Human pose estimation tasks require the use of visual cues and anatomical relationships between joints to locate key points, but convolutional neural network-based approaches have difficulty focusing on remote contextual cues and modeling the dependencies between distant joints. In this paper, we propose an implicit modeling method based on an attention mechanism to implicitly model the constraint relationships between key points by computing feature correlations between joints through multi-stage iterations. To address the problem that the network may weaken the invisible key points during training, a focal loss function is used to make the network more focused on complex key points. The experimental results show that the implicit modeling network can improve the performance of the human posture estimation network, and the algorithm with the HRNet network as the backbone network can improve the performance of the human posture estimation network in the MPII dataset, the benchmark dataset for human posture estimation, compared with the original network. The accuracy of the algorithm with HRNet network as the backbone was improved by 1.7% compared to the original network.