摘要
针对核相关滤波器跟踪算法(Kernel Correlation Filter,KCF)在特征提取单一以及尺度估计不足而导致跟踪效果不佳的问题,本文提出了一种多特征融合的尺度自适应核相关滤波目标跟踪算法。首先,使用帧差法将相邻帧图像对应像素值相减得到差分图像;其次,提取差分图像的方向直方图特征,与目标的均一局部二值纹理特征以及颜色特征进行线性加权融合;最后,加入一种尺度估计策略,将尺度滤波器的估计值与分块算法的估计值融合计算得出目标的尺度和位置。实验数据表明,本方法能有效改善核相关滤波器的跟踪性能,且与其他主流算法相比,在尺度变换下也有较好的跟踪效果。
Aiming at the problem of single feature extraction and insufficient scale estimation of kernel correlation filter tracking algorithm,which leads to poor tracking effect,this paper proposed a multi-feature fusion and scale adaptive kernel correlation filter target tracking algorithm.First,the frame difference method was used to subtract the corresponding pixel values of adjacent frame images to obtain differential image features;secondly,direction histogram feature extraction was performed on the difference image,and then linear weighted and fused with the uniform local binary pattern features and color names features.Finally,an adaptive scale estimation strategy was proposed,which combined the estimated value of the scale filter and the block algorithm to calculate the scale and position of the target.On 16 test videos with scale changes,the algorithm in this paper has a better improvement than the KCF algorithm,with a distance accuracy of 99.3%,an increase of 5.7%.Our method can improve the tracking performance of the kernel correlation filter,and compared with other mainstream algorithms,it also has better tracking effect under scale transformation.
作者
戴煜彤
陈志国
傅毅
DAI Yutong;CHEN Zhiguo;FU Yi(School of Artificial Intelligence and Computer Science,Jiangnan University,Wuxi,Jiangsu 214122,China;Wuxi Research Center of Environmental Science and Engineering,Wuxi,Jiangsu 214153,China)
出处
《信号处理》
CSCD
北大核心
2021年第1期49-58,共10页
Journal of Signal Processing
基金
江苏省高等学校自然科学研究面上项目(17KJB520039)
江苏省“333高层次人才培养工程科研项目”(BRA2018147)
江苏省高校“青蓝工程”(2020年)。
关键词
目标跟踪
尺度估计
帧差法
均一局部二值纹理特征
特征融合
target tracking
scale estimation
frame difference method
uniform local binary pattern features
feature fusion
