An Improved Discriminative Tracking Algorithm Based on Spatial Information and Model Updating

Abstract

Tracking object occluded partially is a difficult problem in video surveillance. Many previous tracking methods fail to track occlusion objects robustly. In this paper, we propose an improved discriminative tracking algorithm based on bag of patches to cope with the partial occlusion as well as drift. In the proposed method, the spatial information is introduced to build the object appearance model and construct the confidence map from three different aspects, which directly determine the ultimate tracking effect. In addition, the context information of the small image patches is also applied. In order to adapt to the variance of the environment, an online model updating strategy is proposed. Contrasting experimental results on several real world scenarios show that our proposed approach can handle partial occlusion and recover from drift. Comparing with four stat-of-the-art tracking methods, our proposed method has better tracking performance.