Parameters Identification Research of Stylized Car-following Models Based on Behavior Clustering

Aiming at the problem that intelligent vehicle decision-making planning in heterogeneous traffic flow does not consider the differences of driving styles, this paper builds a driving behavior data acquisition system to obtain the natural driving data of drivers with multi-attribute characteristics, and excavates the car-following behavior of multi-source heterogeneous data based on multi-constraint extraction method. The time-frequency domain feature parameters of car-following association are extracted and the radical, conservative and calm categories are separated based on PCA and K-means methods. Three classical theoretical car-following models, Gipps, FVD and IDM, are selected. Based on genetic algorithm, the data of stylized labeling behavior are input for parameter identification, and finally three stylized car-following models are obtained. In order to evaluate the stylized identification models, the conventional single identification car following models are compared. By inputting the test set data and calculating the MSE values of the models, it can be obtained that the MSE value of IDM is smaller than that of Gipps and FVD, which has the best prediction effect. The MSE value of the stylized IDM is reduced by 31.3% compared with the non-style identification IDM, and the behavior prediction accuracy is higher.