EROAD
This project addressed a critical gap in road safety analysis: identifying dangerous curves before crash statistics make the risk obvious. Leveraging EROAD's anonymised telematics data, we processed 14 billion GPS records to extract and classify over 1.2 million curve traversal events on New Zealand's state highway network.
Approach
- Isolation of Turning Events – Applied spatial and bearing-change filters to detect curve traversals from raw telematics data.
- Trajectory and Centripetal Analysis – Modelled each traversal using entry, apex, and exit points to calculate velocities and centripetal acceleration ratios.
- Spatial Clustering (DBSCAN) – Grouped trajectories into discrete curve IDs, even in complex, consecutive-curve environments.
- Feature Engineering – Derived 94 behavioural and geometric metrics per curve, then applied PCA to retain the most predictive features.
- Crash Data Integration – Overlaid Waka Kotahi NZ Transport Agency Crash Analysis System records to label curves with observed crash history.
- Predictive Modelling – Trained and tested Support Vector Machine, GLM, Stochastic Gradient Boosting, and Neural Network models, with SVM achieving the best balance of sensitivity (76.4%) and specificity (65.2%).
- Relative Risk Ranking (Monte Carlo) – Generated a national curve risk index by simulating thousands of weight combinations across predictive features.
Impact
- Developed the first nationwide, telematics-based curve risk index to guide targeted safety audits and infrastructure investment.
- Identified consistent links between driver behaviour metrics — particularly centripetal acceleration patterns and trajectory variability — and crash presence.
- Highlighted specific curves with elevated risk, enabling road controlling authorities to act proactively rather than waiting for crashes to occur.
- Delivered international recognition through peer-reviewed conference presentations, reinforcing EROAD's profile in applied transportation research.
This work demonstrated how large-scale telematics can complement traditional road safety assessments, shifting from reactive crash analysis to predictive risk management.