Research On Vehicle Behavior Detection Based On HEVC Compressed Domain

At present, researches on the detection of vehicle behavior events are mainly focused on the pixel domain, and researches on the moving target detection and the moving target tracking which are related to the vehicle behavior event detection are also focused on the pixel domain. In the early video coding standards, there were already scholars studied how to divide and detect the moving target in the compressed domain, but the technologies they had used were still belong to those technologies applied to the pixel domain. The latest HEVC video coding standards are studied in this paper, and according to the motion vector, coding unit mode and residual information contained in the compressed stream, a set of vehicle traffic behavior event detection method based on HEVC compression domain are proposed Event detection method. The three research points of this thesis are as follows:(1) Motion vehicle target detection. A motion vehicle target detection algorithm based on threshold is proposed by using motion vector, coding unit division and prediction residual data as vehicle target characteristic information. First, through the partial decoding analysis of HEVC video compression stream, we can extract the motion vector, coding unit division mode and residual information, and get a more reliable motion vector field after the motion vector assignment, motion vector normalization and median filtering in the motion vector field. Then we propose a motion vector assignment method for a PVM intra-coding unit based on coding unit partitioning and intra prediction residuals, which can reduce the misinterpretation of intra-frame coding units. Finally, we propose the target detection algorithm of the moving vehicle target based on the pre-processed motion vector field, and realize the target detection of the moving vehicle in HEVC compression domain. Experiments show that the proposed algorithm can detect the moving vehicle target more accurately in the relatively fixed road monitoring video of the camera.(2) Target vehicle tracking. The motion vector is used as the moving information and characteristic information of the vehicle target, and we propose a target vehicle tracking algorithm based on block matching. The algorithm makes full use of the motion vector as the motion information of the moving vehicle target in the adjacent video frame, and the tracking of the moving vehicle target in the continuous video frame is realized by a certain degree of motion prediction and the motion target region block matching strategy. Experiments show that the algorithm proposed can be more effective to achieve the tracking of moving vehicle targets in the relatively fixed background of the road surveillance video.(3) Vehicle behavior detection and rapid detection of parking events. Firstly, the center coordinates of the vehicle target are extracted by using the target tracking results in the continuous video frame. And then through the analysis of vehicle travel trajectory, and the compressed domain in the vehicle behavior event detection criteria, we successfully detected the vehicle behavior events including vehicle behavior events are mainly vehicle movement direction detection, retrograde,U-turn, etc; Then we propose a fast detection algorithm for parking events based on motion vectors.The algorithm does not need to carry out vehicle target detection and tracking, and only uses the motion vector information of the coding unit to judge the parking event. The experimental results show that the proposed algorithm can effectively detect vehicle behavior events and quickly detect parking events.