一种基于多特征融合的核相关滤波目标跟踪算法

A target tracking algorithm based on kernel-correlation filtering based on multi-feature fusion

针对核相关滤波算法(KCF)难以处理目标遮挡、旋转、快速移动等问题,本文在KCF的基础上提出了一种基于多特征融合的核相关滤波目标跟踪算法。首先,对密集采样后的样本建立多个特征模型,通过岭回归分类器对不同特征模型进行训练得到相应的滤波模板,然后,在频域中求其最大响应,分配各个特征模型权重系数。最后引入一个遮挡判断机制,实现抗遮挡处理,以获得更为精确的目标位置。实验结果表明:改进后的算法能较好的满足视频跟踪的实时性要求,具有良好的鲁棒性和精确性。

It is difficult for the kernel-correlation filtering target tracking algorithm (KCF) to deal with problems, such as occlusion, rotation, and fast movement. In this paper, based on KCF, a kernel correlation filter strategy based on multi-feature fusion is proposed. Firstly, a number of feature models are built for the densely sampled samples, and the corresponding filtering templates are trained by the on-line ridge regression classifier for different feature models. Then, the maximum response is obtained in the frequency domain using Fourier transform, and the weight coefficients of each feature model are allocated. Finally, an occlusion judgment mechanism is introduced to achieve anti-occlusion processing to obtain more accurate target position. The experimental results show that the improved algorithm can better meet the real-time requirements of video tracking and has good robustness and accuracy.