Research On Occlusion Target Tracking Algorithm Based On Improved Siamese Network

In recent years,with the persistents innovation of artificial intelligence(AI)and deep learning(DL)technology,the field of machine vision has made advance progress in intelligent perception,intelligent computing and intelligent information processing,and has gradually become an indispensable research result of intelligent military in the future.In order to prevent vicious incidents such as sudden attacks on security borders and military strongholds,it is necessary to conduct more effective video target tracking in security monitoring.Therefore,target tracking technology has become the key research content in the field of security,especially when targets occur.When occluded,it is necessary to accurately track the target through intelligent,so as to provide security protection for the environment.Therefore,how to use DL technology to track the target,especially under the condition of occlusion or after the target reappears,has great research value and application value in the actual military environment.In the process of target tracking,as a prediction task,the target is often interfered with different degrees,such as background environment interference,self-deformation and other unfavorable factors.The occlusion of the target has always been a significant challenge in the field of target tracking.In this paper,according to the target occlusion,the degree of occlusion is different,and the occlusion is divided into three basic cases: partial,serious or short-term complete occlusion,long-term complete occlusion or beyond the field of view.On the basis of DL,in view of the above three problems,the following three aspects are improved respectively.(1)Siamese network target tracking combined with dual attentionAiming at the problem of inaccurate tracking after siamese network target tracking is interfered by partial occlusion,a dual attention mechanism based on the infrastructure is proposed.The lightweight ECA-Net channel attention mechanism is used to optimize the effective feature weight of template branches;The improved lightweight global attention mechanism GEC-Net is used to improve the feature mapping ability of search branches from two aspects of space and channel,and comprehensively improve the anti-interference ability in the tracking process.Through qualitative and quantitative analysis,it is proved that the improved algorithm has been optimized in the indicators of overlap rate(increased by 5%)and center position error(reduced by 7.34 pixels).(2)Improving the target tracking of three branch siamese networkAiming at the problems that Siam RPN++ model template can not be updated,can not adapt to severe deformation or short-term complete occlusion,an improved target tracking algorithm based on three branch siamese network is proposed.Firstly,based on the original two branches,a new template branch of the previous frame is added to update the motion state of the target.The updated motion state of the target and the initial template can locate the target more accurately;Secondly,in the classification,the triple loss is used to replace the original cross entropy loss to improve the ability to distinguish the target in the case of three branches;Finally,so as to cope with the situation that the target is completely occluded for a short time,the local expansion search is started according to the moving speed of the target,so that the position information and scale information of the target can be determined in time after the target is seriously occluded or completely occluded for a short time.Through qualitative and quantitative analysis,it is proved that the improved algorithm is optimized in the indicators such as overlap rate and center position error.(3)Shuffle Net V2 lightweight global searchAiming at the case that the target is completely blocked or even beyond the field of view for a long time in the process of target tracking,the Shuffle Net V2 lightweight network is used to optimize the global search strategy on the basis of SPLT algorithm.The problem of the slow speed of the network on GPU is solved by optimizing Shuffle Net V2,and the network is put into use global search module.Furthermore to building up the accuracy of redetection,the speed of global slide window search is gained,and experiments further verify the superiority of the improved algorithm in tracking accuracy(increased by 0.8%)and tracking score(increased by1.1%).