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

An important goal in highway traffic management is event detection – finding ways to use real-time data to alert authorities to potential problems as they begin to emerge. One approach is to estimate the average behaviour of steady-state traffic, and then use this as a baseline, detecting events in real-time data as sufficiently large deviations from this average. In a recent paper, LML Fellow Colm Connaughton and University of Warwick graduate student Kieran Kalair test this approach using time-series data from the UK National Traffic Information Service (NTIS).
As they note, one difficulty in any such analysis is that not all traffic anomalies correspond to traffic “events” – things such as collisions, obstructions or lane closures – which are of significant interest to road users and other stakeholders. A broken-down car, for example, may cause no traffic flow deviation at all during quiet periods. Meanwhile, some traffic flow disruptions during busy periods – those associated with phantom traffic jams, for example – don’t actually have any identifiable underlying cause. For this reason, the authors use labelled event data to first verify that the correspondence between important events and anomalies is sufficient to be of use in practice. They then exploit a large volume of data to obtain a robust understanding of the range of typical fluctuations about the average traffic density-flow relationship, interpreting large excursions from this normal range as proxies for significant events.
Altogether, the research offers a purely data-driven method for identifying when the collective behaviour of the traffic on a road section is unusual in some sense. The technique doesn’t detect individual events such as collisions, stationary vehicles or lane closures, but changes in flow patterns which might follow from such events. To validate the approach, the authors measured how periods identified as atypical relate to existing labels of incidents on roads provided by NTIS, finding that this unsupervised, data-driven methodology offers comparable or better performance than some existing models based on learning patterns from labelled data.
In practical applications, the researchers suggest, this approach should be thought of as a filter which may help human operators of smart motorway infrastructure focus their attention more effectively. Regardless of cause, it may help operators understand where in the infrastructure interventions might efficiently return the system to normal.
The paper is available at https://www.sciencedirect.com/science/article/abs/pii/S0968090X21001947
 

Leave a Reply

Your email address will not be published. Required fields are marked *