Research On Robust Visual Tracking Algorithm Based On Correlation Filters

Visual object tracking has always been one of the most important research topics in computer vision with many applications,such as human computer interaction,video surveillance,autonomous driving and so on.It aims to locate a target,only the initial location of which is given in a video sequence.In general,existing tracking approaches utilize either generative methods based on sparse representation,or discriminative methods based on correlation filters to learn the target appearance model.Despite their efficiency and promising tracking performance,traditional tracking approaches based on correlation filters are usually equipped with only single representation and decision model.Such modeling methods make trackers unable to recourse to target rotation and deformations,and easily lead to tracking drift.To tackle the above issue,this thesis proposes a dynamic decision fusion strategy which utilizes multiple representation models to comprehensively depict the target appearance.The proposed approach adaptively adjusts the decision fusion and model updating schemes in different conditions,which helps the tracker to be more robust to both occlusions and color changes.Besides,considering that conventional correlation filter-based trackers often lack strong discriminative ability due to their shallow structure,a novel end-to-end compact architecture with multiple correlation filters is proposed for visual tracking.The proposed network extends the dimension of traditional correlation filter to make the full use of deep convolutional multi-channel features.It also introduces the densely connected structure and Dirac weights into the visual tracking community,and then incorporates the model with nested filters into the deep learning framework,which effectively learns the data-specific target appearance.By doing so,the proposed tracking approach significantly enhances the discriminative ability and robustness against challenging factors such as target deformations.Moreover,to alleviate problems caused by target scale variations,this thesis investigates a scale estimation strategy with kernel method,which exploits the nonlinear relationship of scale samples with different sizes.Qualitative and quantitative evaluations on two widely used benchmarks demonstrate that the proposed tracking approaches achieve outstanding performance in terms of precision and success rate,and are robust against target appearance variations.The proposed methods also significantly outperform existing state-of-the-art trackers according to the experimental results.