Development and evaluation of three learning based personalized driver models for pathtracking behaviors

The human-like control system of autonomous vehicle is an essential technical way to improve the system applicability and control performance by learning the driver's knowledge and experience. In this paper, three stochasticlearning-based personalized driver models that represent the path-tracking behaviors are proposed and compared. The main framework of these three models are all based on Gaussian Process (GP), and the difference lies in the hierarchical relationship of the driving data. The general GP driver model does not consider the layering relationship, and the GP driver model based on velocity and steering wheel angle respectively takes the clustering results of the corresponding parameters as the basis for layering. All the models are trained based on the driving data which is collected from the intelligent vehicle platform in Beijing Institute of Technology. Estimation accuracy results of the three methods are analyzed with different numbers of Gaussian Mixture Model (GMM) components. The consequences show that the steering wheel angle-based GP driver model achieves a better prediction performance than the other two driver models with the recommended number of GMM components 8-12.