Human pose estimation is a core computer vision task with broad applications, yet its performance degrades significantly in crowded scenes and under heavy occlusion due to missing or unreliable visual evidence. To address this limitation, this work reformulates occluded pose estimation as a structured pattern restoration problem and proposes a graph-based framework that models the human body as a relational skeletal graph. Starting from noisy or incomplete keypoint detections, the proposed method employs a graph neural network to propagate contextual information from visible joints to occluded ones through iterative message passing. Geometry-aware constraints on bone lengths and joint angles are integrated to enforce anatomical plausibility, while an occlusion-aware prediction mechanism distinguishes visible from missing joints during inference. Experiments on COCO-Keypoints, CrowdPose, and OCHuman demonstrate consistent improvements over strong baselines, particularly under moderate and severe occlusion, confirming the effectiveness of structural reasoning for robust pose estimation in real-world environments. These results confirm that explicit structural reasoning enables more accurate, stable, and reliable human pose estimation in real-world, occlusion-heavy environments.
Iqbal et al. (Tue,) studied this question.