Abstract: Human pose prediction refers to the prediction of human pose after movement according to the observable human pose sequence. The existing methods lack the description of the location range of the joint after motion. Therefore, a joint location relation spatiotemporal variation model is proposed. The spatial location relation learning module learns the location relation between different joints constrained by analyzing the relative location distribution among joints in multiple time video frames. The rationality of position relation is described, and the unreasonable joint estimation is reduced. The temporal location relation learning module learns the change relation of single-joint position of action class constraint with time by counting the variation of joint location of two adjacent frames. The reasonable range after the motion of the human joint is described, and the unreasonable joint estimation is reduced. The spatiotemporal relation fusion module uses the Gaussian distribution variance of the location relation as the weight to carry on the weighted fusion of the spatial location relation and the temporal position relation. The module considers the adjacency relation in human pose, and selects the integrated spatio-temporal location relation in the space range to obtain a more reasonable spatio-temporal location relation. The spatiotemporal variation feature learning module can dot multiply the global/local position relation of the node into the self-attention relation. The module use self-attention to find the effective sphere of influence of the joint to improve the accuracy of human pose prediction. The experiment was verified on two databases, Human3.6M and Penn Action. The experimental results show that the proposed model is superior to the existing human pose prediction methods.