Research Of Visual Object Tracking Methods Under Complex Scenes

As a hot topic in the field of computer vision,object tracking has various applications such as smart city,intelligent transportation systems and national defense,etc.Although much progress has been obtained in recent years,object tracking remains a challenging problem in complex scenes.In complex tracking scenes,large appearance changes of the target object and the occlusion problem are key factors to limit the tracking performance.Based on the above analysis,this thesis proposes a series of object tracking algorithms to deal with complex appearance changes of the target object and address the occlusion issue.The main contributions of this thesis are summarized as follows:1)To deeply mine the appearance characteristics of different local patches for accurately tracking the target object,this thesis proposes a novel local sparse appearance model.The proposed method divides all local patches of a candidate target into stable patches,valid patches,and invalid patches,and assigns different weights to all local patches.First,a local sparse score is proposed to effectively mine stable patches.To alleviate the effects of background patches,this thesis designs a discriminative local sparse coding method to decrease the weights of background patches.Then,by utilizing the locality of sparse representation,a local linear regression method is developed to distinguish the valid patches from the invalid patches.Finally,to make the weight assign operations more reasonable,this thesis proposes a weight shrinkage method to determine the weights for valid patches.By mining different kinds of local patches and assigning weights to them,the proposed method can effectively capture large appearance changes of the target object and deal with the occlusion problem.Experiment results on several public visual tracking benchmarks demonstrate that the proposed local sparse appearance model can achieve state-of-the-art tracking performance in complex tracking scenes.2)To effectively employing multiple visual features for better modeling the appearance of the target object,this thesis proposes a dynamic feature fusion method with correlation filters framework.As multiple features generally represent the target object from different views and have different importance in complex tracking scenes,the proposed algorithm dynamically fuses gradient and color features to successfully localize the target object.First,the proposed method learns two independent correlation filters for gradient and color features,and adaptively adjusts the weights of them to deal with significant appearance changes of the target object.The weights are decided by the consistency of the final tracking result and the predicted results obtained from two correlation filters.Then,this thesis proposes a failure detection scheme to alleviate the model drift issue caused by undesirable model update.If a tracking result is identified as a failed case,a re-detection operation is performed to accurately localize the target object.Numerous experiments demonstrate that the proposed feature fusion method can effectively capture large appearance changes of the target object and address the occlusion issue to obtain better tracking performance.3)Considering the target object generally has several different appearance patterns during tracking,this thesis proposes a novel multi-pattern correlation tracking method for object tracking.First,the proposed method employs multiple correlation filters to model different appearance patterns of the target object,and each correlation filter represents one specific appearance pattern.Then,a two stage selection algorithm is designed to select a suitable correlation filter to localize the target object.The reliable score is utilized to exclude unreliable filters in the first stage while the matching score is used to choose an ideal filter for detection in the second stage.Finally,an online evaluation algorithm is designed to learn multiple correlation filters to model the appearance of the target object.By using multiple filters for tracking,the tracking model has better diversity and robustness to deal with complex appearance changes of the target object and the occlusion issue.Experiment results on several visual tracking benchmarks prove that the proposed multi-pattern correlation tracking method can achieve state-of-the-art tracking performance in dynamic tracking scenes.4)As ensemble learning can greatly improve the diversity and generalization power of the tracking model,this thesis proposes a novel ensemble correlation tracking method based on the correlation filters framework for object tracking.First,the proposed ECT tracker employs multiple correlation filters to deeply model different appearance patterns of the target object,and performs ensemble operations with these filters to accurately localize the target object during the tracking process.Then,to effectively learn different filters for ensemble tracking,this thesis designs a novel backtracking algorithm to generate filters for modeling the appearance of the target object.Finally,an online weight assign algorithm is proposed to assign different weights for multiple correlation filters with consideration of historical appearance information and recent appearance changes of the target object.By taking advantage of ensemble learning and multiple correlation filters,the proposed method can successfully capture large appearance variations of the target object and handle occlusion.Extensive experiments on several visual tracking benchmarks demonstrate that the proposed ensemble correlation tracking method can obtain state-of-the-art tracking performance.In summary,from the perspectives of local sparse representation,dynamic multiple feature fusion,multi-pattern learning and ensemble learning,this thesis deeply analyzes the difficulties and challenges,introduces some concepts,designs several effective tracking algorithms and achieves impressive tracking performance.Therefore,this thesis will be a good reference for future research in this field.