典型相关滤波跟踪算法的比较与分析

Comparison and analysis of typical correlation filter tracking algorithm

为综合评价现有的相关滤波类算法,对典型的相关滤波跟踪器进行全面的比较与分析,从而为进一步完善相关滤波器的设计提供指引。从相关滤波跟踪理论的一般框架切入,重点对当前四种具有代表性的相关滤波跟踪器即KCF,DSST,HCF和ECO展开研究,分别从理论分析以及在大规模公开数据集OTB100上的实验表现详细地比较各算法的优劣。比较与分析结果表明,使用卷积特征的算法在跟踪准确性和鲁棒性上相比单纯使用人工特征的算法具有显著优势,然而跟踪速度也会急剧下降,具有尺度估计模块的跟踪器能够得到更优的跟踪成功图表现。最后对深度学习结合相关滤波方法存在的实时性不足、长时跟踪等问题进行分析,并对未来的发展趋势进行了展望。

Comprehensive comparison and analysis on typical correlation filter trackers are performed to synthetically evaluate the existing correlation filter algorithm,thereby providing guidance for further improving the design of correlation filter.Firstly,the general framework of correlation filter tracking theory is taken as the start;and then,the current four representative correlation filter trackers of KCF,DSST,HCF and ECO are researched focally,and the advantages and disadvantages of each algorithm are compared in detail from both theoretical analysis and experimental performance on large-scale public data set OTB100.The results show that the algorithms with convolutional features have significant advantages in tracking accuracy and robustness in comparison with the algorithm with artificial features,but the tracking speed will decrease sharply;the tracker with scale estimation module can obtain superior performance on tracking success plot.Finally,the poor real-time performance and long-term tracking in the deep learning combined with the correlation filtering algorithm are analyzed,and the development trend in the future is prospected.