Research Of Object Tracking Based On Dictionary Encoding Model

In recent years, object tracking has been active in the intelligent visual field, and is cru-cial in many practical life and industrial applications such as airport, station and shopping mall. Due to the challenging factors like occlusion, deformation, rotation, scale and background clut-ter, developing an effective and efficient tracker is still a difficult task. This paper summarizes the research purpose, status and difficulties of object tracking, introduces some related state-of-the-art tracking methods and dictionary based coding models, and analyzes the advantages and disadvantages of these algorithms. On this basis, we propose two kinds of object tracking meth-ods based on the dictionary coding, i.e., object tracking with discriminative dictionary based appearance learning and object tracking based on the dual group structure.This paper first analyzes the limitations of the previous related works, and then proposes a discriminative dictionary learning model. We cast the appearance learning and target match-ing into a single objective framework, and jointly trains the dictionary and classifier, where the interaction between them facilitates the tracking performance. In addition, we adopt the differ-ential tracking strategy to optimize the motion variation of the target directly, which makes the bounding box move purposely and avoids the processing of large number of useless candidates, improving the running efficiency. Moreover, this paper also presents an online updating mecha-nism to capture the appearance change during the tracking process, and weakens the impact of occlusion and inaccurate location.This paper also analyzes the deficiencies of the original sparse representation method, and formulates it in the group level. We exploit the structural information among both the dictionary atoms and candidate samples, and jointly model the commonality and characteristic at the same time. The clustering of samples enables that all data could be encoded simultaneously, while the clustering of atoms encourages the grouping effect. Further, we give an efficient optimization procedure to obtain the solution of the proposed objective function. This solution contains a matrix thresholding and a vector threshoulding process, naturally yielding the expected inter-group and intra-group sparsity patterns. In order to explicitly handle the occlusion, we also add a set of trivial templates to model the unknown noise.We evaluate the two proposed trackers on large amount of challenging image sequences, including the comparison with state-of-the-art methods, and the comparison between these two trackers. The experimental results demonstrate that our methods achieves more satisfactory and excellent performance against other competing methods.