Research On Visual Object Tracking Method Based On Sparse Representation

Object tracking in video sequences is the research hotspot in computer vision field. It combines the research findings in many fields, such as machine learning, pattern recognition, and has been widely applied in video surveillance, intelligent transportation and modern military. So the researchers at home and abroad carry out a lot of researches on object tracking and present many effective tracking algorithms. However, it is still a challenging problem to track an appearance-changing target in complex natural scenes. The common tracking difficulties include scene illumination change, target scale change, part occlusions, target's non-rigid deformation, as well as posture change etc. And these difficulties cause the nonlinear changes of the target appearance, which make the object tracking more complicated. To address above problems, this dissertation does a deep research to improve the veracity and robustness of tracking algorithm. The main contributions of this dissertation are shown as follows:Firstly, a two-step tracking algorithm based on particle pre-filter is proposed. To the problem that most related algorithms need to build observation models for all samples, this algorithm models the particle pre-filter as sparse representation of holistic templates in the first phase. By the particle pre-filter, it gets rid of these particles which are far away from target's real state. As a result, the number of the samples is decreased and the efficiency of the algorithm is enhanced. In order to alleviate drifting during tracking, the proposed algorithm uses both the target initial state and current state as observation reference in the second phase, which improves the accuracy of the tracking results. The experimental results in multiple test sequences demonstrate the effectiveness of the proposed method.Secondly, a tracking algorithm based on local discriminative sparse representation is proposed. In order to enhance the discriminative ability between the target and background, local background images are added into dictionary learning as negative samples, which makes the dictionary consider both expression capability and discriminative strength. As most tracking methods based on local sparse model lack target's global information, the target is modeled as a sparse coding histogram of specified local images. These local images are those compose the dictionary. Thus, the model effectively combines the local features and global structure of the target. To improve the veracity of observation model, a similarity coefficient based on target structural information is designed. Besides, the target model is updated automatically to adapt to target's appearance changes. The experimental results show that the proposed method is able to handle most tracking difficulties, and has a higher tracking precision.Thirdly, a tracking algorithm based on weighted structural sparse representation is proposed. The target is described by structural sparse representation to make full use of structural information between local images, which can avoid the model degradation. The background information is added into the structural dictionary to enhance the discriminative capability of target model. In order to improve robustness of the model, the importance weight is assigned to each local image according to its role in target expression. Then, the target is modeled as weighted structural sparse model. In addition, to reduce the influence of the local occlusion, an occlusion detection module is added into observation modeling. The experimental results in multiple test sequences show that the weighted structural sparse model has a good adaptability to target apparent changes. And the proposed method shows better robustness and accuracy in tracking.Lastly, a multi-task tracking algorithm based on local joint sparse representation is proposed. To address the issue that the structural information between samples are underused by related algorithms, the sparse code of each local image is viewed as an individual task and all the local images of samples are coded jointly in multi-task learning framework. This joint sparse representation is able to extract thoroughly the structural information within sample and the structural information between samples, which not only enhances the expression capability of the model but also improves the efficiency of the algorithm. Furthermore, a function combines global and local similarity is presented to measure the similarity between target image and sample, which can improves the reliability of observation model. The experiments demonstrate the proposed method obtains a more accurate tracking result.In conclusion, object tracking in sequence is realized in this paper based on holistic sparse representation, local discriminative sparse representation, weighted structural sparse representation and local joint sparse representation, respectively. And the tracking performances of these four algorithms are tested in experiments. It is demonstrated by the experiment results that the proposed tracking algorithms are able to improve the veracity and robustness of tracking gradually.