改进的核相关自适应目标跟踪算法及其实验验证

Improved Adaptive Target Tracking Algorithm Based on Kernel Correlation and Its Experimental Verification

为了解决目标跟踪中出现的快速运动、尺度变化、遮挡等问题,提出基于遮挡检测的核相关自适应目标跟踪。该方法首先利用核函数对正则化最小二乘分类器求解获得核相关滤波器;其次利用核相关滤波器计算特征响应图,同时学习一维尺度滤波器对尺度进行估计;最后,通过响应图的最大值和振荡程度来判断目标是否被遮挡;在未受到遮挡的情况下,更新学习目标的外观模型和尺度模型,实现自适应目标跟踪。在公开的标准数据集上的实验结果表明,相比原始核相关滤波算法,平均中心位置误差降低15%,平均重叠率提高10%;且在目标尺度发生变化、遮挡、光照变化、快速运动等复杂场景下有较强的鲁棒性、适应性。

In order to the problem of the fast motion,scale and occlusion in target tracking,a kernel correlation adaptive target tracking approach based on occlusion detection was proposed. Firstly,the algorithm uses a kernel function to solve the regularized least square classifier to obtain the kernelized correlation filters. Secondly,the kernel correlation filter is used to compute the convolution feature response map,and meanwhile,the scale of target was estimated by one-dimensional scale correlation filter. Finally,by judging whether the target is obscured by the maximum of the response and the degree of oscillation,the appearance model and the scale model of the target were updated without occlusion to realize the adaptive target tracking. The results of the target tracking experiment in the public datasets show that compared with the original comparative filter algorithm the proposed filter algorithm has better performance such that the average central position error was reduced by 15% and the average overlap rate was increased by 10%,and furthermore,the proposed approach maintains the good robustness and better adaptability in the elaborate scenes such as target scale changing or fast moving,occlusion,various illumination conditions and so on.