Assessing multiple construction workers' physical fatigue risks in 3D space

This study introduces MultiPose3D, a lightweight deep learning model for simultaneously assessing the ergonomic risks of multiple construction workers in 3D space using monocular surveillance video. Existing vision-based ergonomic methods are limited to single-worker, person-centric pose estimation and cannot reconstruct workers' absolute positions in a shared 3D coordinate system in real time. MultiPose3D addresses the challenge through a three-module pipeline. DetPose localizes each worker via an object detector. SinglePose then estimates each worker's person-centric 3D skeleton. DepthPose, a lightweight network with depthwise separable convolutions, predicts the absolute root depth of each worker by learning a pose-conditioned scaling factor derived from the geometric relationship between a worker's projected image area and physical size. A two-phase training strategy is adopted: SinglePose is first trained on both 2D and 3D datasets, after which its parameters are frozen and DepthPose is trained on 3D datasets alone. Experimental validation demonstrates mean per joint position errors of 49.4 mm on H36M and average root depth errors of 225–383 mm across multiple datasets. On a construction-specific dataset, the model achieves a mean average precision of 81.0. A real-world case study confirms the model's ability to simultaneously assess physical fatigue risks across multiple workers, enabling continuous ergonomic monitoring in collaborative construction environments. Moreover, with a lightweight detection module, MultiPose3D is capable of real-time inference at 25.8 frames per second, demonstrating its practical effectiveness for real-world applications.