The high-intensity physical labor inherent in construction activities frequently leads to musculoskeletal disorders (MSDs), posing significant threats to worker health and safety. Traditional physical workload assessment methods suffer from limitations including subjectivity, invasiveness, or insufficient spatial accuracy in complex construction environments. This study proposes an automated, non-invasive physical workload assessment method that integrates LiDAR-based 3D pose estimation with biomechanical inverse dynamics analysis. A two-stage sparse convolutional network is developed, where the first stage performs global human detection and coarse keypoint estimation on sparse Bird’s Eye View (BEV) features, and the second stage refines keypoint localization through volumetric Region-of-Interest (ROI) grid pooling that enables explicit 3D spatial sampling within each detected bounding box, specifically improving the estimation accuracy of lower-body keypoints that are critical for workload computation. The 14 estimated body keypoints are classified into three priority tiers (Primary, Secondary, Tertiary) based on their contribution to the downstream biomechanical analysis. Plantar pressure data collected via smart insoles are fused with the estimated 3D joint coordinates to compute joint moments through a bottom-up inverse dynamics model. A multimodal dataset comprising 3478 annotated records was established on real construction sites, synchronously capturing 3D LiDAR point clouds and plantar pressure information from 20 participants. Experimental results demonstrate that the proposed model achieves an overall mean per joint position error (MPJPE) of 103.1 mm, with primary (lower-body) keypoints achieving an average error of 92.3 mm, including a 30% accuracy improvement at the ankle joints compared to the single-stage baseline. A comparative analysis of joint moments computed from predicted and ground-truth skeletons confirms that the system preserves correct biomechanical load patterns across all gait phases, with a systematic conservative bias (average overestimation of 3.97 N m) that favors safety in practical monitoring, enabling continuous non-invasive monitoring of construction workers’ biomechanical loads. • Non-invasive workload assessment integrating LiDAR pose estimation and insole sensing. • Two-stage sparse convolutional network with ROI refinement for lower-body keypoints. • Multimodal construction site dataset with 3478 annotated point cloud and pressure records. • Priority-based keypoint classification with 12.8% lower-body accuracy improvement. • Predicted joint moments preserving load patterns with a 3.97 N m conservative bias.
贾宜之 et al. (Fri,) studied this question.