Abstract: Propose a motion state recognition method based on 3 d dynamic characteristics, first USES a star skeleton structure characterization of the human body 3 d motion, through GMM for image sequence foreground image, ISOMAP nonlinear dimension reduction of manifold learning method is used to generate a low dimensional data subspace, the motion of 3 d data projection to nonlinear flow pattern in the low-dimensional subspace, and identify the motion state of inner structure of human body movement is conducted for each key points of the study, based on the integration of learning have hidden markov study so that can be used to automatically identify the motion state. The simulation experimental results show that the recognition speed and accuracy of this method compared to traditional methods have obvious improve, confirmed the feasibility of research method.