Transferability of multivariate extreme value models for safety assessment by applying artificial intelligence-based...

Area of Science:

• Transportation Engineering
• Road Safety Analysis
• Traffic Engineering

Background:

• Traffic conflict techniques are advanced methods for road safety assessment.
• Limited research on the transferability of conflict-based crash risk models hinders large-scale traffic safety evaluations.
• This study addresses the need to assess the transferability of these models to new locations.

Purpose of the Study:

• To investigate the transferability of multivariate peak-over threshold models for crash frequency-by-severity estimation.
• To propose and evaluate two transferability approaches: uncalibrated and threshold calibration.
• To compare these approaches against a complete re-estimation method.

Main Methods:

• Developed multivariate peak-over threshold models for crash frequency-by-severity.
• Implemented an uncalibrated transfer approach by direct application of base models.
• Applied a threshold calibration approach, adjusting conflict thresholds (Modified Time-To-Collision and Delta-V) using local data.
• Compared transferability approaches with complete re-estimation using data from signalized intersections in Southeast Queensland, Australia.
• Utilized AI-based Computer Vision for automated extraction of road user trajectories and rear-end conflicts.

Main Results:

• The threshold calibration approach yielded the most accurate and precise crash frequency-by-severity predictions for target sites.
• Calibrating the threshold parameter in peak-over threshold models significantly enhances base model performance.
• Complete re-estimation for individual sites resulted in inferior fits and less precise crash estimates compared to transferability approaches.

Conclusions:

• Threshold calibration is a highly effective method for transferring traffic conflict models to new sites.
• This approach significantly improves the applicability of traffic conflict models for crash risk estimation at transport facilities.
• The findings support the broader use of transferable conflict-based models in traffic safety evaluations.