Improving Crash Frequency Prediction from Simulated Traffic Conflicts Using Machine Learning Based Microsimulation

Traffic microsimulation is a method used to model vehicle movements and predict crash occurrences based on historical data. Traditional models rely on simplified rule-based behavior, which can lead to unrealistic traffic conflict dynamics and inaccurate crash predictions.

This study employed a machine learning (ML) model to analyze driving behavior from large-scale trajectory datasets at five signalized intersections in Leeds, UK. The researchers used a two-dimensional Time-to-Collision metric to identify conflicts and applied Extreme Value Theory to predict crash frequency.

Findings revealed that the ML model's conflict data yielded crash predictions consistent with real-world statistics, while the rule-based model did not provide meaningful predictions. However, using ML-generated simulated crashes directly for real-world predictions showed poor results, indicating that while ML can simulate conflicts realistically, it still struggles with generating accurate crash scenarios.

These insights suggest that ML-based models could enhance traffic safety predictions without needing extensive calibration for specific locations.