Research On Object Tracking And Refinement Algorithm Based On Template Updating And Trajectory Guidance

Visual object tracking is a fundamental task of computer vision.In recent years,with the widespread application of deep learning,the performance of trackers have been significantly improved.The trackers based on the Siamese network have attracted wide attention due to their good performance and high tracking speed.However,when faced with a target that suffers complex conditions,the performance of the Siamese track-ers could decrease significantly.The main reason for the decrease is that the tracking template cannot accurately represent the current appearance of the target.Therefore,the intuitive solution is to update the template.However,the simple template update method cannot achieve the goal well,and may even cause performance degradation.After analysis the previous work,this paper proposes a target tracking algorithm based on dual-mask template update,which uses the semantic information and motion information of the target to generate a more comprehensive and robust target embedding representation,and explores the trend of the target appearance.This paper proposes a simple template update network.On multiple benchmarks,the performance of the algorithm in this paper can be compared with the current state-of-the-art trackers.For actual tracking,it is important to maintain the continuity of the tracking process.Once the tracker loses the target at a certain moment,it may fail to track successively in subsequent frames.This paper proposes a tracking result refinement method based on target trajectory guidance,which refines the obtained tracking results to maintain the tracking process and improve the robustness of the tracker.This method mines the target trajectory information,proposes a trajectory prediction method that can reduce the influence of camera motion,and uses the target trajectory to guide the tracker and the refinement.This method filters out the tracking results that need to be refined through the IoU prediction network,and uses the refinement network to perform refinement.After the method is installed on two trackers,the performance on multiple benchmarks are significantly improved.