Incident prediction: a statistical approach to dynamic probability estimation : application to a test site in Barcelona

Abstract

Real-time models for estimating incident probabilities (EIP models) are innovative methods for predicting the potential occurrence of incidents and improving the effectiveness of incident management policies devoted to increasing road safety. EIP models imbedded in traffic management systems can lead to the development of control strategies for reducing the likelihood of incidents before they occur. This paper presents and discusses the design, implementation and off-line testing of an EIP model in the PRIME (Prediction of Congestion and Incidents in Real Time for Intelligent Incident Management and Emergency Traffic Management) Project of the “Information Societies Technology Programme” of the EU. A statistically-oriented approach based on Generalized Linear Regression models with polytomous responses is developed: geometry, traffic and weather conditions are taken as explanatory variables at a road section level and a binary variable related to incident occurrence or otherwise for the prevailing conditions is taken as a response variable on the first level of decision. Once the probability of a generic incident has been predicted, the lower level models in the selected hierarchical approach will predict the probabilities of incidents in a set of categories defined at a design level. The EIP model has been incorporated in the AIMSUN microscopic simulation environment (developed by TSS ). AIMSUN is able to emulate a traffic management system, since it simulates traffic evolution including the replication of observed incidents and incorporates different modules of incident and traffic management in such a way that the impact of traffic management strategies can be evaluated by simulation. A test site in Barcelona, located in a 15-km portion of the Ronda de Dalt ring road provided the data for calibrating and testing the EIP module. The selected site is equipped with 12 CCTV cameras for traffic monitoring, 18 local controllers, 12 detection stations, 10 variable message panels and 13 variable speed signals. Detection stations provide measures of different traffic variables in lane detail every minute.