Abstract:

To tackle the issues that vehicle driving risk influencing factors are complex and unknown in the cooperative vehicle-infrastructure environment， and that traditional data-driven and physics-based model-driven prediction models lack sufficient interpretability and adaptability， a method is proposed in which an ego vehicle driving potential field is constructed using surrounding vehicles' motion state information acquired by onboard sensors， and a generalized force model is established by integrating ego vehicle driving state information to construct a vehicle-to-vehicle interaction information network. A generalized force model integrates the vehicle's driving state to create a vehicle-to-vehicle interaction network， while a vehicle-to-road network is formed using road sign and marking data. These networks are then fused into a dynamic interaction information network. To handle the dynamic nature of these forces， a routing mechanism is introduced. A switching Transformer model， combining the Transformer framework with a hybrid expert model， forms the foundation of the proposed vehicle driving risk prediction model， which is enhanced by the interaction information fusion network. Results show that the switching Transformer model outperforms the standard Transformer， with a 3.48% reduction in Mean Absolute Error（MAE）， a 2.19% reduction in Root Mean Square Error（RMSE）， and a 1.39% increase in the coefficient of determination （R2） in the comprehensive test set. In the lane-changing test set， the model shows improved adaptability， reducing MAE by 11.1%， RMSE by 9.0%， and increasing R2 by 5.4%. Additionally， in risk prediction scenarios， the model demonstrates significant improvements in Inverse Time to Collision （ITTC）， with increases of 0.288 7， 0.200 5， 0.714 2， and 0.288 8 for car-following， parking， exiting， and mixed lane-changing and exiting scenarios.

Key words: connected and automated vehicles, interactive information fusion, traveling risk, dynamic potential field, mixture of experts