Anomaly detection and classification in traffic flow data from fluctuations in the flow–density relationship

We describe and validate a novel data-driven approach to the real time detection classification of traffic anomalies based on identification atypical fluctuations in relationship between density flow. For aggregated data under stationary conditions, flow are related by fundamental diagram. However, high resolution obtained from modern sensor networks is generally non-stationary disaggregated. Such consequently show significant statistical fluctuations. These best described using bivariate probability distribution density–flow plane. By applying kernel estimation high-volume UK National Traffic Information Service (NTIS), we empirically construct these distributions for London’s M25 motorway. Curves plane then constructed, analogous quantiles univariate distributions. curves quantitatively separate typical states. Although algorithm identifies general rather than specific events, find that outside 95% curve correlate strongly with spikes travel associated congestion events. Moreover, size an excursion region provides simple, real-time measure severity detected anomalies. benchmarking its ability identify labelled events historical NTIS against some commonly used methods literature. Detection rate, time-to-detect false alarm rate as metrics found be comparable except situations when speed bi-modal. In such situations, new achieves much lower without suffering degradation other metrics. This method has additional advantages being self-calibrating adaptive: marking behaviour different each section road can evolve changes, example, due long-term roadworks.