基于相关滤波的目标快速跟踪算法研究

Research on target fast tracking algorithm based on correlation filtering

在实现高精确度和快速的目标跟踪过程中,相关滤波是一个非常好的选择,但是目前所有的相关滤波跟踪方法仍然无法解决遮挡和光照变化等因素造成的干扰。因此,在传统核相关滤波器(KCF)的基础上,提出多特征图核相关滤波器(MKCF)的目标快速跟踪方法。首先,由初始化目标区域生成多个特征图,并通过对正则化最小二乘(RLS)分类器学习获得位置和尺度核相关滤波器(KCF);然后,随机选取一个特征图,寻找位置和尺度KCF输出响应的最大值,完成目标位置和尺度的检测;最后,随机选择需要在线更新的目标模型。经过试验测试,对比KCF,MKCF的平均中心位置误差(CLE)减少了5像素,平均成功率(SR)提高了10.9%,平均距离精度提高了6.7%;MKCF在目标发生尺度变化、光照变化、形态变化、目标遮挡、轻度旋转及快速运动等复杂情况下均有较强的适应性,具有重要的理论和应用研究价值。

Correlation filtering is a very good choice to achieve fast target tracking with high accuracy,but currently all the correlation filtering tracking methods are still unable to eliminate the interference caused by factors such as occlusion and illumination change. Therefore,a fast target tracking method using the multi-feature graph kernel correlation filter(MKCF)is proposed on the basis of the traditional kernel correlation filter(KCF). First,multiple feature graphs are generated by initializing the target area,and the location and scale KCF are obtained by means of studying the regularized least squares classifier.Second,a feature graph is randomly selected to look for the maximum output response value for the position and scale KCF,and complete the location and scale detection of the target. Finally,the target model that needs to be updated on line is randomly selected. The experiment was carried out. Compared with KCF,the average center location error(CLE)of MKCF reduces 5 pixels,the average success rate(SR)is increased by 10.9%,and the average distance accuracy is increased by 6.7%. MKCF has strong adaptability in complex conditions when the scale,illumination and form changes,as well as target occlusion,slight rotation and fast motion occur. It has important value in theory and application research.