Can surrogate safety measures explain crash patterns at signalized intersections? evidence from network-wide...

Area of Science:

• Transportation Engineering
• Traffic Safety
• Data Science

Background:

• Connected vehicle (CV) data offers potential for proactive, network-wide safety management.
• Existing guidance on transforming CV trajectory data into actionable safety indicators is limited.
• This study introduces a framework to extract surrogate safety measures (SSMs) from CV data.

Purpose of the Study:

• To develop and validate a reproducible framework for extracting SSMs from CV trajectory data.
• To quantify the relationships between extracted SSMs and police-reported crashes.
• To assess the influence of modeling techniques on predicting crash outcomes.

Main Methods:

• Extracted three SSMs: critical deceleration rate to avoid a crash (DRAC), time-to-collision (TTC), and post-encroachment time (PET) from 54.8 million trajectory points.
• Mapped SSM counts to 139 signalized intersections in Tucson, Arizona, and linked them to five years of crash data.
• Employed Negative Binomial regression, Random Forest, XGBoost, and a diffusion graph neural network to model crash-SSM relationships.

Main Results:

• DRAC emerged as the most influential predictor across all model families, significantly associated with rear-end, left-turn, and angle crashes.
• An increase of one thousand DRAC events correlated with a 5-10% rise in expected crashes of these types.
• XGBoost models with DRAC achieved a coefficient of determination above 0.60 for rear-end and no-injury crashes; TTC showed smaller effects, PET was negligible.

Conclusions:

• DRAC is a reliable network-wide screening tool for identifying crash risk.
• TTC provides supplemental insights for specific crash types (turning, low-injury), while PET offers limited value at a 1-second sampling rate.
• The framework demonstrates the utility of large-scale CV data and modern modeling for data-driven crash prevention.