Abstract:

In general, Kernel Correlation Filter（KCF） algorithm is vulnerable to actual detection including occlusion and other conditions. In order to make the tracking result more accurate, an optimized KCF approach combined with improved corner detection is proposed in this paper. With the appropriate number and strong robustness of adaptive Harris corner points, problems of slow extraction of redundant edge points and incomplete edge points caused by illumination variation in generalized Hough algorithm are solved. At the same time, the introduction of adaptive threshold approach minimizes the influence of noise on corner extraction. Subsequently, targets are segmented and tracked respectively. The problem that KCF algorithm is easy to lose targets when they have scale variation is solved according to relative positions of sub-blocks. The learning rate parameter is updated adaptively to reduce learning rate of KCF algorithm and the error of model updating when the target is occluded. Finally, to eliminate the drift phenomenon of KCF when the target moves rapidly, intersection over union and Hungarian algorithm are combined to correlate multiple targets, associated coordinates are taken out one by one and the final position is found using the outline of the target drawn by generalized Hough algorithm. Experiments show that the approach can improve the reliability of target tracking effectively.

Key words: Kernel Correlation Filter（KCF） algorithm, Harris corner detection, Hungarian algorithm, generalized Hough algorithm

摘要：

通常，核相关滤波（KCF）算法易受遮挡等实际检测情况的影响。为使跟踪结果更为准确，提出了结合改进角点检测的优化核相关滤波方法。由自适应Harris角点数量适宜且鲁棒性强的特点，解决了广义霍夫算法提取冗余边缘点速度慢，以及因光照变化导致的边缘点提取不完整的问题。同时，自适应阈值法的引入将噪声对角点提取的影响降为最低。将目标分块并对每一目标子块单独跟踪，由子块间相对位置解决KCF算法在尺度发生变化时目标易丢失的问题。此外，对学习率参数进行了自适应更新，降低了KCF算法的学习率，减少了在目标被遮挡时的模型更新误差。结合交并比与匈牙利算法关联多个目标，逐一取出对应坐标并由广义霍夫算法描绘的目标轮廓得出最终位置，抑制了目标快速运动时KCF算法的漂移现象。实验表明，所提方法有效提高了目标跟踪的可靠性。

关键词: KCF算法, Harris角点检测, 匈牙利算法, 广义霍夫算法