Use of Support Vector Machine Models for Real-Time Prediction of Crash Risk on Urban Expressways

Abstract

This study adopted a novel methodology—a support vector machine (SVM) with two penalty parameters—for the evaluation of real-time crash risk on urban expressway segments by using dual-loop detector data. The purpose of this study was to develop a model that can effectively identify traffic conditions prone to crashes and support implementation of proactive traffic safety management. On the basis of crash data and the corresponding detector data collected on expressways of Shanghai, China, different combinations of dual-loop detector data and time segments before crashes were used to develop the optimal crash risk estimation model by SVM. The transferability of the SVM model was assessed by examining whether the model developed on one expressway was applicable to other similar ones. In addition, the prediction results and transferability of the SVM model were compared with those given by other frequently used classification algorithms, including logistic regression, Bayesian networks, naïve Bayes classifier, k-nearest neighbor, and back propagation neural network. The results showed that the SVM model was more suitable to the prediction of real-time crash risk with small-scale data than other algorithms, with its accuracy in classifying crashes reaching a best of 80%. A comparison to similar studies by other researchers implied that the proposed model achieved better prediction accuracy.