结合特征融合和尺度自适应的核相关滤波器目标跟踪算法研究  

Object Tracking Algorithm Based on Feature Fusion and Adaptive Scale Kernel Correlation Filter

作　　者：马康 娄静涛 苏致远 李永乐 朱愿 MA Kang;LOU Jing-tao;SU Zhi-yuan;LI Yong-le;ZHU Yuan(Fifth Team of Cadets,Army Military Transportation University,Tianjin 300161,China;Institute of Military Transportation,Army Military Transportation University,Tianjin 300161,China)

机构地区：[1]陆军军事交通学院五大队,天津300161 [2]陆军军事交通学院军事交通运输研究所,天津300161

出　　处：《计算机科学》2020年第S02期224-230,共7页Computer Science

基　　金：军队重点学科专业建设项目(智能无人系统关键技术前沿跟踪研究)。

摘　　要：在目标跟踪过程中,改进尺度自适应策略、选择辨别能力强的特征是提高跟踪算法性能的重要途径。为解决核相关滤波算法(Kernel Correlation Filtering,KCF)不能适应目标尺度变化、采用单一的方向梯度直方图(Histogram of Oriented Gra-dient,HOG)特征对目标判别能力有限的问题,通过研究同一目标在不同尺度下相关响应值的大小,在分析大量统计数据的基础上发现其变化规律,提出了一种新的尺度自适应策略,并采取HOG和颜色属性特征(Color Name,CN)线性加权融合的方法提高对目标的判别能力。在OTB数据集上的实验结果表明,所提算法的准确率和成功率相比KCF算法分别提高了8.5%和28.9%,在尺度变化属性视频序列上的准确率和成功率相比KCF算法分别提高了8.1%和38.5%,在其他属性视频序列上的表现也有较大提高,并且跟踪速度达到37.68 fps,可满足实时性要求。In the process of object tracking,an important way to improve the performance of the tracking algorithm is to improve the scale adaptive strategy and select features with strong discrimination ability.In order to solve the problemthat Kernel Correlation Filtering(KCF)can't adapt to the condition of object scale variation,andonlyusesthe single feature of Histogram of Oriented Gradient(HOG)whose discrimination ability to object is insufficient,a new scale adaptive strategy is proposed by studying the correlation response value of the same object at different scales,and finding the changing rule based on the analysis of a large number of statistical data,the method of linear weighted fusion of HOG and Color Name(CN)is also adopted to improves the object discrimination ability of the algorithm.Experimental results on OTB dataset show that the precision and success rate of the proposed algorithm are 8.5%and 28.9%higher than those of KCF algorithm,8.1%and 38.5%higher than those of KCF algorithm on scale variation attribute video sequence,and the performance on other attribute video sequence is also greatly improved,and the tracking speed reaches 37.68 FPS,which meets the real-time requirements.

关 键 词：特征融合 自适应尺度 核相关滤波 目标跟踪 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]