Learning from geometry-aware near misses to real-time COR: A corridor-wide grouped random parameters GEV framework

Real-time corridor-wide crash-occurrence risk (COR) prediction is challenging, since existing near-miss EVT models oversimplify collision geometry, neglect vehicle-infrastructure (V-I) interactions, and fail to adequately account for spatial heterogeneity in traffic and roadway conditions. To do so, this study develops a geometry-aware 2D-TTC near-miss extraction and integrates it with a hierarchical Bayesian structure grouped random parameters (HBSGRP-UGEV) to estimate short-term COR in urban corridors. Building on prior grouped EVT formulations while explicitly accommodating both V-V and V-I near-miss processes within a unified corridor-wide modeling framework. High-resolution trajectories from the Argoverse-2 dataset were analyzed across 28 sites on Miami's Biscayne Boulevard to extract extreme near-miss events. The model incorporates vehicle dynamics and roadway features as covariates, with partial pooling across segments and intersections to capture corridor-wide heterogeneity. Results show that the HBSGRP-UGEV framework outperforms fixed-parameter HBSFP-UGEV models, reducing DIC by up to 7.5% (V-V) and 3.1% (V-I). Predictive validation using ROC-AUC confirms strong accuracy (0.89 for V-V segments, 0.82 for intersections, 0.79 for V-I segments, and 0.75 for intersections). Grouped random-parameters (HBSGRP) framework indicate that relative (speed, distance, and deceleration) dominate V-V near-miss risk on segments, whereas V-I segment risk is primarily associated with relative distance; at intersections, V-V risk is driven by relative (speed and distance), while V-I dynamics exhibit no statistically significant effects. These findings demonstrate the value of a geometry-aware, spatially adaptive framework for proactive corridor safety management, supporting both real-time interventions and long-term Vision Zero goals.