Tsoutsanis A., Ziakopoulos A., Vasileiou Z., Yannis G., “Unsupervised Detection of Harsh Cornering Behavior Using Smartphone based Telematics and Infrastructure Data”, at the Proceedings of the Transportation Research Board (TRB) 105th Annual Meeting, Washington, 11-15 January 2026.

This study presents an unsupervised approach for detecting harsh cornering behavior using smartphone-based telematics enriched with infrastructure data. Unlike traditional methods that require fixed device orientation or labeled datasets, the proposed framework combines GPS-derived yaw and heading changes, orientation-invariant inertial sensor magnitudes, and spatial validation using OpenStreetMap to isolate turning maneuvers with safety relevance. A preprocessing pipeline identifies candidate turns based on angular thresholds and proximity to mapped intersections. A multi-variate feature set is extracted from time-series windows around each turn, capturing translational and rotational dynamics. Outlier detection is performed using DBSCAN, with hyperparameters tuned via k-distance, and complemented by Isolation Forest to enhance robustness. The ensemble of detections shows strong separability when evaluated through supervised classifiers, achieving ROC-AUC scores above 0.98. Visual case studies further confirm the accuracy and interpretability of the identified events. This orientation-agnostic pipeline operates without labeled data, making it highly suitable for real-world deployment in driver behavior analysis, risk profiling, and road safety analytics. The study highlights the feasibility of combining sensor fusion and map based filtering to detect unsafe driving behaviors at scale.