Advancing real-time assessment for work-related musculoskeletal disorders in construction using large foundation models

• Automated synthetic dataset generation using foundation model orchestration • Worker7K: 22,000 annotated instances, largest construction pose dataset • 72.3 mean average precision on a real-world construction dataset • 69.8% improvement over laboratory-trained models on real construction data • Real-time operation at 40 FPS on consumer hardware with 6GB GPU Work-related musculoskeletal disorders present significant challenges in construction environments, necessitating high-accuracy and real-time monitoring solutions. This study introduces a novel deep learning-based framework for real-time ergonomic risk assessment by leveraging large foundation models to overcome critical data scarcity and computational constraints. A synthetic data generation pipeline integrating ChatGPT, Stable Diffusion, and SMPLer-X-H32 was developed to automatically create Worker7K, a construction-specific dataset comprising over 7,000 images and 22,000 annotated worker instances with precise 3D pose information. The new dataset is the largest domain-specific dataset of its kind. Statistical validation confirms that the synthetic dataset exhibits distribution characteristics highly consistent with real-world construction scenarios, with annotation accuracy reaching 61.3% mean average precision. A lightweight human pose estimation model was pre-trained on the public Human3.6M dataset and fine-tuned on Worker7K, achieving 72.3% mean average precision on real-world construction data, representing a 69.8% improvement over models trained solely on laboratory data. Comparative analysis with state-of-the-art methods demonstrates that the proposed model achieves competitive accuracy with fewer parameters than baseline approaches. The system operates at 40 frames per second on consumer-level hardware with only 6 GB GPU memory. Expert validation yields 97.4% overall accuracy in ergonomic risk classification and a Cohen's Kappa coefficient of 0.94 for identifying critical high-risk postures. Error propagation analysis confirms that pose estimation dominates system uncertainty, with REBA scoring remaining stable under typical error conditions. This research contributes a generalizable methodology for leveraging foundation models in resource-limited engineering applications, with broad implications for safety informatics and intelligent monitoring systems.