Modelling highway crashes using Bayesian belief networks technique and GIS

Modeling traffic crashes is a complex undertaking. Previous research studies have used a variety of techniques to analyze crashes. Conventionally, traffic crashes have been modeled using regression models. Recently, intelligent systems have been applied in highway safety modeling. Such methods include artificial neural networks, decision trees, nearest-neighbor rule, Bayesian methods, and clustering algorithms. One method that has not been well used in analyzing highway safety data is Bayesian Belief Network technique. This research investigated the use of Bayesian Belief Networks technique in highway safety modeling.; A prediction model using Bayesian Belief Networks technique is proposed as part of the efforts to enhance traffic safety data analysis. The technique takes advantage of the knowledge of causal relationships or statistical dependencies (or independencies) among the model variables. A simple hypothetical Belief Network that comprised of six variables i.e., annual average daily traffic (AADT), section length, number of lanes, surface width, maximum posted speed limit, and number of crashes per year for each road segment was constructed. The model allows for the prediction of number of crashes per year at a roadway segment given a set of values of each of the model variables.; Geographical Information Systems (GIS) was incorporated in the model for displaying model results. A stand alone GIS application was developed using MapObjects software package. Programming was done in Visual Basic environment. The final output of the model was the map of the roadway network showing predictions of crash category for each roadway section.; Two different datasets were used in modeling---state roadways with a maximum number of lanes of 6 (subset 1) and exclusively six lane highways classified as high crash locations (subset 2). The performance of the proposed model was evaluated using the prediction accuracy. The prediction accuracy is hereby defined as the percentage of the roadway sections whose crash occurrence was correctly predicted. The results obtained in this study yielded the prediction accuracy of 68.08% and 78%, for subset 1 and 2, respectively.