Abstract: In order to control the wide range motion of virtual human, an action recognition method was proposed in the motion process of human's lower limbs. Firstly, according to the motion characteristics of human's lower limbs motion parameters used for action recognition were chosen and an action characteristic extraction method based on wavelet fractal and least square fit was presented. Then, the action recognition of each motion parameter of human's lower limb kinematic chain was realized based on support vector machine. After achieving action recognition result of each motion parameter, the evidence theory was employed to fuse the recognition results. The experimental results of motion capture data show that the method only needs a small amount of training samples and has good real-time performance which can meet the action recognition need of people with different body types and has a good application prospect.