A Study Of Robust Visual Tracking Method Based On Sparse Representation

Visual tracking is an important research topic in the field of computer vision. It has been widely used in defense and civilian areas. However, many factors, such as the cluttered background, the image blur caused by camera dithering, illumination variations, object occlusions, scale and view changes etc., increase the difficulty of modeling the object appearance in video images. These factors usually lead to the drifting of the tracker. Therefore, visual tracking has long been a difficult research topic in the field of computer vision. Based on the systematically analysis of existing visual tracking methods, this paper mainly collaborates the color attention and the sparse representation to improve the robustness and speed of visual tracking. The major contributions are summarized as follows:(1) For the drifting problem, we combine the sparse representation and color attention mechanism to propose a robust tracking algorithm based on the generative model. This method utilizes a dense grid to crop the target image into patches, which maintain the structural information of the object. The reconstruction error of each patch based on sparse representation determines whether the patch is blocked. Since color attention has good discriminability in distinguishing the target from non-targets, a robust similarity measurement is designed to track the object based on both color names and sparse representation of image patches. The object template is updated by linearly combining the appearance of the object in the initial frame and the current frame, which maintains the original appearance and the incremental variation of the object at the same time. Experimental results show that the proposed method is more robust compared with several state-of-the-art algorithms.(2) To solve the problems of heavy computation load and low speed of tracking algorithms, a fast tracking method combining compressed sensing and Bayesian classification is proposed. Since Gaussian random compressed measurement matrix can reconstruct the original signal with high probability, our method utilizes this character to extract the compressed Haar-like features of the target and the background in the initial frame to train a discriminative classifier. The object is tracked with the previously trained Bayesian classifier, which detects samples with the “coarse to fine” strategy and regards the sample with the maximum response as the tracked object. The classifier isupdated according to the historical results. Experimental results show that this method runs in real-time and its accuracy is superior over traditional methods.The proposed algorithms in this paper can effectively handle illumination variations and scale changes, object occlusions, clutter backgrounds etc., which enrich the algorithms of robust tracking.