Visual tracking has been an active research topic which has a wide application inmany computer vision tasks such as intelligence surveillance, vehicle navigation, humancomputer interaction, and so on. Sparse representation accords with people’s sense ofvision characteristic and reduces the dependence on feature selection. Therefore,it hasbeen received extensive attention in recent years.Sparse representation involves two significant components: the collection ofredundant dictionary, the analysis of sparse approximation coefficients. The coefficientscontain rich information including classification,noises, which can be utilized forimproving efficiency. But studies of this aspect still are very few so far.Introducing the background into the template set as the negative sample, therepresentative ones combine with positive samples as to build a new template set which isused to classify the sampling particles. The classification method filters out many sampleswhich are largely dissimilar to the target, avoids the following computationally expensivecost of matching. Experiment shows that it cut down the redundant calculations.The information captured in the approximation coefficients can be utilized forocclusion analysis. From the result of experimental analysis, the coefficient distribution isin a clutter when the target is occluded by an object. It can be applied for occlusiondetection. When an occlusion is detected, it prevents an improper addition to the templateset and reduces the dimensions of template according to the occluded area. Theexperimental results indicate that the proposed method somewhat reduces the algorithmcomplexity and minimizes the risk of introducing the occluded target into the template set. |