基于检测的人体跟踪算法

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。

A detection-based person tracking algorithm

The traditional object tracking algorithms require manually annotated tracking area, and suffer from the problem of drift. To address these difficulties, we focus on the problem of person tracking in this paper, and propose a new detection-based tracking algorithm. To reduce failure in tracking, we employ multiple detectors to locate multiple body parts, and then map their detection results to a common body area. To adapt for quickly moving objects, we employ KLT tracker and agglomerative clustering for linking the detection windows to form person body trajectories. The experimental results reveal: using multiple detectors improves the tracking performance significantly; KLT tracker is adaptable for quickly moving objects. This algorithm is nearly real-time.