| In recent years,with people's increasing attention to the field of security and the continuous development of computer vision technology,the surveillance video obtained by camera has been widely recognized for its intuitive and specific information expression.A large number of cameras have been placed in every corner of our lives.The explosive growth of video surveillance data brings great challenges to the subsequent storage and retrieval.As an effective method to solve the above problems,video synopsis technology has become a hot issue in the field of security.This paper analyzes and optimizes the four main parts of the video synopsis system,namely the detection of moving target,tracking of moving target,combined optimization of trajectory,and fusion of trajectory and background to generate concentrated video.Firstly,the method of deep learning is introduced into the detection and tracking of moving objects.The Deepsort algorithm fused with YOLOv3 target detector was used to realize target tracking in the video,which improved the problems of low detection accuracy and poor tracking effect in the traditional algorithm.Secondly,in order to improve the effect of trajectory combination optimization algorithm and solve the problem of frequent ID switching in complex scenes,proposes a strategy of reidentifying the trajectory where ID switching occurs,makes a qualitative analysis of the strategy,designs a tracking system for verification,and then proposes a trajectory reidentification algorithm.Then,using the extracted tracking results by analyzing and comparing the occlusion,timing sequence and correlation between the targets,the objective function is set,and the problem of obtaining the optimal trajectory combination is transformed into the problem of obtaining the minimum value of the objective function,so as to improve the effect of trajectory combination optimization.Finally,according to the optimal location of the trajectory obtained by the trajectory combination optimization algorithm,the extracted trajectory is fused with the background one by one to generate a concentrated video.Based on the human tracking data set and the video data from the laboratory camera,the experimental test of each part of the video concentration system shows that the video synopsis system developed in this paper can effectively remove the temporal and spatial redundancy in the original video.For videos with complex scenes,the concentration ratio is about 65%,while for videos with simple scenes and sparse personnel,the concentration ratio can be reduced to 4.1%,greatly shortening the video length.Compared with the original video,the generated concentrated video retains the motion information between the targets completely,and the frame is smooth and natural with high truthfulness.Especially for the surveillance video under complex scenes,it effectively reduces the problem of the same target appearing many times in the concentrated video,and improves the video's browsing ability.Therefore,the video synopsis system proposed in this paper has good engineering application value. |
