A Surrogate Approach for Real-Time Stress Assessment of Urban Drivers Using Driving Video Data

Assessing driver stress can enhance Autonomous Vehicles’ (AVs) ability to recognize potential hazards in human-machine cooperative driving. In this study, we use a scene graph technique from the driver’s perspective as an intermediate representation for stress assessment, effectively modeling the complexity of real driving scenarios. We develop a dynamic scene graph readout model based on Global Context-Aware Attention (GCAA) to capture the spatiotemporal variability of the environment. Our approach integrates a Multi-Relational Graph Convolution Network (MR-GCN) and a Long-Short Term Memory Network (LSTM) to evaluate driver stress using field data from Nanjing, China. Results show that our model significantly improves driver stress classification, particularly for critical stress levels, with a 65.3% increase in balanced accuracy and a 38% rise in F-score compared to the baseline. The addition of GCAA enhances accuracy further, with an average improvement of 51.7%. This demonstrates our model’s superior ability to understand the contextual relationships within driving scenarios, making it highly effective for AVs to assess driver status in complex situations and interact more efficiently with human drivers. Furthermore, our model achieves a balance between accuracy and inference speed, making it suitable for real-time assessment tasks.