Human Motion Analysis from Depth Video Sequences Using Multi-scale and Multi-directional Features

摘要

The emerging cost-effective depth sensors have made easier the action recognition task significantly. In this paper, we propose an effective method to analysis human actions from depth video sequences based on multi-scaling and multi-directional transformation which provide additional body shape and motion information for action recognition. In our method, corresponding to the front, side and top projection views, we generate three Depth Motion Maps (DMMs) over the entire video sequences. More specially, the multi-scaling and multi-directional transformations are implemented on the generated DMMs of a depth video sequence. Finally, the concatenation of these features is used as a feature descriptor for the depth video sequence. With this new feature descriptor, the l2-regularized collaborative representation classifier (l2- CRC) is utilized to recognize human actions. The recognition results of Microsoft Research (MSR) Action3D dataset show that our method significantly outperforms than the other existing methods, although our representation is much more compact.