Research On Synthesis Of Human Movement With Spatial-temporal Variations

Nowadays motion capture technology has been widely used in computer animation and some MoCap databases which contain thousands of realistic human motion clips have been shared on Internet. Recently, how to synthesize crowd animation with human-like motion variations for each character becomes a hot topic in practice.This paper proposes a framework to synthesize spatial-temporal human motions after given a few of motion samples. The framework can be divided into two parts: spatial variations synthesis based on nonlinear manifold learning and temporal variations synthesis based on key frame interpolation.Given a group of sparse 3D human movement samples, a sparse motion semantic space is constructed with nonlinear manifold learning method. Then the sparse motion semantic space is intensively resampled to guarantee accurate synthesized result. Therefore, spatial motion variations could be synthesized by user's interaction.Given a few samples of a particular type of human movement, the temporal control points can be extracted with key frame detection and motion unit segmentation. All the samples' temporal control points are considered as prior knowledge. Then the temporal motion variations of candidate movement can be generated by changing the duration of each movement unit.The experiments show that our approaches achieve wonderful results, and the spatial-temporal synthesis is efficient.