Multimodal Graph Attention Networks for Real-Time Action Prediction and Accident Prevention in Industrial Environments

One of the major challenges faced by the manufacturing industry is the prevention of workplace accidents. In this context, ensuring compliance with safety regulations, such as the use of personal protective equipment (PPE) and the proper execution of specific tasks according to safety protocols, is essential, especially when supervisors or safety personnel are not present on site. To address this issue, we propose a multimodal method for real-time human action prediction in industrial environments, aimed at supporting accident prevention systems. Our approach integrates two parallel Graph Attention Networks (GATs): one based on human skeleton pose estimation, and another built from scene object detection graphs. By combining these two complementary modalities, the model captures both human motion dynamics and contextual environmental information. To the best of our knowledge, this approach has not yet been explored in the literature. The proposed method will be evaluated on two benchmark datasets: Kinetics-400 (a large-scale video dataset with diverse real-world actions), and UnsafeNet (a dataset featuring factory-recorded videos annotated with safe and unsafe behaviors). The expected results aim to demonstrate the feasibility of applying multimodal GAT-based architectures to enhance occupational safety through intelligent action recognition systems.