Multi-pattern correlation tracking

In this paper, we propose a novel multi-pattern correlation tracker (MPCT) which deeply models the appearance of the target object for robust tracking. Specifically, multiple correlation filters are learned to capture different appearance patterns of the target object during the tracking process and each filter represents one specific appearance pattern. With the proposed reliable and matching score, a two stage selection algorithm is developed to select a suitable correlation filter to localize the target object. To effectively obtain different filters, we design an online evaluation algorithm to generate filters for different appearance patterns. By taking advantage of multiple filters to model different appearance patterns, the proposed MPCT tracker can not only capture dynamic appearance changes under complex scenes but also deal with severe occlusion and model drift problems to achieve better tracking performance. Extensive experimental results prove that the proposed tracking algorithm performs superiorly against several state-of-the-art tracking methods on challenging tracking benchmarks.