Tree-Structured Correlation Filters for Robust Visual Tracking

Abstract

In recent years, correlation filter based trackers have significantly advanced the state-of-the-art in visual tracking. However, most existing correlation filter based tracking algorithms update target object model assuming that target appearances change smoothly over time. This assumption may not be appropriate for handling more challenging situations such as occlusion, deformation, illumination variation, and abrupt motion, which may break temporal smoothness assumption. To address these issues, in this paper, we propose a novel treestructured correlation filters (TCF) for diverse target object appearance modeling, where multiple correlation filters collaborate to estimate target states and determine the desirable paths for online model updates in the tree. As a result, the proposed TCF tracker has the advantages of both CNNs and correlation filter based trackers. Furthermore, our TCF tracker can preserve model reliability by smoothly updating deep correlation filters along the path in the tree, and make the learned appearance models sufficiently diverse and discriminative. Extensive experimental results on two challenging benchmark datasets demonstrate that the proposed TCF tracking algorithm performs favorably against the state-of-the-art trackers.