特征权值与尺度自适应的核相关跟踪算法  

Feature-Weight and Scale Adaptive Algorithm for Kernel Correlation Tracking

作　　者：朱宏基 于凤芹[1] Zhu Hongji;Yu Fengqin(School of Internet of Things, Jiangnan University, Wuxi, Jiangsu 214122, China)

机构地区：[1]江南大学物联网工程学院

出　　处：《激光与光电子学进展》2019年第10期219-225,共7页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61573168);中央高校基本科研业务费专项资金(JUSRP51733B)

摘　　要：提出了一种特征权值与尺度自适应的核相关跟踪算法。提取目标搜索区域的方向梯度直方图(HOG)特征和颜色名(CN)特征进行自适应权值融合,通过融合特征的相关滤波响应图的峰值找到目标位置;利用权值较大特征的相关滤波响应图的峰值和峰值旁瓣比的乘积作为尺度评估依据,对目标尺度进行粗略估计和精确估计,从而得到目标的最佳尺度。通过在目标跟踪标准(OTB-2013)数据集上的仿真实验,结果表明相比核相关滤波跟踪算法以及其他5种跟踪算法,所提算法在跟踪精度和成功率方面都有明显提高,跟踪精度为0.799,成功率为0.723,能较好地适应目标尺度的变化。A kernel correlation tracking algorithm exhibiting feature-weight and scale adaptation is proposed. The histogram of oriented gradient ( HOG) and the color name (CN) features of the target search area are extracted for performing adaptive weight fusion, and the target position is estimated using the peak value of the correlation filter response map of the fusion feature. Further, using the product of the peak value of the correlation filter response map and the peak sidelobe ratio of the large weighted feature as the basis for scale estimation, the rough and accurate estimations of the target scale are performed and utilized to obtain the optimal scale of the target. The results of the simulation experiments performed using the object tracking benchmark (OTB-2013) dataset show that the proposed algorithm exhibits obvious improvements in terms of tracking precision and success rate compared with other five tracking algorithms. The tracking precision and success rate obtained using the proposed algorithm are 0.799 and 0.723, respectively. Furthermore, the proposed algorithm can well adapt to the change of target scale.

关 键 词：机器视觉 目标跟踪 核相关滤波 特征融合 尺度自适应 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]