A novel joint tracker based on occlusion detection

A challenging problem that needs to be faced in visual object tracking is occlusion, which includes partial occlusion, complete occlusion and blur. Some tracking methods use motion models to predict the location of the occluded objects, and others use patches or parts of the object as a tracking unit to deal with occlusion. These methods seem to solve the occlusion problem indirectly, however, avoiding its negative influence. In this paper, we propose a novel method to solve the occlusion problem directly. First, we propose a new mechanism which can predict occlusion accurately and sensitively with MIL and SVM classifiers. Second, we combine the discriminative method and the generative method in a joint-probability model and use the occlusion information to adjust the weights of the methods, which are complementary. Third, we propose a classification-based template updating method, in which we divide the templates into two groups according to occlusion information and use opposite probability distributions to update the two groups. The experiment results demonstrate that our method is effective and outperforms the state-of-the-art approaches on several benchmark datasets.