Research On Correlation Filter-based Real-time Object Tracking Algorithms

As an important part of the research in computer vision,object tracking has made a lot of significant progress in recent years,while numerous tracking solutions have been successfully embedded into the industrial systems.Correlation filter-based tracking method has shown great potentials in recent years.The tracking framework in this method utilizes the characteristics of cyclic matrix,so that it can effectively solve the intractable sampling problem and localization accuracy in traditional methods.However,the unreal boundary in cyclic samples latently under-limit the performance of the algorithm,the tracker will degrade drastically when the targets encounter with deformation and occlusion.In this paper,the correlation filter-based tracking framework is separated into two aspects:the training stage and the locating stage and our main works can be summarized as follows:At first,the working mechanism of correlation filter-based tracker will be discribed in details.To improve the performance of correlation filter,a dual filter constraint is introduced to improve the over-fitting of unreal samples,by which the improved filter can quickly converge to a suboptimum by dual iterations.Feasibility analysis about our proposed method is carried out by two sets of ablation experiments,the results of which is able to show that the effectiveness of proposed method in suppressing the boundary effect.In the second chapter,the superiority of training with structured output are compared and analysed.Considering about the class imbalance problem of samples in correlation filter-based method,a margin maximized constraint combining with cyclic samples is proposed.To figure out the hard optimization problem,this paper proposes a dual term with margin loss based on the one in last chapter.The advanced correlation filter can be solved by several alternating iterations.The experimental results show that when the proportion of target in search area is within a certain range,the proposed method could further enhance the tracking performance of correlation filter.Finally,aiming at the tracking drift problem in online-update tracking method,a re-detection mechanism based on convolutional feature responses is proposed.In order to efficiently improve the adaptability of the filter to target scale change after adding convolutional feature extraction,an object scale search scheme is proposed,which can effectively approximate the results of tracker with scale pyramid strategy.Considering the real-time requirement of our algorithm,an online dimensionality reduction approach to the convolutional features is carried out during the experiments.The experimental results show that the proposed algorithm can improve the tracking robustness of the correlation filter.All experiments in this paper have been evaluated by the standard dataset OTB100 with 100 video sequences.In order to verify the advancement of this algorithm,comparisons with several other state-of-the-arts are also made in this paper.Eventually,the results show that the advanced algorithms can sufficiently improve both tracking accuracy and robustness of the correlation filter.