一种核相关滤波器的多尺度目标跟踪方法

A Multi-Scale Target Tracking Method Based on Kernelized Correlation Filter

针对传统的核相关滤波器(KCF)跟踪算法无法解决目标尺度变化并导致目标丢失的问题,文中提出了一种尺度自适应核相关滤波分类器。首先利用核岭回归方法对由循环移位得到的大量目标图像样本进行训练得到核相关滤波分类器;然后建立多尺度待检测图像集,通过相关滤波器求取最大响应以得到当前目标位置与尺度信息;最后利用新目标图像为训练样本在线更新目标的尺度和外观信息。为了验证算法的有效性,在数据集中选取10组测试序列进行验证,并同时与KCF、DSST、CN等优秀算法进行对比。实验结果表明,所提算法能更好的适应尺度变化的跟踪,且跟踪精度有所提升。

Aiming at the problem that traditional Kernelized CoiTelation Filter （KCF） tracking algorithm failedto solve the target scale change and target loss, a scale adaptive kernel correlation filter classifier was proposed in thestudy. Firstly, a large number of target image samples obtained by cyclic shift were trained by using kernel ridge re-gression to obtain a kernel - dependent filter classifier. Then, the multi - scale image set to he detected was estab-lished, and the cunent target position and scale inibmlation were obtained by the COiTelation filter. Finally, the newtarget images were used to update the target＇s scale and appearance intbmlation online. In order to verity the validityof the algorithm, 10 sets of test sequences were selected for verification in datasets. Besides, its peromlance wascompared with other competitive trackers such as KCF、DSST、CN. Experimental results showed that the proposed al-gorithm could better applied in tracking with scale variation and improved the tracking accuracy.