Target Domain Adaptation In Visual Tracking

Existing visual tracking algorithms mainly use visual features of targets from independent frames as target representations.Due to the diversity and complexity of tracking videos in the real world,there is a domain gap in the representation capability of visual features(i.e.,visual features present differ-ent representation capability in the source and target domains),which limits the performances of existing trackers.To alleviate the effects of the domain gap problem,existing trackers perform target domain adaptation from two as-pects:enhancing target representations and performing online learning.These trackers combine different types of visual features to enhance target represen-tations but pay less attention to exploit the spatial structural information and temporal coherence in the target domain effectively.These trackers learn the tracking models online in the target domain to facilitate the feature extractors to adapt to target appearances and improve the utilization ability of the decision-making modules to target representations.However,existing online learning strategies often encounter the overfitting problem,which causes the decision-making modules to fall into local attention and the feature extractors to forget the objectness information.In order to perform effective target domain adapta-tion,this thesis proposes four solutions in terms of enhancing target represen-tations and improving existing online learning strategies:(1)This thesis pro-poses to use the spatial structural information to enhance target representations and proposes a structure and appearance preserving network flow algorithm for multi-object tracking.(2)This thesis proposes to incorporate the tempo-ral coherence to enrich target representations and proposes a recurrent memory activation network for visual tracking.(3)This thesis proposes a reciproca-tive learning algorithm to constrain attention.The deep attentive tracker based on this algorithm can avoid the local attention problem which is caused by the overfitting.(4)This thesis proposes an objectness transfer network for visual tracking.This network incorporates objectness transfer learning to avoid the problem of forgetting the objectness information due to the overfitting.The above algorithms are executed on the standard benchmark datasets for perfor-mance evaluations.Extensive experimental results verify the effectiveness and robustness of the above algorithms.This thesis aims to research target domain adaptation in visual tracking,proposes to use spatial structural information and temporal coherence to enhance target representations,proposes to constrain at-tention and transfer objectness to improve the robustness of models over a long temporal span.The proposed algorithms effectively facilitate target domain adaptation,improve tracking performances,and promote the development of start-of-the-art trackers.