Improving road incident detection algorithm performance with contextual data

Abstract

Road Incident Detection Algorithms (IDAs) help Traffic Management Centres (TMCs) detect, and hence respond to road incidents more quickly and effectively, minimising road network disruption, injury, and risk of secondary incidents. The focus of this project is on developing novel incident detection algorithm techniques to contribute to the field of incident detection. A major problem faced by state of the art IDAs is the differentiation of incidents from contextual factors. Contextual factors (contexts) are factors that can be expected to cause disruption in traffic conditions in the future. Examples include sporting events, public holidays, weather conditions etc. Although some studies have addressed this problem, none have done so effectively on real-world data. TMCs commonly find that IDAs raise too many false alerts from contexts, and complain of IDAs requiring too much time, effort or expertise to implement. This research project focuses on how incident detection algorithms can better differentiate incidents from contexts, in an effective and simple enough way to be used in TMCs. The proposed approach incorporates contexts within a traffic forecasting algorithm, which creates forecasts of traffic conditions that can be expected if no incident were to occur. The forecasting algorithm is found to be more accurate than a commonly used historical average predictor in forecasting average speed and flow data from loop detectors, by 4.4% and 4.0% respectively. Incidents are then detected by an IDA that compares real-time traffic conditions with the forecasts. The IDA is evaluated in offline and online tests in order to ascertain whether incident detection algorithm performance can be improved with the incorporation of contexts. In the offline test, the IDA was shown to improve its performance by using contextual data, in detection rate from 94.4% to 96.7%, and in false alert rate from 1.75% to 1.50%. When tested online, in a TMC, 75 alerts were raised that were confirmed to correspond to incidents, and 49 of these alerts elicited a response from operators to manage the incident. Post-test interviews found that the majority found the developed IDA to a useful addition to their current incident detection methods, and would choose to continue to use the system. These results show that contextual data can be used to improve the performance of incident detection algorithms, in a way that is suitable for use in TMCs.

More information

Identifiers

ORCID for Benedict Waterson: orcid.org/0000-0001-9817-7119

Catalogue record

Export record