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

Area of Science:

• Transportation Engineering
• Traffic Safety
• Data Science

Background:

• Existing crash-occurrence risk (COR) models lack accuracy due to simplified collision geometry and ignored vehicle-infrastructure (V-I) interactions.
• Spatial heterogeneity in traffic and roadway conditions is inadequately addressed by current near-miss event (EVT) models.

Purpose of the Study:

• Develop a geometry-aware 2D-TTC near-miss extraction method.
• Integrate this with a hierarchical Bayesian structure grouped random parameters (HBSGRP-UGEV) for short-term COR prediction in urban corridors.
• Accommodate both vehicle-vehicle (V-V) and V-I near-miss processes within a unified framework.

Main Methods:

• Utilized high-resolution trajectories from the Argoverse-2 dataset for near-miss event extraction.
• Developed a HBSGRP-UGEV model incorporating vehicle dynamics and roadway features.
• Employed partial pooling across segments and intersections to capture corridor-wide heterogeneity.

Main Results:

• The HBSGRP-UGEV framework outperformed fixed-parameter models, reducing DIC by up to 7.5% (V-V) and 3.1% (V-I).
• Achieved strong predictive accuracy (ROC-AUC: 0.89 for V-V segments, 0.82 for intersections).
• Identified key risk factors: relative speed, distance, and deceleration for V-V near-misses; relative distance for V-I near-misses on segments.

Conclusions:

• The geometry-aware, spatially adaptive framework enhances proactive corridor safety management.
• Findings support real-time interventions and long-term Vision Zero goals.
• The model effectively captures complex V-V and V-I interactions for improved COR prediction.