Publications / 2024 Proceedings of the 41st ISARC, Lille, France
Work-related Musculoskeletal Disorders (WMSDs) are one of the prominent challenges facing the construction industry, and their effective management requires a risk-quantification strategy. A comprehensive method to assess WMSDs for the construction sector is the Rapid Entire Body Assessment (REBA). The conventional risk-quantification frameworks relying on manual analysis of postures and images are resource-intensive and not scalable. Hence, automated approaches using Computer Vision (CV) are gaining attention. Current CV-based methods to automatically estimate REBA scores do not track all the key points necessary for accurate computation; they rely on heuristics. The present study is part of a wider research effort to develop a CV-based, fully automatic REBA risk calculator. An important task in this effort is to develop and validate a key-point annotation strategy for accurately estimating various body parts and joints necessary for REBA score calculation. This study re-annotated 149,813 human images present in the open-access COCO dataset. About 0.5 million key points were added for three body parts: the head, neck, and hip midpoint. Then, this data was used to train the state-of-the-art CV algorithms, MMPose and Alphapose. A comparison of the ML models developed on the newly annotated data with the pre-existing heuristics-based approaches demonstrates significant performance gains in precision and accuracy. The proposed model has the potential to be widely adopted for the precise and quick estimation of the REBA WMSD risk-assessment framework in different industries.