RGB-Only in-Bed Pose Estimation Under Occlusion via Image Reconstruction
Sewon Kim, Hyebin Kim, Seohyun Lee, Taeyong Lee · IEEE Access · 2026
In-bed pose estimation is important in aging societies, where the growing population requires continuous monitoring. A key challenge in in-bed pose estimation is occlusions caused by bedding or medical equipment. While prior approaches have relied on multimodal data or complex models to address occlusions, such solutions require additional sensors, limiting their practical applicability. In contrast, this study proposes an RGB-only pose estimation framework under occlusion that integrates image reconstruction. Two image reconstruction models, MAE and AOT-GAN, were combined with three pose estimation models, OpenPose, ViTPose, and HRNet. Both image reconstruction models were fine-tuned on the Simultaneously-Collected multimodal Lying Pose dataset. Integrating image reconstruction prior to pose estimation improved pose estimation performance, reducing mean per joint position error (MPJPE) by approximately 30% compared to estimation performed directly on occluded images. To evaluate generalization, the proposed framework was further validated on an independently collected dataset. The MAE-ViTPose configuration achieved a 35% reduction in MPJPE and a 25% increase in the area under the percentage of correct keypoints curve (PCK-AUC) for lower-limb keypoints. However, this improvement was model-dependent, as AOT-GAN increased MPJPE by 21% on the unseen dataset. These results demonstrate that the MAE-ViTPose framework enables enhanced in-bed pose estimation performance under occlusion using camera-based images alone, underscoring its potential for in-bed monitoring without multimodal data.