Prediction of Car-Following Risk Status Based on Car-Following Behavior Spectrum

Car-following risk status reflects the degree of risk situation in the car-following process. For the sake of real-time prediction of car-following risk status (CFRS), this paper, by extracting three CFRS indicators: the reciprocal of time to collision, the transverse oscillation coefficient, and the velocity instability coefficient, determines the status division thresholds based on car-following behavior spectrum. Six machine learning prediction models: the precise decision tree, the lifting tree, the linear regression fitting, the support the vector machine, the K-nearest neighbor, and the ensemble tree, are used for CFRS prediction. The accuracy rate, the recall rate of four-level risk status, and the average recall rate are selected to evaluate the prediction effects of the model. The prediction effects in different duration of prediction and different duration of feature index sequence are compared. Based on the driving behavior data of six typical car-following scenarios in different road types and traffic state combinations collected by the driving simulator, the analysis indicates that the precise decision tree is the best way to predict CFRS. The prediction effect of car-following on the branch in congested traffic flow is significantly better than that in other scenarios, and there is no significant difference between the other five scenarios. By increasing the duration of the feature index sequence, the problem that the deterioration of the prediction effects due to the increase of the duration of prediction can be alleviated. The research results provide technical support for early warning and prevention of vehicle active safety. © 2021 Editorial Department of Journal of Tongji University. All right reserved.