Abstract: Occlusion is an important factor affecting the robustness of human estimation. This paper proposes a robust lightweight human pose estimation model based on Lightweight OpenPose. It utilized multi-resolution representation fusion module to enhance extraction of local features. It obtained long-distance context information through GCABlock to judge coordinates of occluded key points, and simultaneously used attitude adjustment machine to balance global and local information to achieve maximum contribution. By making lightweight improvements to initial and refinement stages, the number of parameters was reduced and positioning accuracy was improved. BoneLoss was used during training to increase network recognition of human body constraints information. Experimental results show that GMNet can effectively detect occluded poses, with parameters reduced to nearly 5.5% of OpenPose model and detection speed improved by about 6 times, and accuracy reaches about 92.6% of original model.