Research On Object Tracking Algorithm Based On Meta Learning And Occlusion Handling

Visual object tracking is an important research area of computer vision and has been widely used in video surveillance,unmanned driving,and other fields.Due to the internal changes of the target,e.g.,scale changes,fast motion and external influences of the scene,e.g.,occlusion,object tracking is still a extremely challenging problem.In recent years,with the development of deep learning,scholars have launched a lot of research and explorations,and the trackers based on deep learning have achieved good performance on multiple standard datasets.However,there are still several problems that have not been solved: the fast and robust learning problems for features,classifiers,and other models in object tracking algorithms;the problem of occlusion detection and handling.Therefore,they may fail when occlusion happens or drastic appearance changes occur during the tracking phase.To address the above issues,this dissertation based on meta-learning to quickly improve the fast and robust learning problem of the object tracking algorithm.And a data-driven occlusion detection method is proposed to further improve some important problems in long-term object tracking.We conduct our research in the following two aspects:(1)We propose a novel localization-aware meta tracker guided with adversarial features.First of all,we design a novel Intersection over Union guided method to effectively balance the problem of classification and localization accuracy.To further improve the robustness of the classifier,we creatively use adversarial features during offline training phase.Those adversarial features can effectively guide the classifier in learning how to better deal with the situation where the discriminative features are occluded or changed.Finally,benefiting from meta learning,the above two components only need to perform one iterative update on the first frame and they can perform well on the tracking sceneries.(2)We propose a data-driven occlusion detection and model coordination object tracking algorithm.This algorithm further proposes a specific solution for severe occlusion or target disappearance in long-term tracking.The occlusion detection module is the core component of the proposed tracker.This module learns in a data-driven manner to ensure its generalization ability.The output of this module is then used to coordinate key decisions such as selecting the appropriate target appearance model,collecting online samples,and deciding when to re-detection.We have performed experiments on the OTB-2015,VOT-2016,and VOT-2018 datasets and have made significant improvements in accuracy and robustness.In addition,the second algorithm is also tested on the VOT-2018 LTB35 and GOT-10 K datasets.Experimental results show that the algorithm also performs well on long-term tracking datasets and can effectively deal with occlusion and target disappearance.