Optimized compressive tracking in co-training framework

As visual tracking algorithms based on traditional co-training framework are characterized by poor robustness in complex environment, an optimized compressive tracking algorithm in a novel co-training criterion is proposed. Firstly, the spatial layout information and the online feature selection technique based on maximizing entropy energy are used to improve the discriminative capacity of compressive sense classifier, and two independent classifiers are constructed by structural compressive features selected from the gray intensity space and the local binary pattern space respectively. Secondly, on the basis of the classifiers collaborative strategy that is acquired by calculating the confidence score distribution entropy of the candidate samples, complementary features can be adaptive fused, which reinforces the robustness of tracking results. Thirdly, as assistant of the cascaded Histograms of Orientation Gradient (HOG) classifier, the collaborative appearance model is updated with accuracy by a novel co-training criterion with sample selecting ability, which solves the updating error of co-training accumulation problem. Comparative experiment results on extensive challenging sequences demonstrate that the proposed algorithm is of better performance than other similar tracking algorithms. © 2016, Science Press. All right reserved.