Abstract: To cope with the problem that the traditional Kernelized Correlation Tracker （KCF） lack of detection for tracking failure, an improved KCF tracking method is proposed. First of all, the improved KCF tracker intercepts the training sample with Gaussian window method on the target location, and this sampling method can obtain more effective signal to noise ratio and at the same time reduce the introduction of background interference information, so that the tracker can have stronger adaptability under complex scene. Then, in the process of video target tracking, the new tracker detects whether target tracking is failure by calculating the correlation peak sidelobe ratio, and the target detector learns in the frame with higher value of correlation matching in kernel function. Once tracking failure is detected, the detector will correct the tracker to restore target tracking. The experiment verifies the robustness of improved algorithm of this paper, shows that the overall performance of this new tracker is improved by 13.2%, compared with the traditional KCF tracker.

Key words: visual object tracking, correlation filters, object redetection

摘要： 针对传统的基于核相关滤波器的跟踪方法（KCF）缺少跟踪失败检测的问题，提出了一种改进的KCF目标跟踪方法。改进的KCF跟踪器采用高斯窗口方法在目标位置上截取训练样本，这种采样方法可以获得更有效的目标信噪比并同时减少背景干扰信息的引入，从而使跟踪器可以在复杂场景下具有更强的适应性。在目标跟踪的过程中，通过相关运算的峰值旁瓣比检测目标跟踪是否失败，并在相关匹配值较高的位置学习目标检测器。一旦检测到跟踪失败，便对跟踪器进行纠正，恢复目标跟踪。通过实验验证了改进算法的鲁棒性，相比传统的KCF跟踪器的总体性能提高了13.2%。

关键词: 视觉目标跟踪, 相关滤波器, 目标重检测