综合信息系统研究中心知识库

Robust visual tracking is an important and challenging problem due to various challenging factors and computational constraints. Recent studies have shown taking advantage of color information is a simple and effective way to improve correlation-based tracker performance. In this paper, we propose a 1-channel color feature called color ratio (CR) feature inspired by mean-shift-based tracking algorithms, which is more efficient and effective than currently widely used 10-channel color-naming features. We then concatenate 1-channel CR, 13-channel HOG and 1-channel gray together to get totally 15-channel features for efficient DCF tracking. During feature concatenation process, we find that weighting between different feature channels can improve the tracking performance notably. Finally, correlation-based responses and CR-based responses are fused to further boost tracker robustness. Experimental results demonstrate that our feature and fusion strategy can achieve superior performance while attaining real-time performance.