Motion recognition and synthesis based on 3D sparse representation

Motion synthesis and recognition based on 3D motion data has been extensively studied in recent years. In this paper, we extract a dimensional representation of human motions from 3D spatial-temporal features and map this representation to low-dimensionality subspaces, which can preserve the intrinsic properties of original data. A method for automatic quantitative synthesis of human motion styles is then proposed. These methods help to make recognition and classification of 3D motion data more efficient, reducing computational complexity whilst preserving the intrinsic properties of original data. This also makes it useful for animation authoring systems and motion recognition. Experimental results show the effectiveness of the proposed methods.