Research On Target Detection And Tracking Technology In Complex Occlusion Environment

In recent years,with the rise of computer vision technology,moving target detection and tracking has been widely used in military guidance,intelligent video surveillance,smart city and many other aspects.However due to the diversity of practical application scenes,the traditional moving target detection and tracking algorithm has poor detection and tracking performance in many complex scenes,such as dynamic background,the presence of moving target in the initial frame,the intermittent movement of target,target deformation,target rotation,background clutter,target occlusion and scale change.Therefore,in order to improve the detection and tracking performance of the traditional algorithm in the above complex scenes,the paper improves the traditional moving target detection and tracking algorithm.The work done in the paper includes:(1)For the traditional Visual Background extraction algorithm(Vi Be)in the dynamic Background,the initial frame of the movement of the target,as well as the target intermittent movement scene,prone to misdetection and miss detection.In this paper,a background modeling method based on spatiotemporal samples is proposed,which fully combines the advantages of Gaussian Mixture Model(GMM)algorithm with better performance in temporal background modeling and Vibe algorithm with better performance in spatial background modeling,and adaptively improves the traditional algorithm on foreground point detection and background model updating.The experimental results show that the proposed algorithm improves the comprehensive evaluation index measure F-by about 8% compared with the traditional algorithm,which greatly improves the robustness of the proposed algorithm in complex scenes.(2)The traditional Kernel Correlation Filtering target tracking algorithm adopts a single feature in appearance model building,and has low tracking accuracy in complex scenes such as target deformation,target rotation and background clutter.In this paper,a target tracking algorithm based on many features adaptive fusion is proposed.Firstly,the peak graphs of the individual responses of different features are analyzed.Then,according to the confidence degree of the response peak graph of different features in different scenarios,the sub-peak main-peak ratio(RSFM)of the response peak graph was used to assign different weights to different features.Finally,the position of the tracking target is determined according to the relevant response of the fused features.The experimental show that the algorithm in this paper improves the tracking accuracy and success rate obviously compared with the traditional algorithm in the complex scene of target deformation,target rotation and background clutter.(3)The target tracking algorithm of traditional kernel correlation filtering is easy to cause tracking loss in the scene of target occlusion and scale change.In this paper,a target tracking algorithm based on block anti-occlusion and scale adaptive kernel correlation filtering is proposed.When the target is partially occluded,the position of the target can be determined by the combination of the unoccluded sub-block and the global block.When the target is globally occluded,the algorithm will automatically expand the target search area to re-detect the target.A new occlusion detection mechanism based on adaptive update model is proposed for tracker model update.When the scale of the target changes,the algorithm can choose the corresponding scale factor according to the aggregation degree of each sub-block,and then adjust the scale of the tracker adaptively.Experimental show that the proposed algorithm improves the tracking accuracy by 17% and 16%,and the tracking success rate by20% and 13%,respectively,compared with the traditional kernel correlation filtering algorithm in the scene of target occlusion and scale change.