Due to the popularity of optical motion capture system

more realistic human motion data can be acquired easily and widely used in various applications such as video games

animation films

sports simulation and virtual reality. This paper proposes a framework and algorithm for 3D human motion synthesis based on nonlinear manifold learning. In this framework

high-dimensional motion samples are mapped into low-dimensional manifold

with nonlinear dimensionality reduction method

to the intrinsic representation of motion semantic features. Furthermore

the sample which is generated by user interactions in low-dimensional manifold can be reconstructed to obtain a 3D motion sequence which owns a new motion semantic feature by reverse mapping. The experimental results show that the method proposed in this paper can not only precisely control the physical features of motions(such as the location of a specific joint)

but also can be used to synthesize new motion data which owns abstract motion semantic

such as motion styles.