基于检测的人体跟踪算法  

A detection-based person tracking algorithm

作　　者：吴建宅[1] 陈芳林[1] 胡德文[1] 

机构地区：[1]国防科技大学机电工程与自动化学院,湖南长沙410073

出　　处：《国防科技大学学报》2014年第2期113-117,共5页Journal of National University of Defense Technology

基　　金：国家重点基础研究发展计划项目(2013CB329401);国家自然科学基金资助项目(61203263)

摘　　要：传统的目标跟踪算法需要人为标定跟踪区域,且受到漂移问题的影响。为了解决这些困难,针对人体跟踪问题,提出了一种新的基于检测的跟踪算法。为了减少漏跟踪,使用了多个检测算子,用来定位多个身体部位,将其检测结果映射到一个相同的身体区域。为了适应快速运动的目标,使用KLT跟踪和凝聚聚类将检测窗口连接起来形成人体轨迹。实验结果表明:使用多个检测算子明显地提高了跟踪性能;KLT跟踪对于快速运动目标具有适应能力。该算法基本满足实时性。The traditional object tracking algorithms require manually annotated tracking area,and suffer from the problem of drift. To address these difficulties,the problem of person tracking was focused on,and a new detection-based tracking algorithm was proposed. To reduce failure in tracking,multiple detectors to locate multiple body parts were employed,and then their detection results were mapped to a common body area. To adapt for the quickly moving objects,the KLT tracker and agglomerative clustering for linking the detection windows to form person body trajectories was employed. The experimental results reveal that using multiple detectors improves the tracking performance significantly,and the KLT tracker is adaptable for quickly moving objects. This algorithm is nearly real-time.

关 键 词：目标跟踪 人体检测 KLT跟踪 凝聚聚类 实时性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]