Pedestrian Particle Filter Tracking And Skeleton Extraction Under Video Surveillance

As the basic work of subsequent gait and action recognition,pedestrian trajectory and skeleton information extraction has always been an important hot topic in the field of computer vision.In recent years,video surveillance has played an important role in crowd tracking,epidemic security and other public occasions.However,the information extraction accuracy and intelligent processing level of monitoring scenes still need to be improved.Traditional tracking and skeleton extraction algorithms are generally affected by the differentiated monitoring environment and monitoring noise.Around this problem,this paper studies pedestrian tracking and skeleton extraction algorithm based on video surveillance scene.The main research work of this paper is as follows:1.In order to solve the problem of drift and loss of tracking results in the basic color particle filter(CPF)tracking algorithm,this paper constructs a dual channel feature representation method based on context weight value and distance weight value,which improves the distinguishability of target features in the apparent model,and introduces particle swarm optimization(PSO)algorithm in the process of particle propagation.According to the characteristics of static and dynamic background video,the particle size update method is selected adaptively,which improves the problems of insufficient searching ability of single particle and lack of communication of group information in motion model.Experimental results show that the color particle filter algorithm based on context information and swarm optimization learning(CSO-CPF)can significantly improve the drift problem in the video sequence tracking results.The tracking pixel error is reduced by 10%-40%,and the tracking overlap rate is increased by 2%-11%.2.In the human detection stage of skeleton extraction,in order to solve the problem of false foreground in visual background extractor(Vi Be)algorithm,this paper uses morphological operation and multi frame initialization combined with pixel region smoothness judgment method to effectively eliminate the particle noise and "ghost" region.The results of video sequence detection in CASIA-A database show that the accuracy and recall of foreground detection are more than 90%.3.In the skeletonization stage of skeleton extraction,aiming at the problems of structure loss,redundancy and bifurcation existing in traditional ZS(Zhang and Suen)thinning algorithm,this paper proposes a thinning algorithm(ST-PTA)based on smooth iteration and template matching.The experimental results of MPEG-7 database show that the thinning rate is improved by 0.05-0.41% compared with other methods,and all kinds of thinning problems are improved.In the process of human joint location based on neighborhood judgment,K-Means clustering and proportional chain code search,the proposed skeletonization method is used to process the real foreground sequence of CASIS-A database,the precision and recall of human joint points are more than 89%,and the algorithm is obviously reduced by the influence of boundary noise.This paper researches the tracking algorithm and skeleton extraction algorithm in monitoring scene.Problems such as poor tracking accuracy of the traditional tracking algorithm and the influence of boundary noise on skeleton extraction algorithm are improved to a certain extent,which can significantly enhance the robustness of the actual monitoring project.