Unsupervised Incident Detection Model in Urban and Freeway Networks

Reducing the effects of incidents by their early detection is one of a crucial requirements for incident management. This paper presents automated incident detection model based on unsupervised approach that uses only traffic observations as a model inputs. First, a novel self-tuning statistic is introduced as a feature generation function to capture spatio-temporal relationship of traffic data in both, urban and freeway networks. Next, these features are used as input in the segment-based mixture model that learns complex data distributions and their parameters. Then, we use the Mahalanobis distance to determine whether the traffic observation corresponds to an incident or recurrent traffic state. The model performance is demonstrated for two network examples, freeway with real data and urban with simulated data. Results show that the developed method achieves high accuracy rates and early incident detection compared to widely used approaches, such as California algorithm series and their extensions.