基于相关滤波器的视觉目标跟踪方法  

Visual object tracking algorithm based on correlation filters

作　　者：张雷[1,2] 王延杰[1] 刘艳滢[1] 孙宏海[1] 何舒文[1,2] 

机构地区：[1]中国科学院长春光学精密机械与物理研究所,吉林长春130033 [2]中国科学院大学,北京100049

出　　处：《光电子．激光》2015年第7期1349-1357,共9页Journal of Optoelectronics·Laser

基　　金：国家"863计划"(2014AA7031010B)资助项目

摘　　要：为了解决视觉目标跟踪中的尺度预测的难题,本文在核相关滤波器的目标跟踪的框架下给出了一种尺度估计策略,并对传统的核相关跟踪方法中目标模型的在线更新方法进行了修改,提出了一种多尺度视觉目标跟踪算法。首先,通过对正则化最小二乘分类器(RLS)学习获得位置和尺度核相关滤波器(KCF);然后,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,在线更新目标模型。实验中,对12组具有挑战性的标准视频序列进行测试。实验结果表明,相对于现有的3种基于相关滤波器的跟踪方法中的最优者,本文方法的平均中心位置误差(CLE)减少了7.0pixels,平均成功率(SR)提高了18.3%,平均距离(DP)精度提高了5.6%;在目标发生尺度、光照、姿态变化、部分遮挡、旋转及快速运动等复杂情况下,本文方法均有较强的适应性,具有重要的理论和应用研究价值。In order to solve the scale prediction problem in visual object tracking,a scale estimation strategy is given in the framework of tracking with kernelized correlation filters,the online updates method of the target model in the traditional kernelized correlation filters based tracking scheme is modified,and a multi-scale visual object tracking algorithm is proposed in this paper.At first,the position and scale kernelized correlation filters are obtained by learning the regularized least-squares classifiers.Secondly,we complete the target position and scale detection by finding the maximum output response of the position and scale kernelized correlation filters,respectively.Finally,the target models are online updated.Corresponding experiment is performed on 12 challenging benchmark video sequences.The results show that the proposed algorithm reduces the median center location error by 7.0pixels,improves the performance by 18.3%in the median success rate,and improves the performance by 5.6%in the median distance precision compared with the best one of the other three existing tracking algorithms based on correlation filters.The proposed tracking algorithm is robust to scale changing,illumination variation,pose variation,partial occlusion,rotation,fast motion and other complex scenes,and it has important research value in theory and application.

关 键 词：视觉目标跟踪 相关滤波器 多尺度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]