Three-Dimensional human motion analysis using LiDAR technology: A systematic review

3D Light Detection and Ranging (LiDAR) has gained increasing attention in the field of human motion analysis due to its capability to non-invasively capture dynamic 3D pose data. However, accurate extraction of motion information from sparse and unordered point clouds remains a considerable challenge. This study aimed to systematically review 3D LiDAR data processing methods for human motion analysis. The search employed five databases (Web of Science, Scopus, Medline, PubMed and Embase) to identify methods for 3D human motion detection using LiDAR data. Following screening of 752 articles, a total of 38 studies were included. Convolutional Neural Networks (CNNs) were the most commonly used algorithm for human detection from LiDAR data, while the PointNet family was widely employed in point cloud feature extraction for human joint position estimation. The CNN-based YOLOv3 model achieved the highest human detection precision (97.7%), while a fusion approach integrating 3D LiDAR data with four inertial measurement units achieved the lowest Mean Joint Position Error (30.0 mm). Use of a CNN-based Federated Learning framework with Dynamic Layer Sharing achieved the highest activity recognition accuracy (98%). Integrating deep learning techniques enables effective extraction of spatiotemporal motion patterns from raw, sparse, and unordered point clouds, enhancing LiDAR-based human motion analysis. Significant challenges still remain in improving the usability of 3D LiDAR systems, particularly in handling multi-person scenarios and ensuring robust human motion analysis in the presence of LiDAR occlusion.