31-08-2016, 11:44 AM
[attachment=70155]
Abstract—In this paper, we describe a novel algorithm that
counts and classifies highway vehicles based on regression analysis.
This algorithm requires no explicit segmentation or tracking
of individual vehicles, which is usually an important part of
many existing algorithms. Therefore, this algorithm is particularly
useful when there are severe occlusions or vehicle resolution is
low, in which extracted features are highly unreliable. There are
mainly two contributions in our proposed algorithm. First, a
warping method is developed to detect the foreground segments
that contain unclassified vehicles. The common used modeling
and tracking (e.g., Kalman filtering) of individual vehicles are
not required. In order to reduce vehicle distortion caused by the
foreshortening effect, a nonuniform mesh grid and a projective
transformation are estimated and applied during the warping
process. Second, we extract a set of low-level features for each
foreground segment and develop a cascaded regression approach
to count and classify vehicles directly, which has not been used
in the area of intelligent transportation systems. Three different
regressors are designed and evaluated. Experiments show that our
regression-based algorithm is accurate and robust for poor quality
videos, from which many existing algorithms could fail to extract
reliable features.
Index Terms—Highway vehicle, image warping, cascaded
regression.
I. INTRODUCTION
VIDEO cameras could be used to record the traffic information
constantly or continuously. We are thus able
to analyze the traffic videos in real time and discover any
information of interest. One fundamental task is to count the
vehicles passing by in a given time period and classify the
vehicles into different categories at the same time. The counting
and classification results could be useful in many different
applications. For example, they could be used to measure traffic
density, traffic flow, and even emissions in terms of pollutants
and greenhouse gases.
Counting and classification also could be done by other
sensors such as radar, infrared, and inductive loop detectors.
Although some sensors could be more accurate, they could also be intrusive and need a higher maintenance cost. For example,
we may need to embed weighing sensors in road to measure
vehicle weight and classify vehicle size. Comparing with other
sensors, vision-based systems could be non-intrusive and could
obtain much richer traffic information. However, current visionbased
systems could be less accurate and more sensitive to
operating conditions (e.g., weather). These problems make
vision-based systems challenging and important research topics
in the area of intelligent transportation systems.
A typical vision-based traffic analysis system could consist
of many components such as foreground segmentation, shadow
removal, feature extraction, and tracking [1]. In order to count
and classify vehicles, there is often a module to detect and
separate individual vehicles for each foreground segment. This
module could be conducted after feature extraction or tracking.
For example, if feature points could be extracted robustly across
multiple image frames, it is possible to fit explicit 2D/3-D
vehicle models [2], [3]. This kind of algorithms usually requires
at least moderate-resolution images without severe occlusions
and motion blur. In this paper, we would like to process lowquality
videos by skipping this module. In our collected videos,
multiple vehicles could be occluded and thus form a large foreground
segment. Separation or inference of individual vehicles
would be a difficult task in this case. Moreover, a 2-D vehicle
shape could be strongly distorted caused by the foreshortening
effect, which means the weak perspective projection used in the
traditional algorithms is not a good approximation. As the video
frame rate could be as low as one frame per second, the vehicle
size could be reduced to less than 10 × 10 pixels for the next
image frame. Therefore, it would also be difficult to detect and
track robust feature points or edges. Fig. 1 shows several image
frames in our collected videos.
There are mainly two contributions of the proposed algorithm.
First, we develop a warping method to detect image
foreground segments that contain unclassified vehicles. In order
to reduce vehicle distortion between two consecutive image
frames, we estimate a non-uniform mesh grid and a projective
transformation. By using this warping method, we do not need
the common used tracking or modeling of individual vehicles.
We either do not need to assume the weak perspective projection
that has been used in some existing algorithms (e.g.,
[2], [4]). In fact, the warping method could be considered as
an approximation of the perspective projection when vehicle
features are not reliable. Furthermore, the warping method
based on the non-uniform mesh grid is designed for both curved
(as shown in Figs. 1© and 7) and straight highway sections.
To our knowledge,
Abstract—In this paper, we describe a novel algorithm that
counts and classifies highway vehicles based on regression analysis.
This algorithm requires no explicit segmentation or tracking
of individual vehicles, which is usually an important part of
many existing algorithms. Therefore, this algorithm is particularly
useful when there are severe occlusions or vehicle resolution is
low, in which extracted features are highly unreliable. There are
mainly two contributions in our proposed algorithm. First, a
warping method is developed to detect the foreground segments
that contain unclassified vehicles. The common used modeling
and tracking (e.g., Kalman filtering) of individual vehicles are
not required. In order to reduce vehicle distortion caused by the
foreshortening effect, a nonuniform mesh grid and a projective
transformation are estimated and applied during the warping
process. Second, we extract a set of low-level features for each
foreground segment and develop a cascaded regression approach
to count and classify vehicles directly, which has not been used
in the area of intelligent transportation systems. Three different
regressors are designed and evaluated. Experiments show that our
regression-based algorithm is accurate and robust for poor quality
videos, from which many existing algorithms could fail to extract
reliable features.
Index Terms—Highway vehicle, image warping, cascaded
regression.
I. INTRODUCTION
VIDEO cameras could be used to record the traffic information
constantly or continuously. We are thus able
to analyze the traffic videos in real time and discover any
information of interest. One fundamental task is to count the
vehicles passing by in a given time period and classify the
vehicles into different categories at the same time. The counting
and classification results could be useful in many different
applications. For example, they could be used to measure traffic
density, traffic flow, and even emissions in terms of pollutants
and greenhouse gases.
Counting and classification also could be done by other
sensors such as radar, infrared, and inductive loop detectors.
Although some sensors could be more accurate, they could also be intrusive and need a higher maintenance cost. For example,
we may need to embed weighing sensors in road to measure
vehicle weight and classify vehicle size. Comparing with other
sensors, vision-based systems could be non-intrusive and could
obtain much richer traffic information. However, current visionbased
systems could be less accurate and more sensitive to
operating conditions (e.g., weather). These problems make
vision-based systems challenging and important research topics
in the area of intelligent transportation systems.
A typical vision-based traffic analysis system could consist
of many components such as foreground segmentation, shadow
removal, feature extraction, and tracking [1]. In order to count
and classify vehicles, there is often a module to detect and
separate individual vehicles for each foreground segment. This
module could be conducted after feature extraction or tracking.
For example, if feature points could be extracted robustly across
multiple image frames, it is possible to fit explicit 2D/3-D
vehicle models [2], [3]. This kind of algorithms usually requires
at least moderate-resolution images without severe occlusions
and motion blur. In this paper, we would like to process lowquality
videos by skipping this module. In our collected videos,
multiple vehicles could be occluded and thus form a large foreground
segment. Separation or inference of individual vehicles
would be a difficult task in this case. Moreover, a 2-D vehicle
shape could be strongly distorted caused by the foreshortening
effect, which means the weak perspective projection used in the
traditional algorithms is not a good approximation. As the video
frame rate could be as low as one frame per second, the vehicle
size could be reduced to less than 10 × 10 pixels for the next
image frame. Therefore, it would also be difficult to detect and
track robust feature points or edges. Fig. 1 shows several image
frames in our collected videos.
There are mainly two contributions of the proposed algorithm.
First, we develop a warping method to detect image
foreground segments that contain unclassified vehicles. In order
to reduce vehicle distortion between two consecutive image
frames, we estimate a non-uniform mesh grid and a projective
transformation. By using this warping method, we do not need
the common used tracking or modeling of individual vehicles.
We either do not need to assume the weak perspective projection
that has been used in some existing algorithms (e.g.,
[2], [4]). In fact, the warping method could be considered as
an approximation of the perspective projection when vehicle
features are not reliable. Furthermore, the warping method
based on the non-uniform mesh grid is designed for both curved
(as shown in Figs. 1© and 7) and straight highway sections.
To our knowledge,