Adaptive Updating Of Target Model In Visual Tracking

As a basic research direction in the field of computer vision,visual tracking has a good application prospect in human-computer interaction,intelligent monitoring and automatic driving.In recent years,with the Siamese network-based target tracking method being proposed,the target model expressed by depth feature is widely used in the tracking framework.The existing tracking algorithms usually use fixed target template or locally updated target template(that is,template update for the target information of the first frame or the previous frame of the tracking target).However,on the one hand,because of the uncertainty of the target itself(such as occlusion,deformation,illumination change,etc.),on the other hand,tracking should be a continuous decision-making process.Therefore,the above scheme can not solve the problem of target template updating in tracking research.Based on the research of human basic memory model and the Siamese network tracking algorithm,this paper proposes an adaptive target template update mechanism based on global pattern,which can alleviate the impact of the uncertainty of the target itself to some extent,and provide support for the continuous tracking decision.The specific work is as follows:(1)An adaptive target model updating mechanism based on global pattern is proposed.In order to improve the robustness of tracking algorithm in complex scenes,based on the Siamese network tracking framework,this paper introduces a global adaptive target template update module,and uses the global information of the target to update under the guidance of human basic memory model.For the dynamic updating problem of adaptive target template,the objective appearance change factor described by the regularized linear regression is set up to guide the updating of target template specifically,and the regularized linear regression can make the calculation in the frequency domain,thus improving the calculation efficiency to a certain extent.(2)In order to verify the effectiveness of the above methods,this paper proposes a variety of specific update methods,which are used to guide the global update of the target model,including constant function,primary function,quadratic function,piecewise function and exponential decay.The global scale sensitivity analysis of the above methods is also carried out.By comparing the accuracy of different tracking challenges under several update modes,the global scale sensitivity of the adaptive update target model under different update modes is analyzed.Finally,the algorithm proposed in this paper is compared with the leading benchmark algorithm on the otb-50 data set.The results show that the overall performance of the adaptive target template updating algorithm based on global pattern is improved by 0.7%,and it can achieve better tracking effect in complex tracking scene challenges.