| With the growth of people's demand for public safety and the rapid development of monitoring system,massive cameras are deployed to people's lives,such as schools,shopping malls,streets and so on,which also leads to exponential growth of surveillance video data.Due to the drawback of manual browsing,such as time-consuming,heavy workload,low browsing efficiency,low video utilization rate and high rate of missing detection,it is necessary to introduce new methods to solve the contradiction that massive video information is rich but hard to excavate.Video synopsis is an effective way to deal with this problem.In this paper,we first design and implement the overall framework of video synopsis,including several key modules,such as generation and updating of background images,trajectory extraction,trajectory combination,trajectory fusion.And then two key processes that affect the final effect of synopsis video,namely trajectory extraction and trajectory combination,and the shortcomings of the previous algorithms are analyzed.For trajectory extraction,we propose a motion detection technology based on full convolutional network,and combined with the actual needs of video synopsis,a more suitable trajectory extraction scheme is designed to preserve the information of moving targets more completely.For trajectory combination,we propose a combination algorithm considering the target similarity to generate more compact and orderly concentrated video.we first calculate the similarity between the target,and then transform trajectory combination into an energy function,which considers the similarity,collision,the length of synopsis video,finally by minimizing the energy function to reorganize each trajectory into the synopsis video to remove the space-time redundancy.Experiments show that the synopsis video obtained by this method is more orderly and more coherent than traditional methods without loss of condensation ratio,which can further improve the efficiency of user browsing analysis. |
PDF Full Text Request