Data-driven Trajectory Planning of Lane Change Maneuver for Autonomous Driving

This paper presents a methodology of trajectory planning for the surrounding-aware lane change maneuver of autonomous vehicles based on a data-driven method. The lateral motion is planned by sampling candidate patterns which are defined based on quintic polynomial functions over time. Based on the cost evaluation among the sampled candidates, the optimal lateral motion pattern is selected as a reference and tracked by the controller. The longitudinal motion is planned and controlled using Model Predictive Control (MPC) which is an optimal control method designed considering the surrounding traffic information. To realize the lane change motion similar to the human driving behavior in the surrounding traffic situation, the human driving pattern is modeled in the form of motion parameters and considered in planning the lateral and longitudinal motion. The motion parameters related to the lane change motion are estimated based on the host vehicle states and surrounding vehicle states, and reflected in the optimization of the lateral and longitudinal motion. The estimator of the motion parameters is constructed using a multi-layer neural network which is designed as a regression function. The neural network is trained using real-world motion information related to the lane change maneuver. The proposed trajectory planning algorithm has been validated via simulation tests in which the dynamic traffic environment is reconstructed. The simulation results show that the proposed trajectory planning algorithm secures collision safety and ride quality in multi-vehicle traffic environment. Also, the motion characteristics induced by the proposed algorithm are shown to be consistent with human driving patterns.