Predicting Vehicle States of Naturalistic Driving Data Based on Adaptive Optimization Gaussian Process Regression

Predicting vehicle motion states plays a critical role in the planning and control of the automated vehicle. To improve the performance of the predicting, a novel Gaussian process regression (GPR) based on adaptive optimization was proposed. First, a large number of naturalistic driving data from 28 drivers was collected. In order to consider the characteristics of the different kind of drivers, a clustering algorithm method, i.e., the Gaussian mixture model Kullback-Leibler (GMM-KL) divergence was developed. Furthermore, the optimal initialization and the boundary of hyperparameters for each type of driver were obtained on the basis of the classification. To realize the accurate, stable and real-time predicting of the vehicle states, the adaptive optimization Gaussian process regression (AOGPR) was developed, which optimizing and updating the hyperparameters online with the help of the optimal initialization and bounded constraints. Finally, the comparison results between the AOGPR and traditional GPR(TGPR), as well as the Bayesian Network (BN) model were presented. The results showed that the proposed method reached a better accuracy, stability, performance for vehicle states predicting than the TGPR and BN model.