Robust object tracking via diverse templates

Robust object tracking is a challenging task in computer vision. Since the appearance of the target changes frequently, how to build and update the appearance model is crucial. In this paper, to better represent the object dynamically, we propose a robust object tracker based on diverse templates. First, we construct diverse multiple templates using the determinantal point process algorithm adaptively, which efficiently detects the most diverse subset of a set. Second, a patch-matching method is employed to propagate every template density to the next frame, and a voting map for each template is constructed by all matching patches. Third, a weighted Bayesian filter framework aggregates all voting maps to optimize target state. Finally, in order to maintain the diversity of multiple templates, we dynamically add, remove and replace the target from templates. Experimental results prove that the proposed method outperforms state-of-the-art tracking algorithms significantly in terms of center position errors and success rates.