3D pose estimation dataset and deep learning-based ergonomic risk assessment in construction
Chao Fan, Qipei Mei, Xinming Li
Abstract
Pose estimation of construction workers is critical to ensuring safe construction and protecting construction workers from ergonomic risks. Computer vision (CV)-based 3D pose estimation for construction workers is increasingly used in ergonomic risk assessment (ERA) due to its considerable practicability and accuracy. Currently, the deficiencies of (1) dedicated datasets for construction activities and (2) informative 3D biomechanical models to both Rapid Entire Body Analysis (REBA) and Rapid Upper Limb Analysis (RULA) impede the performance of CV-based ERA in construction sectors. Therefore, this study introduces a deep learning-based ERA by introducing a new dataset, ConstructionPose3D (CP3D), that follows a proposed 3D biomechanical skeletal model to support REBA and RULA. This dataset contains approximately 421,000 accurate 3D poses and annotations for construction activities. The results indicate that the proposed deep learning ERA models trained with CP3D outperform those without CP3D in accurately estimating the poses of construction workers, leading to improved ERA.