Research On The Techniques Of Vision-Based Traffic Data Collection

With the rapid development of economy in our country, the urban traffic is becoming very busy, which caused more serious traffic problems, such as traffic congestion, environmental pollution and traffic accident. The scientific and systematic analysis of traffic data is the effective way to solve traffic problems. In order to analyze and improve the efficiency of traffic system, real-time and accurate traffic information should be extracted. Vision-based vehicle detection technique becomes a hotspot of research due to their easy installation and maintenance, low cost, and ability to collect rich traffic information and monitor a wider area.Vision-based traffic data collection relates to many common research topics including moving vehicle detection, vehicle classification and traffic parameter extraction etc. This paper respectively design and implement traffic data collection system for day and night. The experimental results demonstrate that the two systems can detect moving vehicle effectively. In addition, this paper makes research on model-based vehicle location and recognition. The main contents of the paper are as follows:(1)We discussed the moving vehicle detection approach by background subtraction for daytime. The background is modeled by Gaussian Mixture Model and the parameters of GMM are discussed in detail. We proposed a shadow removal method by combining a shadow model in RGB space with edge feature. Finally, traffic data is counted by setting region of interests on the road. The experimental results show that the algorithm is effective to vehicle detection at daytime.(2)Model-based vehicle detection method is more robust to view point change, image noise and occlusion etc. Besides, it can obtain three dimensional information of object. We focus on three-dimensional deformable vehicle model matching problem and proposed a convex cube model. The3D model is matched to the2D image by FDA-based algorithm. The results demonstrate our method leads to good performance with respect to efficiency, accuracy and robust to occlusions. (3)We proposed a feature-based approach to robustly detect vehicles at night time. We developed an adaptive thresholding method to extract headlights and initialize vehicle hypotheses by pairing and grouping headlights. Then, these vehicle hypotheses are associated among frames by spatial and shape analysis to confirm vehicle presence. Experimental results demonstrate that the proposed approach has low computing cost and high detection rate.