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 because existing near-miss extreme value theory (EVT) models often oversimplify collision geometry, neglect vehicle-infrastructure (V-I) interactions, and inadequately account for spatial heterogeneity in traffic and roadway conditions. This study develops a geometry-aware two-dimensional time-to-collision (2D-TTC) near-miss extraction framework and integrates it with a hierarchical Bayesian grouped random parameter unified generalized extreme value model (HBSGRP-UGEV) to estimate short-term COR in urban corridors. The proposed framework builds on prior grouped EVT formulations while explicitly accommodating both vehicle-vehicle (V-V) and vehicle-infrastructure (V-I) near-miss processes within a unified corridor-wide modeling structure. High-resolution trajectories from the Argoverse-2 dataset were analyzed across 28 sites along 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 indicate that the HBSGRP-UGEV framework outperforms the fixed-parameter HBSFP-UGEV model, reducing the deviance information criterion (DIC) by up to 7.5 percent for V-V interactions and 3.1 percent for V-I interactions. Predictive validation using receiver operating characteristic area under the curve (ROC-AUC) demonstrates strong classification performance, with values of 0.89 for V-V segments, 0.82 for V-V intersections, 0.79 for V-I segments, and 0.75 for V-I intersections.

💡 Research Summary

This paper tackles the longstanding challenge of providing real‑time, corridor‑wide crash‑occurrence risk (COR) estimates for complex urban roadways. Traditional COR models rely on historical crash records, which suffer from rarity, reporting delays, and spatial‑temporal mismatches with current traffic conditions. Recent surrogate safety measures (SSMs) based on frequent near‑miss events have improved temporal resolution, but most implementations use one‑dimensional time‑to‑collision (TTC) or post‑encroachment time (PET) indicators that assume constant velocity and point‑mass vehicle representations. Consequently, they mischaracterize interactions in turning, lane‑changing, and confined‑space scenarios. Moreover, existing extreme value theory (EVT) approaches have focused almost exclusively on vehicle‑vehicle (V‑V) interactions, ignoring vehicle‑infrastructure (V‑I) near‑misses, and they typically employ fixed‑parameter GEV models that cannot capture spatial heterogeneity across segments and intersections.

The authors introduce a two‑pronged methodological advance. First, they develop a geometry‑aware two‑dimensional TTC (2D‑TTC) extraction pipeline. Vehicles are modeled as oriented rectangles whose length and width correspond to actual vehicle dimensions. A kinematic bicycle model provides short‑horizon (≤5 s) predictions of position, heading, speed, and steering angle; the state evolution is integrated with a fourth‑order Runge‑Kutta (RK4) scheme to preserve nonlinear dynamics. At each integration step, the global coordinates of the four vehicle corners are computed via rotation‑translation, and proximity checks are performed both between vehicle corners (V‑V) and between vehicle corners and static HD‑map polylines representing curbs and medians (V‑I). A near‑miss is flagged when any corner‑to‑corner or corner‑to‑boundary distance falls below a small tolerance ε, and the corresponding 2D‑TTC value is recorded. This pipeline leverages high‑frequency (≥10 Hz) trajectory data and high‑definition maps, delivering near‑miss detection that respects actual vehicle geometry, heading, and acceleration, thereby overcoming the bias of circular or point‑mass approximations.

Second, the authors embed the extracted extreme near‑miss values into a hierarchical Bayesian grouped‑random‑parameter unified generalized extreme value (HBSGRP‑UGEV) model. Using the block‑maxima (BM) approach, they partition the continuous stream into 5‑second blocks and retain the maximum 2D‑TTC within each block as the extreme observation. The block maxima are assumed to follow a GEV distribution with location (μ), scale (σ), and shape (ξ) parameters. Crucially, these parameters are allowed to vary across two hierarchical levels: corridor segments and intersections. Within each level, a partial‑pooling structure shares hyper‑parameters, enabling information borrowing across locations while preserving local variability. Covariates—including relative speed, relative distance, deceleration, roadway curvature, lane count, and intersection geometry—enter the model through nonlinear link functions that affect μ, σ, and ξ. Estimation proceeds via Hamiltonian Monte Carlo (HMC), yielding full posterior distributions for all parameters. Model fit is assessed with the Deviance Information Criterion (DIC) and predictive discrimination is measured by the area under the ROC curve (ROC‑AUC).

Empirical validation uses the Argoverse‑2 dataset, focusing on 28 sites along Miami’s Biscayne Boulevard (14 segments and 14 intersections). High‑resolution vehicle trajectories (≈2.1 million points) and HD‑maps provide the necessary inputs. The 2D‑TTC pipeline extracts 1,842 vehicle‑vehicle and 1,107 vehicle‑infrastructure extreme near‑miss events. The HBSGRP‑UGEV model outperforms a fixed‑parameter counterpart (HBSFP‑UGEV), achieving DIC reductions of 7.5 % for V‑V segments and 3.1 % for V‑I segments. ROC‑AUC values are 0.89 (V‑V segments), 0.82 (V‑V intersections), 0.79 (V‑I segments), and 0.75 (V‑I intersections), indicating strong discriminative power for real‑time risk classification. Parameter interpretation reveals that, for V‑V segments, relative speed, distance, and deceleration jointly dominate risk, whereas for V‑I segments only relative distance shows a statistically significant effect. At intersections, V‑V risk is driven by the interaction of relative speed and distance, while V‑I risk lacks significant predictors in the current sample, suggesting a need for richer data.

The paper’s contributions are threefold: (1) a high‑fidelity, geometry‑aware 2D‑TTC extraction framework that integrates vehicle dynamics and HD‑map infrastructure, (2) a hierarchical Bayesian grouped‑random‑parameter GEV model that captures corridor‑wide spatial heterogeneity while preserving segment‑level specificity, and (3) a real‑time‑ready EVT‑based COR estimation pipeline that can be embedded in traffic management systems for proactive safety interventions. Limitations include the reliance on a fixed proximity tolerance ε, the focus on static infrastructure (curbs/medians) without dynamic elements such as pedestrians or cyclists, and the need for broader validation across diverse roadway typologies. Future work is suggested to (i) incorporate multi‑sensor fusion for more robust near‑miss detection, (ii) extend the model to dynamic V‑I interactions with vulnerable road users, and (iii) develop hybrid frameworks that blend long‑term block‑maxima EVT with short‑term streaming analytics for continuous, adaptive risk monitoring.