Occlusion Robust and Environment Insensitive Algorithm for Vehicle Detection and Tracking Using Surveillance Video Cameras

With the decreasing price of video cameras and their increased deployment on roadway networks, traffic data collection through video imaging has grown in popularity. Numerous vehicle detection and tracking algorithms have been developed for video sensors. However, most existing algorithms function only within a narrow band of environmental conditions and occlusion-free scenarios. In this study, a novel video-based vehicle detection and tracking algorithm is developed for traffic data collection under a broader range of environmental factors and traffic flow conditions. This algorithm employs a scan-line approach to generate spatio-temporal maps representing vehicle trajectories. Vehicle trajectories are then extracted by determining the Hough lines of the obtained ST-maps and grouping the Hough lines using the connected component analysis method. The algorithm is implemented in C++ using OpenCV and BOOST C++ libraries and is capable of operating in real-time. Over five hours of surveillance video footage was used to test the algorithm. Detection count errors ranged from under 1% in the relatively simple situations to under 15% in highly challenging scenarios. This result is very encouraging because the test video sets were taken under demanding conditions that ordinary video image processing algorithms cannot deal with. This implies that the algorithm is robust and able to produce reasonably accurate vehicle detection results under scenarios with adverse weather conditions and various vehicle occlusions. However, this algorithm requires approximately constant vehicle speed to perform well. Further research is necessary to extend the capabilities of the current algorithm to stop-and-go traffic conditions.