Monitoring sleep posture plays an important role in identifying sleep related issues, minimizing pressure related injuries, and improving comfort for patients. Traditional wearable-based systems often disrupt sleep and reduce comfort. This paper presents a non-intrusive, vision-based framework that combines Detectron2 for keypoint detection with a Graph Attention Network (GAT) for sleep posture classification. A custom dataset of four sleep postures—supine, prone, left lateral, and right lateral—was created, annotated with 17 COCO-format keypoints. Detectron2 extracts high-confidence human keypoints, which are structured into skeleton graphs using anatomical connectivity. These graphs are input into a three-layer GAT architecture with attention mechanisms, residual connections, and normalization to model spatial relationships between joints. Unlike traditional approaches that rely on handcrafted features, our method learns representations directly from keypoint coordinates, enhancing generalization and reducing manual intervention. Extensive experiments using 10-fold cross-validation and an unseen test set of 1,400 images demonstrate an overall classification accuracy of 98.57%. The framework remains robust under occlusion, lighting variations, and static poses, making it practical for deployment in real-world clinical and home settings.