多特征融合的自适应加权采样上下文感知相关滤波跟踪算法  

Adaptive Weighted Sampling Context-aware Correlation Filter Tracking Algorithm Based on Multi-feature Fusion

作　　者：李勇锋 谢维信[1] LI Yongfeng;XIE Weixin(Key Laboratory of ATR National Defense Science and Technology,Shenzhen University,Shenzhen,Guangdong 518060,China)

机构地区：[1]深圳大学ATR国防科技重点实验室,广东深圳518060

出　　处：《信号处理》2022年第1期211-222,共12页Journal of Signal Processing

摘　　要：针对传统相关滤波器使用特征单一、背景信息不足等缺点,提出一种多特征融合的自适应加权采样的上下文感知相关滤波算法。首先,对于灰度图像序列,采用方向梯度直方图(FHOG)特征、局部二值模式(LBP)特征以及灰度特征相融合;对于彩色图像序列,则采用方向梯度直方图(FHOG)特征、局部二值模式(LBP)特征以及颜色(CN)特征相融合。其次,自适应采样响应图中的高响应区域,将其作为负样本引入滤波器训练中,并对最高响应区域赋予更高的抑制权重。对于目标尺度变化问题,引入尺度池进行尺度估计。最后,在OTB100数据集上实验的结果表明,相比原算法,本文算法的距离精度提高了8.5%,成功率提高了14.7%,并将本文算法与其他几种主流算法对比,验证了其有效性。Aiming at the shortcomings of traditional correlation filter such as single feature and insufficient background information,an adaptive weighted sampling context aware correlation filtering algorithm based on multi feature fusion is proposed.Firstly,for gray image sequence,Histogram of Oriented(FHOG)feature,Local Binary Patterns(LBP)feature and gray feature are fused;for color image sequence,Histogram of Oriented(FHOG)feature,Local Binary Patterns(LBP)feature and ColorNaming(CN)feature are fused.Secondly,the high response region in the response graph is adaptively sampled,which is introduced into the filter training as negative samples,and the highest response region is given higher suppression weight.For the problem of target scale change,scale pool is introduced to estimate the scale.Finally,the algorithm is tested on OTB100 data set.Compared with the original algorithm,the tracking accuracy and success rate of the proposed algorithm are improved by 8.5%and 14.7%respectively,and compared with other mainstream algorithms,its effectiveness is verified.

关 键 词：目标跟踪 相关滤波 上下文感知 多特征融合 加权采样 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]