> (accessed 07.11.14).
Gavrila, D.M., Munder, S., 2006. Multi-cue pedestrian detection and tracking from a moving vehicle. Int. J. Comput. Vis. 73 (1), 4159, <http://link.
springer.com/10.1007/s11263-006-9038-7> (accessed 30.05.13).
Gettman, D., Head, L., 2003. Surrogate safety measures from trafc simulation models. Transport. Res. Rec. 1840 (1), 104115, <http://trb.metapress.com/
index/T264J4213G6V863Q.pdf> (accessed 19.07.12).
Hasegawa, O., Kanade, T., 2005. Type classication, color estimation, and specic target detection of moving targets on public streets. Mach. Vis. Appl. 16 (2),
116121, <http://link.springer.com/10.1007/s00138-004-0163-4> (accessed 25.07.13).
Hastie, T., Tibshirani, R., 1996. Discriminant adaptive nearest neighbor classication. IEEE Trans. Pattern Anal. Mach. Intell. 18 (6), 607616, <http://
ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=506411> (accessed 24.07.13).
Hsieh, J.-W. et al, 2006. Automatic trafc surveillance system for vehicle tracking and classication. IEEE Trans. Intell. Transport. Syst. 7 (2), 175187,
<http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=1637673> (accessed 25.07.13).
Hyden, C., Amundsen, F.H., 1977. No Title. Proceeding of First Workshop on Trafc Conicts. Institute of Transport Economics, Oslo/Lund Institute of
Technology, Oslo, Norway.
Ismail, K., Sayed, T., Saunier, N., 2009a. Automated collection of pedestrian data using computer vision techniques. Transportation Research Board Annual
Meeting. <http://nicolas.saunier.conns.net/stock/ismail09automated.pdf> (accessed 07.12.12).
Ismail, K., Sayed, T., Saunier, N., 2009b. Automated pedestrian safety analysis using video data in the context of scramble phase intersections. Annual
Conference of the Transportation Association of Canada. <http://n.saunier.free.fr/saunier/stock/ismail09automated-tac.pdf> (accessed 07.12.12).
Ismail, K., Sayed, T., Saunier, N., et al, 2009c. Automated analysis of pedestrianvehicle conicts using video data. Transport. Res. Rec.: J. Transport. Res.
Board 2140, 4454, <http://trb.metapress.com/index/5N8257G765U252H8.pdf> (accessed 18.07.12).

176

S. Zangenehpour et al. / Transportation Research Part C 56 (2015) 161176

Ismail, K., Sayed, T., Saunier, N., 2010. Automated analysis of pedestrianvehicle conicts context for before-and-after studies. Transport. Res. Rec.: J.
Transport. Res. Board 2198, 5264, <http://trb.metapress.com/openurl.asp?genre=article&id=doi:10.3141/2198-07> (accessed 20.04.12).
Kanhere, N., Bircheld, S., 2008. Real-time incremental segmentation and tracking of vehicles at low camera angles using stable features. IEEE Trans. Intell.
Transport. Syst. 9 (1), 148160, <http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4405641> (accessed 25.07.13).
Kembhavi, A., Harwood, D., Davis, L.S., 2011. Vehicle detection using partial least squares. IEEE Trans. Pattern Anal. Mach. Intell. 33 (6), 12501265, <http://
www.ncbi.nlm.nih.gov/pubmed/20921579> (accessed 29.07.13).
Klein, L.A., Mills, M.K., Gibson, D.R.P., 2006. Trafc Detector Handbook, third ed., vol. I.
Lipton, A., Fujiyoshi, H., Patil, R., 1998. Moving target classication and tracking from real-time video. In: Proceedings Fourth IEEE Workshop on Applications
of Computer Vision. WACV98 (Cat. No. 98EX201), pp. 814. <http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=732851> (accessed 25.07.13).
Lowe, D.G., 2004. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis. 60 (2), 91110, <http://link.springer.com/10.1023/B:VISI.
0000029664.99615.94> (accessed 21.05.13).
Messelodi, S., Modena, C.M., Zanin, M., 2005. A computer vision system for the detection and classication of vehicles at urban road intersections. Pattern
Anal. Appl. 8 (12), 1731, <http://link.springer.com/10.1007/s10044-004-0239-9> (accessed 25.07.13).
Miranda-Moreno, L.F., Strauss, J., Morency, P., 2011. Disaggregate exposure measures and injury frequency models of cyclist safety at signalized
intersections. Transport. Res. Rec.: J. Transport. Res. Board 2236, 7482, <http://trb.metapress.com/openurl.asp?genre=article&id=doi:10.3141/223609> (accessed 24.09.13).
Mohamed, M.G., Saunier, N., 2013. Motion prediction methods for surrogate safety analysis. Transport. Res. Rec.: J. Transport. Res. Board 2386, 168178,
<http://trb.metapress.com/openurl.asp?genre=article&id=doi:10.3141/2386-19> (accessed 05.03.14).
Morris, B.T., Trivedi, M.M., 2008. Learning, modeling, and classication of vehicle track patterns from live video. IEEE Trans. Intell. Transport. Syst. 9 (3),
425437, <http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4534830> (accessed 25.07.13).
Ojala, T., Pietikainen, M., Maenpaa, T., 2002. Multiresolution gray-scale and rotation invariant texture classication with local binary patterns. IEEE Trans.
Pattern Anal. Mach. Intell. 24 (7), 971987, <http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1017623> (accessed 22.05.13).
Saunier, N., Sayed, T., 2006. A feature-based tracking algorithm for vehicles in intersections. In: The 3rd Canadian Conference on Computer and Robot Vision
(CRV06), pp. 5959. <http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=1640414> (accessed 22.07.13).
Saunier, N., Sayed, T., Ismail, K., 2010. Large-scale automated analysis of vehicle interactions and collisions. Transport. Res. Rec.: J. Transport. Res. Board
2147, 4250, <http://trb.metapress.com/openurl.asp?genre=article&id=doi:10.3141/2147-06> (accessed 28.07.13).
Saunier, N. et al, 2011. Estimation of frequency and length of pedestrian stride in urban environments with video sensors. Transport. Res. Rec.: J. Transport.
Res. Board 2264 (1), 138147, <http://n.saunier.free.fr/saunier/stock/11-01-trb-stride-poster.pdf> (accessed 09.07.13).
Saunier, N., n.d. Trafc Intelligence. <https://bitbucket.org/Nicolas/trafcintelligence/> (accessed 01.04.13).
Shi, J., Tomasi, C., 1994. Good features to track. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition CVPR-94, pp. 593600.
<http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=323794> (accessed 20.09.13).
Somasundaram, G. et al, 2013. Classication and counting of composite objects in trafc scenes using global and local image analysis. IEEE Trans. Intell.
Transport. Syst. 14 (1), 6981, <http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6291788> (accessed 26.07.13).
St-Aubin, P., Miranda-Moreno, L., Saunier, N., 2013. An automated surrogate safety analysis at protected highway ramps using cross-sectional and before
after video data. Transport. Res. Part C: Emerg. Technol. 36, 284295, <http://linkinghub.elsevier.com/retrieve/pii/S0968090X13001794> (accessed
23.01.14).
Strauss, J., Miranda-Moreno, L.F., Morency, P., 2013. Cyclist activity and injury risk analysis at signalized intersections: a bayesian modeling approach. Accid.
Anal. Prevent. 59, 917, <http://linkinghub.elsevier.com/retrieve/pii/S0001457513001905> (accessed 23.05.13).
Tola, E., Lepetit, V., Fua, P., 2010. DAISY: an efcient dense descriptor applied to wide-baseline stereo. IEEE Trans. Pattern Anal. Mach. Intell. 32 (5), 815830,
<http://www.ncbi.nlm.nih.gov/pubmed/20299707> (accessed 12.06.13).
White, H., 1989. Learning in articial neural networks: a statistical perspective. Neural Comput. 1 (4), 425464, <http://www.mitpressjournals.org/doi/abs/
10.1162/neco.1989.1.4.425> (accessed 24.07.13).
Yilmaz, A., Javed, O., Shah, M., 2006. Object tracking. ACM Comput. Surv. 38 (4), 13-es, <http://portal.acm.org/citation.cfm?doid=1177352.1177355>
(accessed 21.05.13).
Zaki, M.H., Sayed, T., 2013. A framework for automated road-users classication using movement trajectories. Transport. Res. Part C: Emerg. Technol. 33,
5073, <http://linkinghub.elsevier.com/retrieve/pii/S0968090X13000806> (accessed 26.07.13).
Zhang, Z. et al., 2007. Real-time moving object classication with automatic scene division. In: 2007 IEEE International Conference on Image Processing, pp.
149152. <http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=4379787> (accessed 01.08.13).