Learning to Represent and Generalize Lateral Control Deviation Based on Driving Data

Abstract

The lateral control deviation caused by insufficient control algorithm and defective actuator response Will lead to the deviation between the actual driving trajectory and planned trajectory. How to use the driving data to represent lateral control deviations and then generalize them to the correction of planned trajectories will further improve the driving safety of intelligent vehicles. In this paper, a statistical model is established to predict the lateral control deviation. The lateral control deviation between the planned trajectory and the actual driving trajectory is represented by velocity, curvature, and heading deviation. We select three kinds of features to train the statistical models: the previous, current, and future state of path point sequences. The Gaussian Mixture Model (GMM) is used to train the statistical model of lateral control deviation, and Gaussian Mixture Regression (GMR) is used to predict the lateral control deviation. Meanwhile, the actual driving trajectory of the vehicle can be generated directly according to the planned trajectory and the vehicle state. Experiment results show that the proposed method in this paper can predict the future lateral control deviation value With higher accuracy by comparing it with the K-means clustering method.