Driving Risk Evaluation Based on Multidimensional Data

摘要

This paper proposes a crash-involvement risk evaluation method using multidimensional data including drivers' license and violation records, vehicle registration information, camera data at intersections, and traffic crash data. A bagging-support vector machine (Bagging-SVM) model is trained using both static and behavioral features to evaluate a driver's crash-involvement risk. A feature selection method based on the F-score is performed before the risk evaluation. The results show that this extraction method can improve the stability of the evaluation, and provides the importance of each feature. A case study on truck drivers from a medium-size city in northern China is presented. It is found that driving behavior such as violations, trips at midnight, and headways are significantly related to crash-involvement.