密度峰值聚类的全尺度相关滤波跟踪方法  

Full-Scale Correlation Filtering Tracking Method Based on Density Peak Clustering

作　　者：肖振久[1] 李鑫 XIAO Zhenjiu;LI Xin(College of Software,Liaoning Technology University,Huludao,Liaoning 125105,China)

机构地区：[1]辽宁工程技术大学软件学院,辽宁葫芦岛125105

出　　处：《计算机工程与应用》2023年第5期131-139,共9页Computer Engineering and Applications

基　　金：辽宁省教育厅科学技术研究项目(LJ2020JCL023)。

摘　　要：针对背景感知相关滤波算法对目标进行尺度预测时采用图像金字塔,在跟踪过程中需要根据多尺度的目标样本训练出跟踪滤波器,导致跟踪过程中存在计算开销大、跟踪速度慢的问题,提出了一种基于密度峰值聚类的全尺度跟踪方法。分别提取搜索区域中目标和背景的特征信息,通过密度峰值聚类方法分别对目标和背景进行聚类;通过单尺度的滤波器对目标位置进行粗预测,并通过交替方向乘子法(ADMM)降低滤波器训练的时间复杂度;对搜索区域中的前景点和背景点进行分类,并根据尺度置信度得到目标的最终位置与尺度。该算法和目前一些主流的跟踪算法在公共数据集通过在OTB2013、OTB2015和DTB70上进行实验,在有效提高跟踪速度的前提下,面对旋转、遮挡等多种复杂情况时的跟踪效果较好、跟踪成功率较高,满足实时性要求。Since image pyramid is adopted in scale prediction of targets with background aware correlation filtering algorithm, it’s required to train a tracking filter with multi-scale target samples during tracking, resulting in excessive computation and slow tracking speed. In order to solve these problems, a full-scale tracking method is proposed based on density peak clustering. First, the feature information of target and background in the search area is extracted, followed by clustering of target and background using density peak clustering;second, the target location is roughly predicted with single-scale filter, after which alternating direction method of multipliers(ADMM)is used to reduce the time complexity of filter training;finally, the foreground points and background points in the search area are categorized to obtain final location and scale of target based on scale confidence. According to the experiments conducted for the proposed algorithm and current mainstream tracking algorithm on OTB2013, OTB2015 and DTB70 with public data sets, the proposed algorithm achieves better tracking effect and higher success rate in face of complex circumstances such as rotation and blocking in addition to effective improvement of tracking speed, thus satisfying the requirement on real-time performance.

关 键 词：目标跟踪 相关滤波 密度峰值聚类 计算机视觉 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]