Development Of Video Synopsis Algorithm And GPU Optimization

With more and more attention people paid to public safety, the requirement of security monitoring systems becomes higher. How to quickly find effective information in mass surveillance video data has become the pressing needs of the industry. Video synopsis technology is highly concentrated on the original video and can keep the key information of the original video at the same time. Video synopsis is the effective measure to solve this problem.This thesis researches the principle and framework of the current mainstream video synopsis and analyzes the key technology needed in video synopsis, including moving target detection and tracking, trajectories extraction, combination optimization, background fusion and so on. This thesis adopts VIBE background modeling algorithm to detect moving objects. According to the method of combining the object detection and tracking, this project achieves multiple target tracking based on the KCF single tracking algorithm, futherly obtains and saves the trajectories data. The combinatorial optimization method based on the energy function is adopted to shift the trajectories on the timeline. Aim to solve the optimal solution of the energy function, the simulated annealing algorithm is improved. This thesis designs the simulated annealing algorithm based on the K-means clustering, solving the problem of the generation of initial solution. The visual effect of video synopsis is optimized. Finally fuse the trajectories and background images to generate the summary video.This thesis tests and analyzes the modules in video synopsis. Design the GPU parallel computing for background modeling, target tracking, background fusion, and implement the thread parallel acceleration.Finally through the experiment of system evaluation, the video synopsis this thesis develops can remove the spatial and temporal redundancy in the video, greatly shortens the time of the original video, and completely retains the original motion information in original video. The video synopsis can meet the needs of quickly scanning the video. In the meantime, verify the influence of the GPU parallel computing on the performance of every module and the whole system. Therefore, the video synopsis this thesis develops has good engineering application value.