Research On Sparse Representation And Spatio-temporal Based Tracking Approach

Moving target tracking in video is an important research field of computer vision and is key technology to video content analysis, it plays an important role in visual tasks like movement analysis, pattern recognition, video surveillance, human-computer interaction and traffic management and so on, and the main task of it is to locate the target and obtain the target trajectory by analyzing and estimating the states of target in the Sequence of video images, and provide basic information for more advanced video processing. target tracking has a very wide range of applications in the civilian, military and Industry.Bayesian methods could be used to achieve coordination and harmonization between prior knowledge and observation. The observation model and motion model in many tracking algorithm are constructed based on Bayesian tracing framework. In tracking,robust of algorithm could be improved by constructing Appearance Model. Appearance model can be divided into generative model and discriminative model according to the construction method. To complete tracking task, we proposed a tracking algorithm based on Bayesian tracking framework, in this algorithm, discriminative model and generative model are combined to complete the tracking task. The discriminative model is constructed by clustering the super-pixel that have be segmented in the training images during training phase; the generative model is constructed by calculating the sparsity histogram of target templates in first frame. In tracking, super-pixel-based confidence map is obtained, and the confidence values of candidates are sampled and calculated; the Sparse representation coefficient of local patches which collected from candidates are calculated to obtain the Sparsity histogram of candidates, then the similarity between the candidates and templates is calculated. The observation model is constructed on the basis of the two parameter that are obtained by the above process. The posteriori estimates of are calculated according to the observation model and motion model, and the tracking result is obtained at last. Further more, online updating of the two appearance model is kept independently in order to improve the adaptive of them. The target could be distinguished from background by discriminative model, and spatial information is included into generative model. The experimental results and evaluations demonstrate that the application of SPS algorithm can obtain accurate and robust track result with theappearance variation of target object.Tracking with temporal-spatial context could extend the resource which could be exploited in practical problem. When the appearance of target had changed a lot due to occlusion?changes of posture and other factors, the temporal-spatial context could still be used for target location. We proposed an improved algorithm based on temporal-spatial context. First, deconvolution method is used to learn spatial context model in the areas of target and background around it, then the spatial context model will be exploit to update the temporal-spatial context in the next frame image. Saliency feature is introduced during the model learning process, and as the feature description to the context area,by doing this the complexity of context area would be reduced, and the accuracy of target detection would also be developed. The saliency of target are extracted, and particle filter and saliency histogram matching are used to predict and modify the tracking result.