Aiming at the problems of large amount of calculation and slow detection speed of the existing soccer pose estimation model, this paper proposes a lightweight improved algorithm based on YOLOv8-Pose model and applies it to the recognition task of soccer player’s foul behavior. Firstly, RepConv and DBB reparameterization modules are introduced to enhance the ability of multi-scale feature fusion and solve the problem of large scale difference of visual objects. A centralized intra-layer adjustment feature pyramid network is designed, which combines the deformable attention mechanism and CASPPF in a parallel way. The pyramid network is globally centralized adjusted in a top–down manner, and the spatial weight of the global representation in the network is increased, so that the improved algorithm can obtain comprehensive and discriminative feature representation. The cross-entropy loss function is replaced by the exponential sliding sample weighting function (EMA-SlideLoss) to enhance the classification ability of the model and improve the stability of training. Using these methods, we propose an improved model RDE-YOLOv8-Pose model. The experimental results show that compared with the YOLOv8-Pose model, this research method improves mAP@0.5 by 0.142 and achieves a detection speed of 47.4 frames/s. This study can effectively improve the motion capture ability of artificial intelligence methods for soccer players, and provide more efficient and accurate technical support for intelligent monitoring of sports events and auxiliary referee decision-making.