Hierarchical Bayesian threshold excess model for real-time vehicle-based conflict prediction in dynamic traffic environments

Vehicle-based collision risk assessment methods often exhibit a tradeoff between simplifying assumptions in physics-based models and the interpretability challenges of learning algorithms. To tackle this, methods based on Extreme Value Theory (EVT) have gained momentum in recent years, but there is a lack of studies employing EVT for vehicle-based applications. This paper proposes a new, context-aware conflict prediction algorithm using a hierarchical Bayesian threshold excess model. Contextual traffic data are integrated with vehicle sensor data to improve the robustness and accuracy of the model. The feasibility of real-time deployment is also examined by optimising computational efficiency, leveraging several implementations of the Hamiltonian Monte Carlo No-U-Turn Solver (NUTS). The results demonstrate that including traffic covariates improves the model goodness-of-fit by 4.80% in terms of Deviance Information Criterion, and generalisability with a decrease of 1.36% in mean absolute error. However, partially pooled models, while enhancing goodness-of-fit, result in a reduction of generalisation capabilities. Additionally, the No-U-Turn Sampler compiled in JAX demonstrated sufficient performance for both online training and inference, thus making this methodology a feasible solution for real-time deployment in vehicle-based applications.