Online hash tracking with spatio-temporal saliency auxiliary

In this paper, we propose an online hashing tracking method with a further exploitation of spatio-temporal saliency for template sampling. Specifically, spatio-temporal saliency is firstly explored to make the sampled templates contain true object templates as much as possible. Then, different from the previous batch modes for hashing, the hashing function in this work is online learned by new pairs of collected templates received sequentially, in which the relationship between the positive templates and negative templates can be appropriately preserved that is more useful for visual tracking. With the hash coding for templates, the between-frame matching can be efficiently conducted. Besides, this work further builds a positive template pool as a memory buffer for object depiction, in which representative truly positive target templates are gathered and utilized to restrain the degradation of the appearance model due to the error accommodation in online hashing. Extensive experiments demonstrate that our tracker performs favorably against the state-of-the-art ones.