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

2025-09-02

AI Summary

Existing corridor-level collision occurrence risk (COR) prediction models oversimplify collision geometry, neglect vehicle–infrastructure interactions, and fail to capture spatial heterogeneity—limiting their utility for real-time safety interventions. To address these limitations, this paper proposes a geometry-aware two-dimensional time-to-collision (2D-TTC) metric that jointly incorporates vehicle–vehicle (V–V) and vehicle–infrastructure (V–I) near-miss events with dynamic driving behaviors. We develop the first dynamic risk assessment framework that jointly models collision geometry and spatial heterogeneity. Leveraging high-resolution trajectory data, we design a Hierarchical Bayesian Spatiotemporal Grouped Random-Parameters model with Nonstationary Generalized Extreme Value distributions and Partial Parameter Sharing (HBSGRP-UGEV). The model achieves 7.5% and 3.1% reductions in Deviance Information Criterion (DIC) for V–V and V–I scenarios, respectively, and attains an ROC-AUC of 0.89. Key risk factors—including relative speed, inter-vehicle distance, deceleration behavior, and proximity to roadway boundaries—are identified, providing an interpretable, deployable statistical foundation for real-time collision warning and “Vision Zero” network optimization.

Technology Category

• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning
• Planning, Routing, and Scheduling: Optimization of Spatio-temporal Systems
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Security and Privacy: Large-scale security measurements

Abstract

Real-time prediction of corridor-level crash occurrence risk (COR) remains challenging, as existing near-miss based extreme value models oversimplify collision geometry, exclude vehicle-infrastructure (V-I) interactions, and inadequately capture spatial heterogeneity in vehicle dynamics. This study introduces a geometry-aware two-dimensional time-to-collision (2D-TTC) indicator within a Hierarchical Bayesian spatiotemporal grouped random parameter (HBSGRP) framework using a non-stationary univariate generalized extreme value (UGEV) model to estimate short-term COR in urban corridors. High-resolution trajectories from the Argoverse-2 dataset, covering 28 locations along Miami's Biscayne Boulevard, were analyzed to extract extreme V-V and V-I near misses. The model incorporates dynamic variables and roadway features as covariates, with partial pooling across locations to address unobserved heterogeneity. Results show that the HBSGRP-UGEV framework outperforms fixed-parameter alternatives, reducing DIC by up to 7.5% for V-V and 3.1% for V-I near-misses. Predictive validation using ROC-AUC confirms strong performance: 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. Model interpretation reveals that relative speed and distance dominate V-V risks at intersections and segments, with deceleration critical in segments, while V-I risks are driven by speed, boundary proximity, and steering/heading adjustments. These findings highlight the value of a statistically rigorous, geometry-sensitive, and spatially adaptive modeling approach for proactive corridor-level safety management, supporting real-time interventions and long-term design strategies aligned with Vision Zero.

Problem

• Addresses oversimplified collision geometry in near-miss models
• Excludes vehicle-infrastructure interactions in crash prediction
• Fails to capture spatial heterogeneity in vehicle dynamics

Innovation

• Hierarchical Bayesian spatiotemporal grouped random parameter framework
• Geometry-aware two-dimensional time-to-collision indicator
• Non-stationary univariate generalized extreme value model