Surveillance Video Synopsis Based On Object Trajectory

Recently the video surveillance network becomes tremendous and generates massive video data. Traditional methods of video surveillance synopsis gradually are unable to meet the application requirements. The intelligent video surveillance technologies come into being. Based on the computer vision and other subjects, the intelligent video surveillance technologies can solve various problems of video surveillance and improve efficiency greatly. As the important field of intelligent video surveillance, video synopsis can extract effective information, reduce the spatio-temporal redundancy and obtain a compact video to describe the original video. It can be used for video storage, browsing and retrieval. However, objects missing is the common problem in most of existing video synopsis methods which degrade the description ability of the compact video, as well as the true collision and false collision among objects, therefore they are not completely applicable to the surveillance video. Aiming at the above problems, this thesis proposed a video synopsis method based on object trajectory optimization. The main work consists of the following two parts:(1).For the disconnection of extracted moving object trajectories and the phenomenon of objects blink resulted by collision between objects, this thesis introduces an improved trajectories extracting method. This method merges the objects, between which the true collision happen, thereby maintain the correlation between the true collision objects and the continuity of object trajectories. Experiments verified the proposed method.(2).The energy function is set to measure the constraints from the original video to the synopsis video, by the Maximum a’ Posterior state of a special Markov Random Field, which can be solved by Relaxed Linear Programming method. The optimal trajectory set is obtained which can enhance the compact ratio with the reduction of the false collision phenomenon. Finally, the object trajectories of the optimal solution are stitched with and the background images together by the Gaussian distribution to generate a synopsis video frame sequences, thereby to synthesize the synopsis video. The comparative experiments have been conducted to verify the described method and analyzed the results at last.The result of experiments suggest that the proposed method outperforms other methods on compact ratio, and have better visual effect while guaranteeing the information integrality of the original video.