Modeling discretionary cut-in risks using naturalistic driving data

Abstract One of the operational issues that intelligent vehicles have to deal with is cut-into and by other vehicles. A vehicle cut-in risk model helps determine how an intelligent vehicle should react to the other vehicle’s cut-in behavior. On the other hand, such a model could also help intelligent vehicles carry out cut-in maneuver in a considerate manner to minimize the impact on following vehicles in the target lane. In this study, a discretionary cut-in risk model for vehicles is developed on the basis of field driving data and machine learning methods, namely, decision trees and Support Vector Machine (SVM). A united algorithm is developed to combine the two machine learning models for achieving enhanced conservativeness to the traffic states with high misclassification costs. To build the naturalistic driving database, the wavelet method is employed for filtering; the K-means approach, an unsupervised data learning method, is used to categorize the cut-in impact on the following vehicles in the target lane into three groups. The impact is indicated by the following vehicle’s average and maximum deceleration. Using this model, intelligent vehicles can assess the risk level during other vehicles’ cut-in process as well as their own impact on the following vehicle in the target lane when carrying out cut-in maneuver.