Online Robust Dictionary Learning For Visual Tracking Via Part-Based Sparsity Model

Visual tracking is an important issue of computer vision. It has been widely used in many applications such as human-computer interaction, cognitive systems and surveillance. Visual tracking has been investigated extensively, but is still an open problem with chanllenges such as occlusion, deformation in real world enviroment. Based on part-based sparsity model, this thesis proposes an algorithm based on online robust dictionary learning for visual tracking.Firstly, a robust tracking method is proposed by using part-based sparsity model. In this model, one object is represented by image patches. The candidates of these patches are sparsely represented in the space spanned by the patch templates and trivial templates. The part-based method takes the spatial information of each patch into consideration, where the vote maps of the multiple patches are used. With the combined vote map, the best estimation of the target location can be acquired.Secondly, the elements of the dictionary can update dynamically according to their weights in sparse representaion model. However, this strategy cann’t guarantee the target sparsely represented by the updated dictionary. In this thesis, dictionary learning is used for template update strategy, which uses the recent tracking results as the trainning set. In this way, the dictionary can be learnt adaptively for the good descriptiveness of the target.Finally, a number of matrix operations of the dictionary learning algorithm have been designed to run on the GPU in this thesis. In order to make full use of both CPU and GPU, a popular computing architecture CUDA is used and sensible resource allocation is achieved to accomplish the computation.On the sequences of infrared object on the sea and public benchmarks sequences, the abundant results of experiments launched demonstrate that our proposed visual tracking method outperforms many existing state-of-the-art algorithms. The results also demonstrate that the speed increased from several times to dozens of times according to CPU by using the GPU to accomplish the computation.