基于重检测机制的核相关滤波跟踪算法  

A Kernel Correlation Filter Tracking Algorithm Based on Re-detection Mechanism

作　　者：孙晓锋 贾子彦[1] 张雷 吴雪涛 SUN Xiaofeng;JIA Ziyan;ZHANG Lei;WU Xuetao(Jiangsu University of Technology,Changzhou 213000,China)

机构地区：[1]江苏理工学院,江苏常州213000

出　　处：《电光与控制》2021年第8期44-47,114,共5页Electronics Optics & Control

基　　金：国家自然科学基金(61701202);江苏省研究生创新基金(20820111928)。

摘　　要：传统的核相关滤波跟踪(KCF)算法不能很好地处理目标快速移动和大面积遮挡,容易导致目标丢失。在KCF算法的基础上,提出了目标丢失检测、第一帧重检测、扩展区域重检测3种机制来解决以上问题。利用最大响应分数和平均峰值相关能量(APCE)来判别目标是否丢失;在目标即将丢失时,采用扩展区域重检测机制;在目标图像与第一帧目标图像相似时,采用第一帧重检测机制。为了能体现出所提算法的跟踪性能,从VOT2016和OTB100数据集中选取了14组视频序列作为测试集,其中7组视频序列含有目标遮挡和快速运动情况。经过定量实验对比,所提算法相比传统KCF算法平均中心位置误差(CPE)减少了20像素,平均重叠率(OR)提高了16.1%。The traditional Kernel Correlation Filter(KCF)tracking algorithm cannot handle well when the target is moving fast or has large-area occlusionwhich may cause the target to be lost.Based on the traditional KCF algorithmthis paper proposes three mechanismsnamelytarget loss detectionfirst frame re-detection and extended area re-detectionto solve the above problems.The maximum response score and Average Peak Correlation Energy(APCE)are used to determine whether the target is missing.When the target is about to be lostthe extended area re-detection mechanism is adopted.When the target image is similar to the first-frame image of the targetthe first frame re-detection mechanism is adopted.In order to reflect the tracking performance of the proposed algorithm14 sets of video sequences were selected from the VOT2016 and OTB100 data sets as the test setsin which 7 sets of video sequences had the scenarios of target occlusion and fast motion.A quantitative comparative experiment shows thatcompared with the traditional KCF algorithmthe improved algorithm reduces the average Center Position Error(CPE)by 20 pixelsand increases the average Overlapping Rate(OR)by 16.1%.

关 键 词：目标跟踪 核相关滤波 重检测 第一帧 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]