Lower Limb Morphological Analysis for Sleep Posture Estimation using 3D LiDAR

Abstract

Accurate sleep posture estimation can be achieved using 3D LiDAR by focusing exclusively on lower-limb morphology, as leg orientation provides significant discriminative features. This targeted approach reduces computational overhead while strengthening privacy by eliminating the need for full-body point cloud reconstruction. This study investigates classification of four important sleeping postures (supine, prone, right lateral, and left lateral) by isolating lower-limb morphology. Spatial data are recorded from subjects (N=15) using Light Detection and Ranging (LiDAR) sensor, with the preprocessing pipeline involving tilt correction, outlier removal, specific Region of Interest (ROI) selection to isolate the legs and sequential sampling to minimize computational latency. The classification was performed using a PointNetbased framework and validated exclusively through Leave- One-Subject-Out (LOSO) cross-validation to ensure model generalization across different individuals. Experimental evaluation confirms that lower-limb morphology provides sufficient spatial features for robust posture discrimination, yielding a mean LOSO accuracy of 99.76% across the four classes. By utilizing the sample size of 1024 points, the model successfully maintained high-fidelity classification while processing a significantly smaller anatomical subset, effectively eliminating the redundancy of full-body point clouds.