The invention belongs to the technical field of three-dimensional dynamic graphic processing, and relates to a three-dimensional human body animation reconstruction method. The method comprises the following process steps of S1, selecting a video segment of a single person, and inputting the video segment into a trained generative adversarial network model to obtain a human body articulation pointthree-dimensional rotation angle sequence, S2, adopting a Savizkg-Golag five-point secondary smoothing algorithm to carry out smoothing processing on each three-dimensional articulation point rotation angle in the movement process, S3, creating an initial three-dimensional cartoon model, and binding 23 articulation points of the human body with the same articulation points of articulation pointsof the three-dimensional cartoon model, S4, updating the obtained three-dimensional rotation angle to a three-dimensional cartoon model, and realizing synchronous change of the attitude of the three-dimensional cartoon model by utilizing the rendering effect of Unity3D. According to the method, a generative adversarial network model is adopted, samples do not need to be calibrated, and the earlier-stage sample preparation workload is reduced; by introducing two-dimensional coordinates, three-dimensional coordinates, three-dimensional rotation angles, deviation values of human body shape parameters and confrontation loss values of joint points, the model precision is greatly improved.