Real-time risk estimation for active road safety: Leveraging Waymo AV sensor data with hierarchical Bayesian extreme value models

Near-miss traffic risk estimation using Extreme Value Theory (EVT) models within a real-time framework offers a promising alternative to traditional historical crash-based methods. However, current approaches often lack comprehensive analysis that integrates diverse roadway geometries, crash patterns, and two-dimensional (2D) vehicle dynamics, limiting both their accuracy and generalizability. This study addresses these gaps by employing a high-fidelity, 2D time-to-collision (TTC) near-miss indicator derived from autonomous vehicle (AV) sensor data. The proposed framework uses univariate Generalized Extreme Value (UGEV) distribution models applied to a subset of the Waymo motion dataset across six arterial networks in San Francisco, Phoenix, and Los Angeles. Extreme events are identified through the Block Maxima (BM) sampling-based approach from each conflicting pair, with 20s block sizes to account for the scarcity of samples in short-duration traffic segments. The framework also incorporates conflicting vehicle dynamics (e.g., speed, acceleration, and deceleration) as covariates within a non-stationary hierarchical Bayesian structure with random parameters (HBSRP) UGEV models, allowing for the effective management of vehicle spatial, temporal, and behavioral heterogeneity. Results show that HBSRP-UGEV models outperform other approaches, with a 6.43-10.56% decrease in DIC, especially for near-miss events in short-duration traffic segments. The inclusion of dynamic vehicle behaviors and random effects substantially enhances the model's capability to estimate real-time traffic risks. This generalized real-time EVT model bridges the gap between active and passive safety measures, offering a precise and adaptable tool for network-level traffic safety analysis.