Vehicle Abnormal Behavior Detection Using Traffic Surveillance Video

In recent years, with China’s rapid economic development, the number of vehicles is rising greatly, which is causing frequent road traffic accidents and has brought great influence to people’s normal life.At present, most of the vehicle behavior monitoring are based on detection of traffic surveillance video retrieval through artificial way, which can only lead to the review of traffic accidents, not advance prevention. In order to further standardize vehicle driving, alleviate traffic congestion, then to reduce traffic accidents, vehicle abnormal behavior detection using traffic surveillance video has become the focus and difficulty on the research field of intelligent transportation system, and will play an important role in people’s daily life, as well as social progress and economy development.In this thesis, vehicle detection using traffic surveillance video, vehicle tracking, vehicle trajectory extraction and vehicle abnormal behavior detection are studied. Then some targeted algorithm improvement and technology solutions are put forward. The main work is as follows:In order to track the target vehicle, then to further extract vehicle trajectory and analyze driving behavior, first we should detect vehicle target from video surveillance data. Based on the analysis of commonly used moving object detection algorithm, this thesis puts forward an improved adaptive threshold Surendra background subtraction algorithm, with the combination of three-frame difference algorithm to achieve moving target detection and recognition. Then the experimental results show that the improved algorithm can combine the advantages of background subtraction algorithm and frame difference algorithm, which is robust for environmental interference. Due to the features of real-time and stability, the proposed algorithm can restore the moving area of vehicles, which can provide target information for the vehicle tracking procedure.CamShift algorithm is classic for moving object tracking. However, it needs to manually select the tracking target, and has failure problem in target occlusion situation. In order to solve the above shortcomings and optimize the tracking effect, this thesis uses vehicle detection results to initialize the first step of CamShift algorithm and introduces Kalman filter to predict the target motion. CamShift vehicle tracking algorithm based on Kalman filter prediction is proposed to further narrow down the search area of the target in the next frame, which can simplify the computational complexity of CamShift algorithm. As for the target occlusion failure problem, predictive value of Kalman filter is used to replace the target location calculated by CamShift algorithm, then update the Kalman filter as observed value. Experiments show that the improved algorithm can effectively resist tracking failure caused by target occlusion, and realize the automatic tracking of the vehicle movement.Through the real-time tracking of targets, moving vehicle center coordinates can be easily gained using the external target tracking rectangle. After the curve line fitting, we can get the vehicle trajectory. By the depth analysis of the trajectory data, this thesis puts forward several distinguish criteria for vehicle motion behavior, including vehicle movement direction recognition, lanes change, switching, retrograde action, and etc. It can be explained by the experimental data that the proposed technology solutions can be widely applied to identify vehicle illegal behavior, and are easy to imply with high stability.