Abstract: A driving risk classification method based on driver's collision avoidance behavior is proposed, and the driving risk discrimination algorithms under different driving modes are established by using support vector machine with concurrent consideration of the effects of driving behavior, road condition and environmental factor on driving risk states. Training and validation on prediction algorithm are conducted based on the “100-car” natural driving data from Virginia Tech in the US. The results show that in driving risk prediction modeling, the consideration of the discrepancies in driver's behavior, road condition and environmental factor, in particular the driver's distraction state, is conducive to increasing the accuracy of prediction model. In addition, under the condition of false positive rate lower than 5%, the prediction of high-risk state for future driving process by using prediction algorithm created has higher accuracy, in favor of giving timely warning or correction aids to drivers in near danger state, providing a new idea for the research on collision avoidance warning strategy and control method

Key words: collision warning, driving risk discrimination, SVM, driving assistance system