Visual object tracking based on appearance model selection

Abstract

Occlusion and appearance variation are two common challenges in visual object tracking. Existing methods may not distinguish occlusion from large appearance variation during appearance model updating, as both of them may cause large appearance transformation inside the bounding box. In this paper, we propose an appearance model selection (AMS) based visual tracking algorithm. In the proposed method, the appearance model will be duplicated and one of them stops updating when there is large appearance change inside the bounding box, led by either allowed appearance variation or unexpected occlusion. According to the appearance information of an incoming video frame, the proposed method will choose the best appearance model in the model pool by the model selection mechanism. The proposed method can track the visual targets with appearance variation accurately and avoid error accumulation from occlusion at the same time. Experimental results demonstrate that the proposed AMS tracking method outperforms other existing related ones on four video database.