Research On Long-term Target Tracking Algorithm Based On RPN Network

Single-object long-term tracking is a technique that models the appearance and motion information of a target using a sequence of images or videos,and accurately predicts and locates the target in a long-term sequence.It is an important part of the field of computer vision and has significant applications in video surveillance,military guidance,and autonomous driving.Longterm object tracking encounters problems such as frequent occlusion,loss and reappearance,and target deformation.However,because long-term object tracking is more closely related to daily life,more and more tracking algorithms are focusing on solving long-term problems.The RPN network has excellent foreground-background classification ability and scale regression ability,and can well judge the target scale in complex long-term tracking sequences.It can achieve realtime tracking with high success rate and accuracy,and has received extensive attention and application in the field of long-term object tracking in recent years.Based on the SiamRPN algorithm,this article focuses on the research of long-term object tracking for target deformation and target loss and reappearance problems.The main research work of this article is as follows:(1)In response to the problem of template failure caused by target deformation,an improved Update-siamRPN algorithm is proposed based on the SiamRPN algorithm.When the maximum response value during tracking is lower than the set threshold,the historical frames are judged based on the temporal information of the meta-updater,and the invalid template caused by target deformation is updated to reduce the cumulative error in the subsequent tracking process.On the OTB100 dataset,compared with the SiamRPN algorithm,the success rate and accuracy of the Update-siamRPN algorithm are improved by 10.7% and 7%,respectively,and it can achieve a tracking speed of 24 frames per second.On the VOT2018 dataset,its accuracy is improved by5.7%,robustness is improved by 6.6%,and the average overlap ratio is improved by 6.1%,which effectively improves the ability to solve deformation problems.(2)In response to the problem of target loss and reappearance in long-term object tracking,an improved algorithm named Update-siamRPNLT is proposed based on the Update-siamRPN algorithm,with a structure optimized for long-term sequences.The local tracking results are evaluated and a response score is generated using the loss judgment network.When the target response score is lower than a certain threshold,the search is expanded to the global level,and the RPN network is used to optimize the location area through scale regression to obtain the tracking result.When the target response score is higher than the threshold,it is sent to the template updater for template updating.On the OTB100 dataset,compared with the UpdatesiamRPN algorithm,the success rate and accuracy of the Update-siamRPNLT algorithm are improved by 7.9% and 7.3%,respectively,in target out-of-view video sequences.In the VOT2018-LT dataset,the algorithm achieves an F-score of 0.644 while meeting the 20 FPS tracking speed,effectively improving the tracking ability of the algorithm for target loss and reappearance problems.