HighLight Heading
June 2002
ITS Institute researcher advances traffic monitoring
ITS Institute researcher and computer science professor Nikos Papanikolopoulos of the U of M Artificial Intelligence, Robotics, and Vision Laboratory is currently working on a number of research efforts involving vision-based sensing systems and their application to management of intersections and public spaces. He and postdoctoral researcher Osama Masoud presented their current work on monitoring freeway entrance ramps and complex intersections.
Vision-based vehicle tracking involves analysis of a live video image by specialized computer programs that can interpret the pixel data of each video frame. Moving vehicles appear to the computer as "blobs" of similarly colored pixels whose positions change from frame to frame. The tracking program uses sophisticated classification algorithms and built-in knowledge of how vehicles move to determine which blobs make up a moving car or truck.
Because blob tracking relies on movement for clues to vehicle boundaries, it is well suited to free-flowing traffic like that usually (we hope) found on freeways. But in areas where vehicles are close together, such as on freeway entrance ramps, blob tracking is more difficult. From the camera's roadside perspective, one vehicle is often partially hidden behind another (occlusion), making it difficult to accurately determine how many vehicles are in a queue. Vehicles on congested ramps also move relatively slowly and often stop, which can further confuse the image-processing system.
Masoud reported on the researchers' efforts to improve their system's ability to perform motion segmentation based on differential optical flow analysis. Although computationally expensive and sensitive to signal noise caused by camera movement, this approach offers a potential solution to the problems of occlusion and low-speed motion. Masoud also speculated that a hybrid approach combining elements of differential analysis and regular blob tracking may be the optimal solution.
Professor Papanikolopoulos found an ideal testing ground for another vehicle-tracking project literally on his doorstep: the intersection of Washington Avenue and Union Street in front of the Computer Science building is a busy "T" intersection similar to those found in many suburban settings. Traffic crossing the Mississippi River on the Washington Avenue bridge has to contend with crowds of pedestrians moving between campus buildings, as well as a steady stream of cars leaving the campus via Union Street. Inevitably, vehicles come into conflict with pedestrians and each other.
In this project, the goal is to detect situations that are likely to evolve into traffic accidents; developing a reliable detection component is a first step toward the eventual goal of building a deployable driver-warning system to help prevent intersection accidents.
Detecting accidents before they happen is essentially a problem in computational geometry. From the pixel data provided by the video sensor, the computer must extract information such as the size, position, trajectory, and velocity of vehicles and pedestrians, then determine if the configuration is likely to lead to a conflict. This in turn requires the computer to estimate the future positions of vehicles and pedestrians based on their current positions and observed movements.
Recent refinements in the software have improved the system's ability to deal with difficult lighting conditions, such as the presence of large shadows around vehicles. Because deep shadows appear to the camera as coherent blobs of dark pixels, they may be interpreted as part of the vehicle which creates them (resulting in an inaccurate vehicle size reading) or as independent vehicles. Papanikolopoulos showed screen captures from the current testing version of the software which demonstrated reliable exclusion of dark shadows.
If they can be incorporated into a deployable system, these vision-based tracking systems offer significant advantages over currently deployed traffic monitoring technologies. The researchers hope that continued development will result in a tracking system that is robust enough to be relied on in safety-critical applications.