Research On Traffic Flow Surveillance Technology Based On Computer Vision

With the boom of China`s economy and the increasing if car ownership, almost each city suffers from the traffic problem, so that how to manage urban traffic effectively becomes a significant part of urban development. Today, solving the traffic problem with computer vision is a popular research field by the maturity of computer vision. The goal of this paper is to realizing an intelligent traffic-flow surveillance prototype system, which could obtain related statistics if traffic flow through algorithm processing real time video provided by surveillance cameras. Thus, traffic management department could adjust traffic management strategy timely due to these statistics. At last, traffic management efficiency could be improved, traffic congestion could be relieved, and traffic accidents could be reduced.The number of vehicles passing through a street over a period of time is a reflection of transportation flux. So the vehicles in video should be tracked by a multi-object tracking algorithm to obtain the number of them. Before realizing vehicle tracking, the test results of some video pre-processing and background modeling algorithms is compared, and we choose the median filter algorithm and the Mixture if Gaussian Model for background modeling, which are applicable for this paper. The moving objects, which could be tracked by multi-object tracking algorithm, would be obtained through the processing of background modeling. First, we propose a multi-object tracking algorithm based on Lucas-Kanade sparse optical flow, which stretches optical-flow tracking from single object to multiple objects; realize the multi-object tracking algorithm based on Cam-Shift; and propose a multi-object tracking algorithm based on contour center of object according to the two algorithms above, which improves the speed of object-tracking algorithm while guarantee the accuracy of it. Finally, we discuss the realizing procedure of the intelligent traffic-flow surveillance prototype system, and the experiment results are presented.Comparative experiments focused on the speed of the algorithms and the accuracy of the numbers of vehicles in different traffic situation and illumination conditions prove that the two algorithms proposed by this paper are faster and have a higher accuracy in better experiment environment. Nevertheless, the success rate of all the three algorithms descends in insufficient illumination or heavy traffic conditions. Studying the experiment results, we find that the main reasons of failure of multi-object tracking are that objects sticking together couldn`t be separated by traditional background modeling algorithms effectively and that objects stopping a long time could be regarded as background.