Synthesizing high-quality human animations from the motion capture data is an important technology. The cost for the motion capture system is quite high, and the motion data cleaning is also an exhausting work. Usually, the existing motion capture data is a long motion sequence. Therefore, in many practical applications, it is difficult to create new animations from the long motion sequence directly. So it is a hot topic to extract the primitive movement from the existing motion capture data for synthesizing new animations. Many existing methods seldom consider the time sequence of motion data and the correlation among the joints. In this paper, a new technology is proposed to extract the primitive movement for synthesizing new animations. Firstly, PCA is adopted to map the high-dimensional motion data into a low-dimensional space, and the squared Mahalanobis distance is used to measure the similarity between different poses. Secondly, dynamic time warping and the sum of mean squared error are combined to segment and label the motion capture data automatically. Finally, a probability transfer model is proposed to construct the motion graph, which can be easily used for synthesizing new animations based on constraints.