Study On Vehicle Detecting And Tracking In Traffic Video Surveillance

With the rapid development of the urbanization, traffic congestion and environment pollution have become more and more serious. As an effective approach to solve the above problems, Intelligent Transportation System has drawn more and more attentions. Intelligent traffic surveillance system based on the image processing is an important subject of the Intelligent Transportation System. Moving vehicle detecting and tracking is a basic and key work in intelligent traffic surveillance. In this dissertation, some key problems in traffic surveillance such as moving vehicle detecting and tracking are well studied. The major contributions of this dissertation are summarized as follows:1.In order to make the variance of the online K means model updating algorithm converge more rapidly, a fast background subtraction algorithm based on the Mixture Gaussian Model is presented.Different learning rate of the mearn and the variance are adopted.The more rapid convergence of the variance and the better robustness to the variation of the environments are obtained by the variable variance learning rate based on the counter. Applying the proposed algorithm to the moving vehicle detection, the information of the vehicle such as position can be obtained accurately.2.Based on the research on the spatial-temporal correlation of the motion vector in the traffic video series, two different search strategies are presented, one being triangle search and the other hierarchic three-step search. According to the temporal correlation of the motion vector, the above two strategies are unified in the algorithm named a fast correlation tracking algorithm based on the classified search. The search strategy is flexible. The tracking precision is guaranteed and the efficiency is promoted.3.The CamShift algorithm only can be applied to tracking the targets in certain color. To deal with this problem, an improved algorithm named adaptive color space vehicle tracking algorithm is proposed. According to the new measurements of the objects and the background, the color space is selected dynamicly. The vehicle position is iterativly calculated in the color probability distribution image. Through the judgment of the color space updating and the Kalman filter, the algorithm efficiency is improved. The application fields of the CamShift algorithm are extended into the moving vehicle tracking by the proposed algorithm. The reliable and stable tracking results are obtained.4.To increase the robustness to the variation of illumination or background and the large overall motion of the vehicle, the adaptive weights multi-feature spatial histogram Mean Shift tracking algorithm is presented. Based on the framework of production fusion, the Mean Shift iterative solution is deduced. According to the similarity of the object and the background of the three features, the weights are adjusted, which make the tracking results no longer depend on the single feature too much. The accurativ tracking in complex background is realized.5.In order to solve the problem of vehicle tracking under occlusions efficiently, an anti-occlusion vehicle tracking algorithm based on the particle filter is proposed. Firstly, two improvements are adopted to improve the particle filter algorithm based on the kernel histogram, one being the adaptive state transition model and the other Gaussian segmentation resampling. Secondly, the blocking similarity and the position judgments are used, through which no-occlusion, partial occlusion and serious occlusion are distinguished. Lastly, the complete modaling Mean Shift algorithm, the partial modeling mean shift algorithm and the improved particle filter algorithm are used to track the vehicles in the different states. The vehicle under long time occlusions or repeated occlusions can be tracked accurately.