Self-expressive tracking

Yao Sui, Xiaolin Zhao, Shunli Zhang, Xin Yu, Sicong Zhao, Li Zhang

Abstract: Highlights • An observation of the inherent relationship and structure among the candidates. • A structured representation with naturally discriminative power for target description. • An evaluation criterion of the likelihood of the candidates belonging to the target. • An update strategy with adaptive update rates for appearance model maintenance. Abstract Target representation is critical to visual tracking. A good representation usually exploits some inherent relationship and structures among the observed targets, the candidates, or both. In this work, we observe that the candidates are strongly correlated to each other and exhibit obvious clustering structure, when they are densely sampled around possible target locations. Thus, we propose a Self-Expressive Tracking (SET) algorithm based on an accurate representation with good discriminative performance. The interrelationship and the clustering structure among the observed targets and the candidates are exploited by using a self-expressive scheme with a low-rank constraint. Further, we design a discriminative criterion of the likelihood for target location, which simultaneously considers the target, background and representation errors. To appropriately capture the appearance changes of the target, we develop an update strategy that adaptively switches different update rates during tracking. Extensive experiments demonstrate that our tracking algorithm outperforms many other state-of-the-art methods.