Research On Video Target Detection And Tracking Algorithm And Apply In Public Security Criminal Investigation System

At present,the issue of public safety has received more and more attention from people.A large number of monitoring systems have been applied to public places such as campuses,plazas,bus stations and railway stations.Traditionally,when public security personnel use surveillance video to solve crimes,they need to artificially watch large numbers of surveillance videos to find target suspects.The workload is very heavy and prone to errors.Therefore,this project intends to conduct research on the target detection and tracking of pedestrians in video surveillance.Through the pedestrian recognition technology,the pedestrian targets in videos are automatically detected,located and tracked,so as to reach the goal of reducing the workload of public security personnel and improving the efficiency of detection.Target pedestrian detection and tracking technologies in video mainly include pedestrian detection,pedestrian re-identification,and pedestrian tracking.The following is the main research content of this article.(1)Analyze and research the classic pedestrian detection algorithm.Pedestrian detection mainly includes pedestrian feature extraction,classifier training and post-processing.The HOG-SVM method and the aggregation channel method are mainly analyzed.In the HOG-SVM method,only the gradient method histogram features are used;in the aggregate channel method,the image is screened by cascading classifier with the characteristics of color(LUV),gradient histogram and amplitude.In order to extract image features quickly,the method of quick image pyramid feature is used to extract image features.Finally,simulate in INRIA pedestrian database and analysis of the results.(2)Research on pedestrian Re-identification algorithm.The algorithm of metric learning and the algorithm of pedestrian re-identification algorithm based on neural network are deeply analyzed,and studies the network layers in the residual network model in detail.In order to improve the recognition rate of pedestrian re-identification,a generation model is introduced in the residual network model,namely a generative anti-network model(GAN)which improve the recognition ability of the network model.In order to further improve the network recognition capability,multiple lables are assigned to generation pictures.Finally,simulate in DukeMTMC-reID pedestrian re-identification database.(3)Research on pedestrian tracking.Based on the analysis of Mean-Shift tracking and fusion methods,we deeply study kernel correlation filtering(KCF)tracking methods,such as cyclic matrix,ridge regression and kernel function.In view of the problem of color interference and the invariable pedestrian window in the KCF method,the color feature and multi-scale detection method are introduced to improve the problem.For the occlusion problem,an occlusion discrimination mechanism is introduced to update the model better.In order to meet the real-time requirements for tracking,the multi-scale detection of the target pedestrians is used once every several frames.Finally,simulation experiment is carried out in the pedestrian tracking video set.