基于核相关滤波算法的层级遮挡检测目标跟踪

Hierarchical occlusion detection tracking based on kernel correlation filters

基于核相关滤波的快速跟踪算法在与除机器学习之外的大部分传统算法相比,该算法在时效性和跟踪精度方面优势明显。但在目标发生严重遮挡或全部遮挡时,KCF的跟踪精度会明显下降。通过对其分析发现,当目标发生遮挡时,分类器会引入错误信息,从而导致跟踪精度下降甚至跟踪失败。因此在KCF算法的基础上引入遮挡检测机制,当跟踪器判断出目标发生遮挡时,停止分类器参数的更新;同时利用局部二值模式（local binary pattern,LBP）特征进行层级遮挡检测可以有效地区分目标变形和遮挡。对于满足线性运动的目标,利用卡尔曼滤波器对目标进行位置预测以解决目标发生遮挡后的定位问题。实验证明,提出的算法对具有遮挡和变形情况的目标跟踪具有较高的鲁棒性。

Compared with most traditional tracking methods,the method of tracking via Kernel correlation filters（KCF）has significant advantage in real time performance and tracking accuracy.However,when the object is seriously or completely occluded,the tracking accuracy will decrease obviously.Through the analysis of KCF,we found that when the occlusion occurred,the classifier would introduce erroneous information,resulting in poor tracking accuracy,or even tracking failure.Based on the KCF framework,it introduced the occlusion detection mechanism to stop the updating of the classifier parameters when the occlusion occurred,and using（local binary pattern,LBP）feature for hierarchical occlusion detection can effectively distinguish target deformation and occlusion.As to the linear motion object,the Kalman filter was used to predict the position of the object in order to solve the problem of occlusion.The results show that the proposed algorithm is robust in object tracking under occlusion and deformation.