| Object tracking is one of the most fundamental tasks in the field of computer vision,which is defined as achieving precise localization of the tracked object in subsequent frames,given its initial object position in a video sequence.As one of the research hotspots in recent years,scholars have proposed a large number of effective object tracking algorithms,most of which are based on deep learning.However,the training of high-performing feature extraction network requires large-scale labeled data,and the collection and labeling of datasets are timeconsuming and labor-intensive,hence self-supervised learning strategy is introduced to object tracking task to alleviate this problem.Nevertheless,the performance of the existing object trackers based on self-supervised learning still need to be improved,as most of them extract shallow information of the tracked objects,and the random sampling strategy fails guarantee to obtain effective motion change information of the objects.Based on this,the research of this thesis is as follows:(1)To address the problem that the benchmark UDT model only extracts the shallow semantic information of the object,and it is difficult to reverse verification when the forward tracking faces tracking challenges such as deformation and occlusion,this thesis proposes a Multi-frame Consistency Correction Based Self-supervised Siamese Network Tracker(MCCSST).Firstly,the visual feature enhancement module is proposed to integrate the global and local feature information of the object in an adaptive weighted manner to improve the discrimination ability of the model.Secondly,the forward multi-frame reverse sequence verification strategy module reasonably selects the optimal path optimization model.Finally,a mixed-order correction module is proposed to correct the tracking offset.Extensive experiments show that the proposed algorithm has a significant improvement in performance compared with the benchmark self-supervised object trackers.(2)For the benchmark UDT model utilizes random sampling and directly cropping the central region as the tracking object,which makes model reduce the attention to the change of obejct appearance.In order to solve this problem,a Self-supervised Object Tracker Based on Multi-Cycle Contrast Enhanced Consistency(Self-MCEC)is proposed.Firstly,the training samples are selected with a fixed frame interval to fully capture the valuable temporal motion information of the object.Secondly,the deep semantic features of the objects are extracted through the attention mechanism,which contributes to accurately locating the object location from the complex background.Finally,the object appearance variation is increased by data augmentation techniques while a contrast-enhanced consistency loss is constructed to optimize model.The experimental results show that our method achieve excellent performance on multiple benchmark datasets. |
