Bedridden Posture Tracking Using Pressure Mapping and Keypoint Detection

Abstract

Significant challenges persist in caring for severely disabled bedridden patients during critical tasks such as back-raising and toileting assistance. Ensuring correct posture is vital for safe care delivery, yet reliance on manual repositioning increases caregiver burden and patient safety risks. To address this, we proposed an automated posture tracking method integrating pressure distribution imaging and keypoint detection. We designed a large-area flexible pressure sensor to capture pressure distribution data, which are processed through smoothing and denoising to generate intuitive images. This enables pressure distribution-based posture recognition, anatomical keypoint localization, and body positioning in the supine posture. The system employs lightweight neural networks: ShuffleNet for posture classification (98.81% $\pm ~0.24$ % leave-one-subject-out (LOSO) cross validation accuracy) and YOLOv8pose-slimneck for keypoint detection (0.072–0.310-pixel error). Clinically validated positioning achieves 13.66 mm horizontal and 18.71 mm vertical accuracy, enabling precise guidance for horizontal movement and establishing a foundation for fully automated back-raising and toileting care robots. Furthermore, keypoint detection effectively identifies bony prominences (or landmarks), offering robust support for pressure ulcer prevention. Implementation requires no facility modifications and demonstrates 92.37% $\pm ~1.2$ % occlusion resistance under bedding. This solution establishes the groundwork for fully automated care robotics, significantly enhancing patient safety and care efficiency.