Adaptively learning background-aware correlation filter for visual tracking

Abstract

Correlation filter (CF) trackers have received more and more attention due to their excellent performance while maintaining high frame rates. However, the limited context information might limit the performance of CF trackers as the presence of background effects in or around target bounding box will corrupt CF learning. In this paper, toward improving background-aware CF trackers, we propose a general algorithm that adaptively incorporates background contexts in CF learning to suppress the distractors effectively. Comparing with existing background-aware CF trackers, our approach can adaptively explore background distractors by employing their correlations to the target object which makes our tracker more effective and efficient. Experimental results on large-scale benchmark dataset demonstrate the effectiveness and efficiency of the proposed approach against recent CF trackers.