Diff-Pre: A Diffusion Framework for Trajectory Prediction.

Area of Science:

• Intelligent Transportation Systems
• Machine Learning for Autonomous Driving
• Computer Vision for Traffic Analysis

Background:

• Accurate vehicle trajectory prediction is vital for intelligent transportation systems, road safety, and traffic efficiency.
• Existing models struggle with complex, high-interaction traffic scenarios like lane changes and overtaking.

Purpose of the Study:

• To propose a novel trajectory prediction model integrating a diffusion framework with vehicle trajectory and intention features.
• To improve the accuracy and robustness of vehicle trajectory prediction, especially in complex traffic situations.

Main Methods:

• Utilized a diffusion model framework incorporating target and neighboring vehicle trajectories, and driving intentions.
• Employed Long Short-Term Memory (LSTM) networks for temporal feature extraction.
• Integrated a multi-head attention mechanism for dynamic interaction modeling and an intention module for lateral offset regulation.

Main Results:

• The proposed model achieved superior performance over representative methods, evidenced by lower Average Displacement Error (ADE) and Final Displacement Error (FDE).
• Demonstrated enhanced robustness and predictive accuracy in high-interaction and uncertain scenarios, including lane changes and overtaking.
• Achieved an average ADE of 0.199 m and average FDE of 0.437 m within 1 to 5 seconds prediction horizons.

Conclusions:

• The diffusion framework offers an effective and efficient solution for vehicle trajectory prediction.
• This work represents the first application of diffusion frameworks to vehicle trajectory prediction, opening new research avenues.
• The model provides a significant advancement for ensuring road safety and optimizing traffic efficiency in intelligent transportation systems.