Real-time Crash Risk Evaluation by AdaBoost Support Vector Machine

Through the real-time evaluation of crash risk, a dynamic traffic management system has the potential to improve safety performance of a traffic corridor or a traffic network. Typically, traffic data on freeways are acquired from loop detector stations and presented as a time series of traffic flows or speeds or other derived variables. These time-series data become the basis for assessing crash risk and potential management options. The primary objective of this study is to describe a study that focuses on the real-time assessment of crash risk based on a machine learning approach - AdaBoost Support Vector Machine (B-SVM). The machine learning B-SVM algorithm, particularly suitable for imbalanced data sets, is applied to a case study with the use of crash and traffic database for a segment of Interstate 210 freeway in California. The evaluation indicated that the 5 minutes speed at the subject location had a relatively higher impact on crash risk, in comparison to nine other variables that were included in the model. The outcome from the adopted machine learning B-SVM approach is further evaluated and compared to three other methods - Decision tree, Kernel function, and Bayesian logistic regression. It was found that the proposed B-SVM algorithm could provide better prediction and higher F-value for crash risk evaluation.