Accurate crash frequency prediction is essential for designing safer road infrastructure, but traditional rule-based traffic simulation models often fail to capture realistic conflict dynamics. A new study published on arXiv demonstrates that machine learning (ML)-based behaviour models can significantly improve crash prediction accuracy by learning human driving behaviour directly from large-scale trajectory datasets.
Study Methodology
The research, led by Liu, Xian, Prato, Carlo G., and Markkula, Gustav, conducted traffic microsimulation for five real-world signalised intersections in Leeds, UK. They compared a standard rule-based model with a state-of-the-art ML model. Simulated vehicle trajectories were analysed using a two-dimensional Time-to-Collision metric to identify simulated conflicts. These conflicts were then modeled using Extreme Value Theory to predict crash frequency. According to the study, the microsimulation combined with surrogate safety measures has increasingly been used as a proactive alternative to historical crash data for predicting crash frequency for current or planned road infrastructure designs.
Key Findings
The results showed that conflicts generated by the ML model yielded crash predictions in line with real-world crash data. In contrast, the rule-based model did not permit meaningful predictions, presumably due to a lack of model calibration to the specific simulated intersections. Interestingly, directly using ML-generated simulated crashes to predict real-world crash frequency also yielded poor results, suggesting that while current ML models can realistically reproduce conflicts, they are not yet able to generate realistic crashes. The researchers noted that existing microsimulation-based safety studies have adopted simplified rule-based behaviour models, which reproduce traffic flow reasonably well but often fail to generate realistic conflict dynamics, limiting crash prediction accuracy.
Comparative Analysis
| Model Type | Crash Prediction Accuracy | Location-Specific Calibration Required |
|---|---|---|
| Rule-based | Did not permit meaningful predictions | Yes |
| ML-based | In line with real-world crash data | No |
The study found that the ML model avoided the need for location-specific calibration, a key advantage over rule-based approaches.
Implications for Traffic Safety
The findings demonstrate that ML-based behaviour models are promising for improving crash prediction from simulated conflicts without requiring location-specific model calibration. This represents a shift from simplified rule-based approaches towards data-driven simulation. The study suggests clear future directions for ML-based traffic microsimulation, including further refinement of ML models to generate realistic crashes directly. For enterprise technology decision-makers involved in infrastructure simulation, this research underscores the value of integrating machine learning into safety assessment tools to achieve more reliable predictions without extensive manual calibration.
