Robust Long-Term Vehicle Trajectory Prediction Using Link Projection and a Situation-Aware Transformer.

Area of Science:

• Intelligent Transportation Systems
• Autonomous Driving
• Machine Learning

Background:

• Vehicle trajectory prediction is crucial for Intelligent Transportation Systems (ITS).
• Urban environments with intersections and traffic signals complicate accurate long-term trajectory forecasting.
• Accumulated errors in long-term predictions lead to significant inaccuracies and off-road deviations.

Purpose of the Study:

• To develop a robust long-term vehicle trajectory prediction method resilient to error accumulation.
• To prevent predictions from deviating from actual road geometry.
• To introduce a novel evaluation metric for trajectory prediction accuracy.

Main Methods:

• Utilized the Transformer model for analyzing and forecasting vehicle trajectories.
• Proposed an extra encoding network to capture external factors' influence on driving patterns.
• Implemented a post-processing 'link projection' method to ensure predictions stay on the road.
• Introduced the Area-Between-Curves (ABC) metric for evaluating trajectory similarity.

Main Results:

• The proposed method demonstrates robustness against error accumulation and off-road predictions.
• The Transformer model with the additional encoding network effectively captures external influences.
• The link projection method successfully guides predictions onto road geometry.
• The ABC metric provides a more comprehensive evaluation of trajectory accuracy.

Conclusions:

• The novel trajectory prediction method significantly outperforms conventional deep learning models.
• Achieved improvements of up to 65.74% (RMSE), 60.13% (MAE), and 91.45% (ABC) on real-world datasets.
• The proposed approach offers a more accurate and reliable solution for long-term vehicle trajectory prediction in complex urban environments.