Researches On Self-correction Single Object Tracking

As a classic research topic in computer vision,visual tracking is widely used in intelli-gent transportation,autonomous driving and surveillance,and has great social and economic value.The success of deep learning has greatly promoted the development of visual tracking.Siamese network based trackers formulate the visual tracking mission as an image match-ing process,and has attracted much attention because of real-time and accuracy.Due to occlusion,distractors and change in viewing angle,the appearance features of the object are unstable.The Siamese network matches the subsequent images according to the initial frame,so it is difficult to effectively distinguish the changes in the appearance of distractors and objects.Therefore,it is of great significance to research the object tracking and reloca-tion algorithm.In this thesis,aiming at the feature confusion caused by occlusion and distractors in the tracking process,we introduce feedback control theory in Siamese network based tracker,construct a tracking-decision-correction object relocation framework.We propose an arbiter based on histogram matching filter,design correction models based on structural similarity(SSIM)and contrastive learning,and explore the object relocation method based on feed-back theory,which effectively improves the anti-interference and long-term performance of single object tracking.The work content of this thesis is summarized as follows.(1)Aiming at the problem of object loss in the visual tracking,we analyze the types and causes of tracking failure firstly,find the U-shaped phenomenon of the histogram by count-ing the histogram of the center distance of the candidate objects.The Band Rejection Filter and Transfer Arbiter(BRT)model is proposed,which can effectively detect the phenomenon of object loss.Secondly,we propose a correction model based on SSIM,which uses the min-imum similarity principle to construct the template set,and selects the correct object accord-ing to the KM algorithm and the maximum structural similarity criterion.The experiment results show that our method can detect tracking failures in time and correct them accurately,which improves the performance of the tracker.(2)Aiming at the few-shot characteristics of visual tracking,we introduce self-supervised learning,propose a correction model based on dynamic contrastive learning.Firstly,two sample generation modes are designed referring to BRT model.Secondly,a data augmen-tation method based on central consistency is proposed to overcome the similar problem of positive and negative samples caused by random crop.Finally,to solve the problem that SSIM is sensitive to geometric deformation.Semantics SSIM(Se-SSIM)is designed to guide the contrastive learning to learn the difference between the semantic features of the object and the distractor online.It can improve the distinguishing ability of the corrector.The experimental results show that the model can effectively identify the distractor,which further improves the performance of the tracker.