Visual tracking via multi-experts combined with average hash model

Model-free online object tracking is an important research topic of a wide range of applications in computer vision. A main challenge for object tracking is the model drift problem. In this paper, we proposed a multi-expert selection tracking algorithm that can not only prevent adding bad examples to object model but also can correct the effect of bad updates even if the bad examples are involved. Multi-expert ensemble is constructed of a base tracker and its former snapshots. We choose compressive tracker as our base tracker and introduce an efficient mechanism based on Hash algorithm to prevent bad model updates. Extensive experimental results show that the proposed algorithm performs favorably against state-of-the-art methods. In addition, experiment results on a newly collected dataset with challenging situations demonstrate the better performance of our method.