Incremental online PCA for automatic motion learning of eigen behaviour

This paper presents an online learning framework for the behavior of an articulated body by capturing its motion using real-time video. In our proposed framework, supervised learning is first utilised during an offline learning phase for small instances using principal component analysis (PCA); then we apply a new incremental PCA technique during an online learning phase. Rather than storing all the previous instances, our online method just keeps the eigenspace and reconstructs the space using only the new instance. We can add numerical new training instances while maintaining the reasonable dimensions. The experimental results demonstrate the feasibility and merits.