Bayesian Updating Approach for Real-Time Safety Evaluation with Automatic Vehicle Identification Data

Abstract

Although numerous studies have attempted to use data from inductive loop and radar detectors in real-time crash prediction, safety analyses that have investigated the use of traffic data from an increasingly prevalent nonintrusive surveillance system have not included the tag readers on toll roads known as “automatic vehicle identification (AVI) systems.” This paper (a) compares the prediction performance of a single generic model for all crashes and a specific model for rear-end crashes that used AVI data, (b) applies a Bayesian updating approach to generate full probability distributions for the coefficients, and (c) compares the estimation efficiency of the semiparametric Bayesian modeling with that of logistic regression with frequentist matched case control. A comparison of AVI data collected before all crashes and rear-end crashes with matched noncrash data revealed that rear-end crashes could be identified with a 72% accuracy, whereas the generic all-crash model achieved an accuracy of only 69% when different validation data sets were used. Moreover, the Bayesian updating approach increased the accuracy of both models by 3.5%.