分层关联的多目标跟踪算法研究

Research on Multi-object Tracking Using Hierarchical Data Association

检测跟踪(Tracking by detection)是近年来多目标跟踪领域的一个主要研究方向。遵循检测跟踪框架,提出一种基于分层关联的全局性的数据关联算法。首先利用目标检测器在整个视频上检测目标,得到检测响应;然后利用广义最小团图在视频片段中对检测响应进行数据关联,得到轨迹片段;最后再在整个视频中对轨迹片段进行分层关联,得到最终的轨迹。在公共数据集上的测试结果表明,该算法能够有效地对多个目标进行数据关联,具有较强的处理遮挡能力。

Tracking by detection is a main research direction in the field of multi-target tracking in recent years. We proposed a global multi-object tracking algorithm using hierarchical data association following the tracking by detection framework. We first obtained the detection response in the whole video using an object detector, and then utilized the to solve the data association problem on detection response in video clip and obtained tracklets. At last we obtained the object track by solving the association problem on traeklets in whole video using a hierarchical method. Experiments on the public datasets show the proposed method can solve data association and handle occlusion effectively.