Safety-Balanced Driving-Style Aware Trajectory Planning in Intersection Scenarios With Uncertain Environment

This paper proposes a two-stage trajectory planning method for self-driving vehicles (SDVs) in intersection scenarios with uncertain social circumstances while considering other traffic participants, which are human-driving vehicles (HDVs) with different driving styles. The mixture-of-experts approach is first utilized to learn from human-driving trajectory data to construct a multimodal motion planner, which uses a Transformer to model the interactions between vehicles by explicitly considering their driving styles to facilitate the integrated network to achieve scene-consistent multimodal trajectory prediction and candidate trajectory generation. Second, based on the generated trajectories for the SDV and the predicted trajectories for the other HDVs, each candidate planning trajectory is evaluated via a safety-balanced value function. After that, the trajectory with the highest value is selected for implementation. Such a method plans a safe and efficient driving trajectory in complex and uncertain scenarios. The experimental results demonstrate the efficiency and effectiveness of the designed method as well as the robustness and reasonableness of the SDVs' maneuver decisions at an intersection considering the behavioral dynamics of HDVs.

Automated summary

This paper proposes a two-stage trajectory planning method for self-driving vehicles (SDVs) in complex and uncertain intersection scenarios, considering interactions with other human-driving vehicles (HDVs) with different driving styles. The method utilizes a mixture-of-experts approach to learn from human-driving trajectory data and construct a multimodal motion planner, which explicitly models the interactions between vehicles and enables scene-consistent multimodal trajectory prediction and candidate trajectory generation. The generated trajectories are evaluated using a safety-balanced value function, and the trajectory with the highest value is selected for implementation. Experimental results demonstrate the efficiency, effectiveness, robustness, and reasonableness of the method in intersection scenarios with HDVs' behavioral dynamics.