Supporting Ergonomic Risk Assessment: An AI-driven Approach for Action Recognition Using Standard Video Recordings

Background: While AI has been widely adopted in medical, social, and industrial fields, its use in ergonomic risk assessment of occupational movements is limited. Existing automated activity recognition methods often rely on controlled conditions or extra sensors, reducing their practicality. This study aims to develop an AI-driven tool to digitize the Occupational Repetitive Actions (OCRA) risk assessment by automatically identifying occupational actions from standard video recordings in real manufacturing settings. The goal is to help ergonomists conduct risk assessments more broadly and support workers’ health. Methods: Video data were collected from 37 workstations across multiple manufacturing facilities, comprising 869 action clips representing 12 distinct occupational activities. Two-dimensional skeletal pose extraction was performed using MMAction2 to anonymize workers and reduce environmental confounding factors. Graph Convolutional Network architectures (ST-GCN and Two-Stream Adaptive GCN) were trained and validated using stratified dataset partitioning (67.3% training, 24.7% validation, 8.0% test). Results: The proposed model achieved approximately 60% classification accuracy, substantially outperforming the frequency-based (12%) and duration-based (24%) baselines. The performance evaluation used multiple metrics, including Mean Average Precision, F1-score, Jaccard/IoU, Dice-Sørensen coefficient, and Hausdorff distance. Conclusions: This work demonstrates the technical feasibility of AI-driven, markerless motion capture systems for digitalizing musculoskeletal disorder risk assessment in real-world occupational settings. The principal innovation is the successful implementation of action recognition from standard video recordings captured in uncontrolled manufacturing environments, which transcends spatial constraints and eliminates the need for specialized sensing equipment. Despite current limitations in dataset size and classification accuracy, this study lays a foundation for developing deployable platforms to support real-time ergonomic assessment and proactive prevention of work-related musculoskeletal disorders.