Abstract:
Autonomous driving heavily relies on accurate trajectory prediction to optimize route planning
and enhance vehicle safety. Current deep learning-based trajectory models have demonstrated
remarkable success on public datasets but often fall short in real-time applications due to
computational limitations in vehicles. In this research, we propose LaneFormer, an optimized
trajectory prediction framework designed to balance high predictive accuracy with
computational efficiency, ensuring its suitability for real-time deployment in autonomous
systems. Our model introduces an efficient attention mechanism to capture complex interactions
between agents and road structures, outperforming state-of-the-art methods while using fewer
resources. We evaluate LaneFormer on the Argoverse dataset, demonstrating its robustness in
predicting future trajectories with competitive metrics across multimodal scenarios.
