Abstract: Due to lack of information in 3-dimension, the problem of generating 3D human motion from 2D images taken by single camera is almost impossible at first sight and is regarded as an ill-conditioned problem. Our idea is based on a propagation model and an algorithm is proposed to acquire the 3D motion information. In contrast to the existing methods, our approach utilizes a set of rules to get the smooth starting points of the model within a predefined space. Then the spatial ambiguity arisen from the fact that there are two possible poses of a link in 3D is eliminated by our expanding algorithm. Finally, several experiment results are presented and validated.