18-12-2012, 04:33 PM
A Video-based Vehicle Detection and Classification System for Real-time Traffic Data Collection Using Uncalibrated Video Cameras
[attachment=44263]
ABSTRACT
Length-based vehicle classification data are important inputs for traffic operation, pavement
design, and transportation planning. However, such data are not directly measurable by singleloop
detectors, the most widely deployed type of traffic sensor in the existing roadway
infrastructure. In this study a Video-based Vehicle Detection and Classification (VVDC) system
was developed for truck data collection using wide-ranging available surveillance cameras.
Several computer-vision based algorithms were developed or applied to extract background
image from a video sequence, detect presence of vehicles, identify and remove shadows, and
calculate pixel-based vehicle lengths for classification. Care was taken to robustly handle
negative impacts resulting from vehicle occlusions in the horizontal direction and slight camera
vibrations. The pixel-represented lengths were exploited to relatively distinguish long vehicles
from short vehicles, and hence the need for complicated camera calibration can be eliminated.
These algorithms were implemented in the prototype VVDC system using Microsoft Visual C#.
As a plug & play system, the VVDC system is capable of processing both digitized image
streams and live video signals in real time. The system was tested at three test locations under
different traffic and environmental conditions. The accuracy for vehicle detection was above 97
percent and the total truck count error was lower than 9 percent for all three tests. This indicates
that the video image processing method developed for vehicle detection and classification in this
study is indeed a viable alternative for truck data collection.
INTRODUCTION
Due to the considerable differences in performance, size, and weight between long vehicles (LVs)
and short vehicles (SVs), length-based vehicle classification data are of fundamental importance
for traffic operation, pavement design, and transportation planning. Highway Capacity Manual
(1) requires adjustments to heavy-vehicle volumes in capacity analysis. The geometric design of
a roadway, such as horizontal alignment and curb heights, is affected by the different moving
characteristics of LVs due to their heavy weight, inferior braking, and large turning radius. The
heavy weight of such vehicles is also important in pavement design and maintenance, as truck
volumes influence both the pavement life and design parameters (2). Safety is also affected by
LVs: eight percent of fatal vehicle-to-vehicle crashes involved large trucks, although they only
accounted for three percent of all registered vehicles and seven percent of total Vehicle Miles
Traveled (VMT) (3). Recent studies (4,5) also found that particulate matters (PM) are strongly
associated with the onset of myocardial infarction and respiratory symptoms. Heavy duty trucks
that use diesel engines are major sources of PM, accounting for 72% of traffic emitted PM (6).
All these facts illustrate that truck volume data are extremely important for accurate
analysis of traffic safety, traffic pollution, and flow characteristics. Unfortunately, most traffic
sensors such as single-loop detectors currently in place cannot directly measure truck volumes.
Although dual-loop detectors provide classified vehicle volumes, there are too few of them on
our current roadway systems to meet the practical needs. Considering that traffic surveillance
cameras have been increasingly deployed for monitoring traffic status on major roadways,
effective utilization of these cameras for truck data collection is of practical significance.
PREVIOUS WORK
Applying image processing technologies to vehicle detection has been a hot focus of research in
Intelligent Transpoatation Systems (ITS) over the last decade. The early video detection research
(7) at the University of Minnesota has resulted in the Autoscope video detection systems that are
widely used in today’s traffic detections and surveillance around the world. Several recent
investigations into vehicle classification via computer vision have occurred. Lai et al. (8)
demonstrated that accurate vehicle dimension estimation could be performed through the use of a
set of coordinate mapping functions. Although they were able to estimate vehicle lengths to
within 10% in every instance, their method requires camera calibration in order to map image
angles and pixels into real-world dimensions. Similarly, commercially available Video Image
Processors (VIPs), such as the VideoTrack system developed by Peek Traffic Inc., are capable of
truck data collection. However, the cost for such systems is significant and they require
calibrated camera images to work correctly.
METHODOLOGY
In order to satisfy the requirements for real-time data collection, the complexity of the approach
has to be balanced against its effectiveness. Some pattern recognition and model-matching
algorithms (17) can not be executed for real-time detection due to their over-expensive
computational cost. A background-based approach that requires less computational work is
therefore employed to meet the practical needs. Without complex calibration processes, several
simple yet effective algorithms are integrated to handle problems frequently encountered in
video-based traffic data collection, such as slight camera vibrations and shadow removal, to
enhance the overall system performance. This section describes the major algorithms of this
computer vision-based vehicle detection and classification approach. Before presenting the
details of each algorithm, the system is briefly overviewed as follows.
TRB 2007 Annual Meeting CD-ROM Paper revised
[attachment=44263]
ABSTRACT
Length-based vehicle classification data are important inputs for traffic operation, pavement
design, and transportation planning. However, such data are not directly measurable by singleloop
detectors, the most widely deployed type of traffic sensor in the existing roadway
infrastructure. In this study a Video-based Vehicle Detection and Classification (VVDC) system
was developed for truck data collection using wide-ranging available surveillance cameras.
Several computer-vision based algorithms were developed or applied to extract background
image from a video sequence, detect presence of vehicles, identify and remove shadows, and
calculate pixel-based vehicle lengths for classification. Care was taken to robustly handle
negative impacts resulting from vehicle occlusions in the horizontal direction and slight camera
vibrations. The pixel-represented lengths were exploited to relatively distinguish long vehicles
from short vehicles, and hence the need for complicated camera calibration can be eliminated.
These algorithms were implemented in the prototype VVDC system using Microsoft Visual C#.
As a plug & play system, the VVDC system is capable of processing both digitized image
streams and live video signals in real time. The system was tested at three test locations under
different traffic and environmental conditions. The accuracy for vehicle detection was above 97
percent and the total truck count error was lower than 9 percent for all three tests. This indicates
that the video image processing method developed for vehicle detection and classification in this
study is indeed a viable alternative for truck data collection.
INTRODUCTION
Due to the considerable differences in performance, size, and weight between long vehicles (LVs)
and short vehicles (SVs), length-based vehicle classification data are of fundamental importance
for traffic operation, pavement design, and transportation planning. Highway Capacity Manual
(1) requires adjustments to heavy-vehicle volumes in capacity analysis. The geometric design of
a roadway, such as horizontal alignment and curb heights, is affected by the different moving
characteristics of LVs due to their heavy weight, inferior braking, and large turning radius. The
heavy weight of such vehicles is also important in pavement design and maintenance, as truck
volumes influence both the pavement life and design parameters (2). Safety is also affected by
LVs: eight percent of fatal vehicle-to-vehicle crashes involved large trucks, although they only
accounted for three percent of all registered vehicles and seven percent of total Vehicle Miles
Traveled (VMT) (3). Recent studies (4,5) also found that particulate matters (PM) are strongly
associated with the onset of myocardial infarction and respiratory symptoms. Heavy duty trucks
that use diesel engines are major sources of PM, accounting for 72% of traffic emitted PM (6).
All these facts illustrate that truck volume data are extremely important for accurate
analysis of traffic safety, traffic pollution, and flow characteristics. Unfortunately, most traffic
sensors such as single-loop detectors currently in place cannot directly measure truck volumes.
Although dual-loop detectors provide classified vehicle volumes, there are too few of them on
our current roadway systems to meet the practical needs. Considering that traffic surveillance
cameras have been increasingly deployed for monitoring traffic status on major roadways,
effective utilization of these cameras for truck data collection is of practical significance.
PREVIOUS WORK
Applying image processing technologies to vehicle detection has been a hot focus of research in
Intelligent Transpoatation Systems (ITS) over the last decade. The early video detection research
(7) at the University of Minnesota has resulted in the Autoscope video detection systems that are
widely used in today’s traffic detections and surveillance around the world. Several recent
investigations into vehicle classification via computer vision have occurred. Lai et al. (8)
demonstrated that accurate vehicle dimension estimation could be performed through the use of a
set of coordinate mapping functions. Although they were able to estimate vehicle lengths to
within 10% in every instance, their method requires camera calibration in order to map image
angles and pixels into real-world dimensions. Similarly, commercially available Video Image
Processors (VIPs), such as the VideoTrack system developed by Peek Traffic Inc., are capable of
truck data collection. However, the cost for such systems is significant and they require
calibrated camera images to work correctly.
METHODOLOGY
In order to satisfy the requirements for real-time data collection, the complexity of the approach
has to be balanced against its effectiveness. Some pattern recognition and model-matching
algorithms (17) can not be executed for real-time detection due to their over-expensive
computational cost. A background-based approach that requires less computational work is
therefore employed to meet the practical needs. Without complex calibration processes, several
simple yet effective algorithms are integrated to handle problems frequently encountered in
video-based traffic data collection, such as slight camera vibrations and shadow removal, to
enhance the overall system performance. This section describes the major algorithms of this
computer vision-based vehicle detection and classification approach. Before presenting the
details of each algorithm, the system is briefly overviewed as follows.
TRB 2007 Annual Meeting CD-ROM Paper revised