Automatic incident detection on Stockholm highways using vehicle-by-vehicle data

Sammanfattning

Traffic incidents on highways cause congestion, increase travel times, and raise the risk of secondary accidents, making fast and reliable automatic incident detection (AID) an important part of modern traffic management. This thesis investigates machine learning methods for AID using vehicle-by-vehicle traffic data from two highway segments in the Stockholm road network, Gröndal and Hallinge. Three research questions are addressed, concerning suitable evaluation metrics, the most informative features, and the best performing detection method. A set of aggregated and lane-based features was constructed containing speed, occupancy, headway, vehicle length, flow, and density measurements, as well as temporal and spatial differences between upstream and downstream sensors. Features were selected using MRMR ranking followed by model-specific forward selection and compared against a theory-guided manual selection. Three machine learning models, logistic regression, random forest, and kernel density estimation (KDE), were implemented and evaluated against the rule-based California Algorithm baseline. Performance was assessed at both minute- and incident-level using detection rate (DR), false alarm rate (FAR), mean time to detection (MTTD), and F1-score, within a five-fold cross-validation framework. The results show that no single metric fully captures detection performance and that a combination is needed. Speed-based features were consistently the most important, with spatial speed difference between sensors being the single most informative feature. Random forest performed best overall, outperforming the California Algorithm substantially. KDE was sensitive to variability in normal traffic and performed poorly at Hallinge. Lane-based features improved random forest performance but degraded simpler models, indicating that lane-level information requires a sufficiently flexible model to be exploited.