A physics-informed attention model for integrated driving risk assessment

Accurate and stable quantification of driving risk is critical for enhancing the safety performance of autonomous vehicles (AVs). Such quantification not only effectively prevents traffic accidents but also, in scenarios where collisions are unavoidable, enables timely and targeted occupant protection by accurately assessing potential severity. This study proposes a physics-informed integrated risk assessment model (PIRAM), which fuses collision probability and severity predictions into a unified integrated driving risk (IDR) metric. First, an integrated driving risk prediction dataset (IDRPD) was constructed using driving simulator experiments. Multiple hazardous driving scenarios were simulated within CARLA's virtual environment to collect data on drivers' safety-critical decision-making behavior. Then, a neural network model combining data-driven methods with physics-based constraints was developed. Specifically, the model employs an attention mechanism to capture spatiotemporal dependencies in vehicle trajectory and map information, and integrates a dynamic bicycle model as a physical constraint to guide predictions in accordance with fundamental physical laws, thereby significantly improving both prediction stability and accuracy. Experimental results and case studies conducted on the IDRPD demonstrate that PIRAM outperforms several baseline models, increasing prediction accuracy for collision probability and severity by 7.9 % and 3.2 %, respectively, and enhancing prediction stability by 10.3 % and 5.9 %, respectively. Furthermore, PIRAM enables earlier risk warnings by an average of 0.5 s. These advancements offer a reliable quantitative basis for occupant protection strategies in AVs and underscore PIRAM's substantial potential to improve safety in autonomous driving applications.