An unsupervised learning approach for analysis of human motion is proposed. In this approach, by learning a set of hidden Markov models under constrains of minimal descript length criterion, a continuous gestures sequence could be segmented and clustered, and thus the segments and labels of the original sequence are automatically extracted. The approach contains two steps. First continuous gestures are discretized and an original solution is found in discrete domain based on MDL criterion. Then coming back to continuous domain, a set of HMMs is learnt under constrains of MDL criterion, The HMMs exploit richer dynamics and thus generate better results. Experimental results by using real human gesture data demonstrate the effectiveness of the approach.