Video-Based 3D pose estimation for residential roofing-dataset

To protect residential roofing construction workers from both fatal and musculoskeletal injuries, it is necessary to assess the musculoskeletal and biomechanical risks in residential roofing tasks. This undertaking requires accurate information of workers’ 3D body positions to analyze kinematics and kinetics of the human body. In this study, we proposed a novel 2- stage motion estimation approach based on a convolution neural network to estimate residential roofer’s body positions using three-view video data. Instead of pursuing end-to- end training, our approach includes two stages: (1) use a multi-view model to estimate the 3D pose in a single frame; (2) use a multi-frame model to apply temporal convolutions to refine the multi-view outputs. The performance of the approach was evaluated by comparing our estimation with the gold-standard marker-based 3D human pose estimation (“ground truth”). The evaluation results show that our marker-free video-based approach can accurately capture the 3D posture of workers during the common roofing task and the proposed multi-frame model can effectively improve the precision of the coordinate sequence. The values of mean per joint position error of estimated human position before and after processing by the multi-frame model are 27.93 and 24.81 mm, respectively. These results prove that the proposed marker-free motion capture estimation approach can efficiently and accurately locate 3D body joints and pave the way for future onsite musculoskeletal motion analysis during roofing activities.