The International Association for Automation and Robotics in Construction
Publications / CCC 2025 - Zadar, Croatia
FEASIBILITY OF AN INTEGRATED MULTI-MODEL APPROACH FOR DYNAMIC MUSCULOSKELETAL DISORDER RISK ASSESSMENT
High-load, repetitive construction tasks pose significant risks for the development of musculoskeletal disorders (MSDs), which adversely affect workers' health, productivity, and quality of life, making accurate risk quantification crucial for timely prevention. In recent years, computer vision (CV) has demonstrated significant potential in assessing MSD risks by offering an automated, contactless, and adaptable monitoring approach in complex environments. Existing approaches that rely on traditional biomechanical frameworks (e.g., REBA)?which assess MSD risks based on individual static frames by averaging the results?fail to capture the nonlinearly accumulating nature of MSDs risks in continuous motion and overlook individual variability in dynamic indicators such as amplitude, frequency, and smoothness. These limitations introduce bias, underscoring the need for an adaptive, personalized dynamic MSD risk scoring technique. To address these limitations, this study investigated the feasibility of a multi-model framework that integrates an enhanced Spatiotemporal Graph Convolutional Network (ST-GCN), a lightweight Transformer, and a diffusion model. It first employed an ST-GCN with dynamic adjacency and adaptive hypergraphs to model short-term dependencies; next, a Transformer refined long-term motion patterns; and finally, a diffusion model generated personalized risk score distributions to track MSDs risk evolution. The framework was trained on the HMR 2.0 dataset and evaluated on both HMR2.0 and MoYo dataset. Joint information was extracted for comprehensive MSDs risk assessment, and the resulting risk scores were compared with those from traditional biomechanical static single-frame methods. The results demonstrated that the proposed approach captured individual variability and nonlinearly accumulating MSDs risks more effectively, confirming the superiority of the proposed dynamic MSDs risk assessment. The findings underscore the significant potential of the proposed model for video-based MSDs risk assessment to enhance the accuracy of automated, real-time MSDs risk monitoring in high-load and dynamic environments such as construction sites.
