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