Computer vision techniques such as 3D human pose estimation can help college athletes obtain more accurate running motion data, such as step count, stride length, and posture. This provides data support for athletes to optimize their running posture and scientifically develop fitness plans. This study delves into the algorithm for 3D human posture estimation of collegiate athletes in running scenarios and suggests ways to improve its performance by utilizing temporal information and human body topology. This research presents a solution to the issue of occlusion-induced incorrect 3D posture estimation: a grouped spatiotemporal attention network for human body modeling based on a grouped human skeletal structure. With the use of a self-attention mechanism, a spatial encoder may improve the model’s grasp of spatial joint information and increase the possibility of predicting unusual poses by extracting local inter-joint linkages within the area and global inter-regional correlations. A temporal encoder is designed to capture the temporal relationships of the input frame sequence, obtaining a posture representation with temporal characteristics and alleviating occlusion issues.