基于动态贝叶斯网络的多人跟踪算法

Tracking Method of Multi-actors Based on Dynamic Bayesian Network

在单目视频多视角下的多人跟踪中,单一特征选取会造成识别困难。该文提出一种基于动态贝叶斯网络的分类特征联合建模的跟踪方法,将视频中基于时空的运动特征和轮廓特征相复合,采用先粗后精的方法解决由于视觉角度不同而造成的跟踪困难,实现同一场景中多视角下的多人跟踪。实验证明该方法有效且具有较好的鲁棒性。

In view of tracking difficulty of multiple view angles about multi-actors from monocular video, a tracking method based on Dynamic Bayesian Network（DBN） is proposed which operates on monocular gray-scale video imagery. It puts forward an united model according to the combination of the actors＇ figure features and the spatial motorial features in order to track multi-actors of multiple view angles in the monocular video. Experiment shows that this method is effective and robust.