Dynamic Graph Neural Network for Vehicle Trajectory Prediction and Driving Intent Recognition.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Robotics

Background:

• Existing vehicle trajectory prediction methods struggle with dynamic inter-vehicle interactions, temporal continuity of driving intentions (e.g., lane-changing), and prediction uncertainty.
• Accurate prediction of vehicle movement is crucial for autonomous driving systems and traffic management.

Purpose of the Study:

• To develop an advanced vehicle trajectory prediction method that overcomes the limitations of current approaches.
• To improve the accuracy, temporal continuity, and reduce uncertainty in predicting future vehicle trajectories, especially in complex traffic scenarios.

Main Methods:

• Integration of Dynamic Graph Neural Networks (DyGNN) and Transformer architectures.
• Construction of a time-varying interaction graph to model dynamic inter-vehicle relationships.
• Utilizing a Transformer encoder to capture temporal dependencies in historical trajectory data.
• Incorporating driving intention as a prior constraint to reduce prediction uncertainty.

Main Results:

• The proposed method achieves a joint representation of spatial interactions and temporal evolution.
• Improved accuracy and continuity in recognizing driving intentions within complex traffic scenarios.
• Demonstrated low prediction errors across various prediction horizons on real-world datasets.
• Exhibited good effectiveness and robustness in vehicle trajectory prediction.

Conclusions:

• The integrated DyGNN and Transformer approach significantly enhances vehicle trajectory prediction.
• Modeling dynamic interactions and driving intentions effectively reduces prediction uncertainty.
• The method shows strong performance and reliability for real-world traffic applications.