Abstract: The current high-precision human pose estimation algorithm is inefficient in training and inference. To solve this problem, a lightweight human pose estimation model RepHRNet is designed. A Rep-shuffle block module based on structural reparameterization was proposed to improve the inference and training speed. The Lite fuse layer feature fusion layer was proposed to improve the efficiency of feature map fusion. Experiments show that compared with the lightweight high-resolution network LiteHRNet, the model achieves faster training and inference speed and higher accuracy on MPII dataset and COCO dataset.