Improved Compressive Tracker via Local Context Learning

This paper presents an improved compressive tracking algorithm via local context learning. There are two primary problems with compressive tracker, occlusion and drifting, both of which are solved by introducing a local context model. The local context information, which are often discarded in generative methods, provides specific information about the configure of a scene. The spatial relationships between the object and its surrounding backgrounds help relocate the object when it undergoes significant appearance changes. Our approach makes full use of context information and models the statistical correlation between the low-level features from the object and its surrounding backgrounds. The tracking problem is formulated by maximizing an object location likelihood function, and obtaining the best object location with the combination of compressive tracker and local context model. Experimentally, we show that our algorithm can greatly improve compressive tracker both in terms of robustness and accuracy and outperform state-of-art trackers on various benchmark videos.