Detection And Tracking Of Illegal Targets Based On Video Stream

Based on the increasing number of traffic violations at intersections in China,the traffic violation evidence can be obtained by using video stream to detect and track traffic violations.The punishment and education by it can effectively improve citizens' awareness of traffic safety and reduce the incidence of accidents.In the detection of traffic violation targets in video stream,there is a situation that the crowd density is too high and easy to be detected by mistake.In the process of tracking,there is a problem that the tracking success rate and accuracy decrease due to occlusion and other factors.For those problems,this paper studies the following aspects:1.In order to solve the problem of false detection of illegal targets in traffic video,this paper proposes a border regression method based on the YOLOv3 target detection algorithm.Through the combination of skin color detection and face detection and then use the position score function to predict the pedestrian non motor vehicle target frame regression operation.2.A Staple tracking algorithm with adaptive weighting coefficient is proposed.Using the similarity between the correlation filter response and the ideal Gaussian response to realize the adaptive weighting coefficient of the final response function.3.For the current situation that the illegal target is easily occluded during the tracking process,we propose an occlusion re-tracking method based on the trajectory prediction equation and the filtered template memory network.A memory filter template is built in the occlusion frame,and the candidate prediction frames obtained by the trajectory prediction equation are respectively calculated with the filter template and the memory filter template.The target position is determined according to the response result,and the filter template and the memory filter template are updated by the filter template memory unit.Through the processing and comparison analysis of some intersection monitoring video data of Ningbo traffic department,it is verified that the proposed method has good processing effect on occlusion tracking,and has low false detection rate,high tracking success rate and high precision,and good spatial robustness.