基于多特征融合的尺度自适应KCF目标跟踪算法

Scale-adaptive kernel correlation filtering tracking algorithm based on multi-feature fusion

摘要:

首先,对核相关滤波（KCF）目标跟踪算法进行了详细推导；然后，针对KCF算法提取单一特征，不能很好地表达目标的外观模型，提出将多种特征融合的方法，增加外观模型的可区分性.同时针对KCF算法不能自适应尺度变化的问题，引入一种尺度自适应变化方法.还对于KCF算法的固定更新率在目标被遮挡的情况下会学习到错误信息的问题，提出一种在线模型更新因子的方法；最后，通过实验对比结果表明，本文提出的算法跟踪精度更高，且对目标尺度发生较大变化和遮挡情况下的跟踪具有较强的鲁棒性.

Abstract:

In this paper, the algorithm of Kernel Correlated Filtering (KCF) target tracking is deduced in detail. Secondly, aiming at the single feature extracted by KCF algorithm, which can not express the appearance model of the target well, a method of fusing multiple features is proposed to increase the distinguishability of the appearance model. For the problem that KCF algorithm cannot adapt to scale change, a scale adaptive change method is introduced. To tackle the problem that the fixed update rate of KCF algorithm learns error information when the target is occluded, an online model updating factor method is proposed. Finally, the experiment results show that the proposed algorithm has higher tracking accuracy and stronger robustness in the case of large changes in target scale and occlusion.