SAPENet: Structure-Aware Pose Estimation with Attention and Skeletal Constraints

Human pose estimation plays a vital role in various fields such as action analysis, human-computer interaction, and sports understanding. However, existing methods suffer from limited feature modeling capabilities, making it difficult to capture long-range dependencies among keypoints, especially under occlusion or in complex backgrounds, often leading to inaccurate localization. In addition, most approaches lack structural constraints, resulting in anatomically implausible poses with unreasonable joint connections. To address these issues, we propose a structure-aware pose estimation framework that jointly enhances the global modeling of keypoints and the structural plausibility of the predicted poses. The proposed framework comprises two core modules: a cross-keypoint attention module that models long-range dependencies between keypoints and significantly improves the perception of overall pose structure in complex scenarios; and a skeletal consistency constraint module that explicitly incorporates human skeletal priors by enforcing joint connectivity and physiological constraints, thereby improving the structural coherence and stability of pose predictions. Experimental results on the Human3.6M and PoseTrack datasets demonstrate that our method outperforms existing approaches in terms of both accuracy and robustness, particularly in challenging conditions involving occlusion and multi-person interactions.