摘要
针对大多数跟踪算法无法解决尺度变化问题和现有尺度解决方案存在冗余、固定的问题,提出一种基于核相关滤波框架的由粗到细快速和新颖地解决尺度估计的方法,考虑到仅利用响应图峰值进行比较存在不稳定性,以检测响应图与期望输出图的欧氏距离作为峰值的可靠程度,以两者的乘积作为最终比较结果;该方法首先使用3个尺度因子确定目标尺度变化方向,然后在尺度变化方向求解最优尺度;在OTB-100的26个带有尺度变化属性的基准序列上进行实验,并与现有其他先进跟踪算法进行定量和定性比较。结果表明:提出的方法能够很好地解决尺度变化问题;与核相关滤波相比,所提出算法的平均距离精度提高了18.8%,曲线下面积提高了19.6%,跟踪速度是稳健视觉跟踪的精确尺度估计的2.5倍,是特征整合尺度自适应核相关滤波跟踪算法的6倍。
Most tracking algorithms cannot solve the problem of scale variation and the existing scale solutions are redundant and fixed.To solve the problems,a fast and novel scale estimation method based on kernel correlation filtering framework is proposed,which is coarse-to-fine.Considering that the peak value of the response graph is not stable,the Euclidean distance of the detection response graph and the expected output graph are used as the reliability of the peak value.The product is taken as the final comparison result.Firstly,three scale factors are used to determine the direction of scale variation,and then solve the optimum in the direction of scale variation.The proposed algorithm is experimented on 26 benchmark sequences with scale variation attribute of OTB-100,and is quantitatively and qualitatively compared with other existing advanced tracking algorithms.The results show that the proposed method can solve the scale variation problem well.The proposed method is 18.8% higher in mean distance precision and 19.6% higher in area under curve than those of the kernel correlation filter.The tracking speed is 2.5times of the accurate scale estimation for robust visual tracking,and is 6times of the scale adaptive kernel correlation filter tracker with feature integration.
作者
何雪东
周盛宗
He Xuedong;Zhou Shengzong(Fujian Institute of Research on the Structure,Chinese Academy of Sciences,Fuzhou,Fujian 350002,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2018年第12期373-379,共7页
Laser & Optoelectronics Progress
关键词
机器视觉
目标跟踪
核相关滤波
尺度估计
遮挡检测
machine vision
target tracking
kernel correlation filtering
scale estimation
occlusion detection
