Road-Aware Trajectory Prediction for Autonomous Driving on Highways.

Area of Science:

• Computer Science
• Robotics
• Artificial Intelligence

Background:

• Autonomous vehicles require accurate long-term trajectory prediction for safe operation.
• Deep learning methods have been used for trajectory prediction but often neglect road geometry constraints.
• Vehicles must adhere to road shapes, making road-aware prediction crucial.

Purpose of the Study:

• To develop a novel road-aware trajectory prediction method for autonomous vehicles.
• To leverage high-definition maps and deep learning for enhanced prediction accuracy.
• To ensure predicted trajectories are feasible and realistic by incorporating road structure.

Main Methods:

• Developed a data-efficient learning framework using a curvilinear coordinate system.
• Implemented a lane assignment for surrounding vehicles.
• Proposed an output-constrained sequence-to-sequence network incorporating road structural constraints into the prediction and loss functions.

Main Results:

• The method effectively uses road structural constraints as prior knowledge.
• Incorporated constraints into both the trajectory prediction network and the maneuver recognition loss function.
• Demonstrated data efficiency and ability to predict reasonable trajectories, particularly at merging sections.

Conclusions:

• The proposed road-aware trajectory prediction method enhances realism and feasibility.
• Leveraging HD maps and structural constraints improves prediction accuracy in complex scenarios.
• The approach is data-efficient and suitable for real-world autonomous driving applications.