Research Of Object Tracking Via Online Dictionary Learning Model Based On Sparse Representation

Visual tracking has been extensively studied because of its importance in practical applications, such as visual surveillance, human computer interaction, traffic monitoring and so on. Despite extensive research in this topic with demonstrated success, it is still a very challenging task to build a robust and efficient tracking system to deal with various appearance changes caused by pose variation, illumination changes, shape deformation and abrupt motion. In this thesis, we address these challenging factors by building several robust appearance models for visual tracking.1?In order to improve the tracking speed,a model based on the local appearance of interest point is proposed. This appearance model consists of two parts: target representation based on the target interest point and a robust matching criterion based on sparse representation. In the target representation, the target dictionary is obtained from the patches around target interest points. Centered on each candidate point of interest from the candidate window of the current frame, and particle filter algorithm is utilized to extract different direction and scale of the candidate target block. In the robust matching criterion, the correspondence between target and candidate points is obtained via sparse representation method, and a robust matching criterion is proposed to screen candidate target block. The target is localized by measuring the displacement of these interest points. The reliable candidate patches are used for updating the target dictionary. In addition, we propose an online dictionary learning algorithm for updating the object templates, so that each learned template can capture a distinctive aspect of the tracked object. Another appealing property of this approach is that it can automatically reject the occlusion and cluttered background in a principled way. Thereby, the probability of drift of target template is reduced.2?In order to enhance the capacity of tracking algorithm to disturbance, firstly, a local adaptive weighting algorithm is put forward to increase degree of differentiation between the good target area and bad target area. Secondly, structure sparse representation algorithm, using the candidate target which contains rich target and background features to build a complete dictionary to reconstruct the target template under the condition of the same dimension target template better sparse coefficient can be obtained, is proposed. Utilizing the internal structural of sparse coefficient matrix to statistics the similarity between the target template and every candidate target, can significantly improve the tracking precision. In the end, an online discriminative double dictionary learning algorithm for updating the object templates so that each learned template can capture a distinctive aspect of the tracked object is proposed. Another appealing property of this approach is that it can automatically detect and reject the occlusion and cluttered background using the discriminative function. To make the updated target template can accurately reflect the changes.In general, the present thesis studies deeply the sparse representation target tracking algorithm based on particle filter framework with the purpose of enhancing robustness, accuracy and instantaneity of the object tracking algorithm.