Privacy-preserving 3D human skeleton reconstruction from ankle-level 2D LiDAR using deep learning

Human skeleton reconstruction enables smart monitoring and biomechanical analysis, yet RGB/RGB-D systems face privacy concerns in public spaces, illumination sensitivity, and occlusion. To overcome these limitations, we propose a privacy-preserving approach for 3D skeleton reconstruction using ankle-level 2D LiDAR-based Motion History Images (MHIs). Our method integrates a direction-guided multi-head deep learning model, leveraging ResNet18 as the backbone to extract features from the MHIs and classify walking directions (front, left, right, back), thereby improving pose estimation across diverse orientations. Experiments comparing ResNet18, ResNet50, and DenseNet121 demonstrate significant improvements in accuracy, with DenseNet121 showing the best performance in terms of joint position error, while ResNet18 offers the best trade-off between accuracy and inference speed for real-time deployment. An ablation study reveals the importance of direction classification, with the removal of this feature leading to increased errors, especially in joints such as the spine base. We further validate anatomical consistency by estimating subject-specific height from predicted skeletons, observing consistent accuracy across individuals. These results position our approach as a robust and privacy-conscious alternative to conventional methods for 3D skeleton reconstruction.