Construction Worker Ergonomic Assessment via LSTM-Based Multi-Task Learning Framework

Work-related musculoskeletal disorder (WMSD) is a critical occupational hazard and among the leading causes of nonfatal injuries in construction. Rapid ergonomic assessment is important to proactively detect and prevent WMSD-related hazards. This study proposes a novel deep learning framework for ergonomic assessment from construction videos. First, continuous skeleton postures of workers are extracted using a deep-learning-based pose tracking algorithm. Second, a long short-term memory (LSTM) based multi-task learning (MTL) model is created to simultaneously classify various ergonomic poses in different body parts using time-series skeleton postures. Finally, Ovako working posture analysis system (OWAS) is applied to assess the ergonomic risk from the identified poses in different body parts. Real-world construction videos are used to demonstrate the efficacy of the proposed method. Compared with existing vision-based ergonomic assessment methods, the novelty and contribution of this study is: (1) this study leverages LSTM network to exploit the temporal dependency among time-series skeleton postures, which effectively mitigates the errors associated with a single-frame posture and improves the accuracy, and (2) MTL is adopted to learn a unified classifier for multiple body parts leveraging the commonality in human pose, leading to improved performance and computational efficiency.