Prototype System for Real-Time Incident Likelihood Prediction

摘要

This research project focused on developing and validating freeway incidentlikelihood models based on traffic, weather and incident data collected by the Indiana Department of Transportation on the Borman Expressway in Northwest Indiana. Likelihood prediction models were developed for the two major types of incidents in terms of associated traffic delays, vehicle crashes and overheating vehicles. These models exhibit high goodness-of-fit to data and provide accurate predictions, they capture the effect of relevant traffic, location, weather and geometric variables on freeway incident probabilities. In the second part of the research the incident likelihood prediction models were incorporated into an incident likelihood prediction simulator. A traffic simulator (INTRAS) is used to generate the traffic characteristic (i.e., volume and speed) inputs to the freeway incident likelihood simulator while the environmental conditions are specified by the user. The simulator combines the sequential outputs from an existing incident detection algorithm and those of our incident likelihood prediction models through Bayesian updating. The likelihood prediction outputs are used as the initial prior probabilities. As the detection outputs are received every minute, the incident likelihoods are sequentially updated. Every fifteen minutes, new incident likelihood predictions are produced by the models and used to update the estimates of freeway incident likelihoods. This sequential fusion of incident detection and prediction produces better estimates of incident likelihoods because more accurate prior probabilities are used and because both traffic and environmental factors are taken into account.