融合全局与局部哈希特征的目标实时跟踪

Fusion of global and local Hash feature for real-time object tracking

为了缓解光照变化、部分遮挡和背景干扰等场景对于目标跟踪算法的影响并实现实时跟踪,提出了基于全局和局部哈希特征的建模方法.这种建模方法能够有效地提取目标全局和局部特征并进行融合,从而获得更加鲁棒的目标表达.为了提升算法的运行效率,采用倒金字塔候选框搜索策略,以去除大量的冗余候选框;另外,使用汉明距离来度量候选框与目标框之间的相似性,并结合哈希特征和汉明距离的特点,将位操作嵌入到了哈希特征的提取与存储及汉明距离计算的过程中.最后,通过在一些复杂场景中与多个经典跟踪算法进行对比实验,验证了本文算法在稳定性、鲁棒性和时效性等方面的优势.

In order to mitigate the effect of complex scenes such as illumination change,partial occlusion,background interference on target tracking algorithm and realize the real-time object tracking simultaneously,the global and local Hash feature-based modeling method was proposed,which could effectively extract the object＇s global and local features,and then fused to obtain more robust object appearance.For the purpose of enhancing efficiency of our tracking algorithm,an inverted pyramid candidate boxes searching strategy was employed to significantly reduce redundancy.In addition,Hamming distance was used as the similarity measure between candidate box and object box.Combined with the characteristics of Hash feature and Hamming distance,some bitwise operations were embedded into the whole process of the extraction and storage of Hash features and the Hamming distance calculation.Finally,a number of comparison experiments against several classical tracking algorithms were tested in some complicated scenes,which demonstrated the superiority of the proposed algorithm in terms of stability,robustness,and effectiveness.