TrajRS Enhances Certified Robustness for Pedestrian Trajectory Prediction

This paper introduces TrajRS, a significant advancement in ensuring the safety and reliability of autonomous driving systems by improving the robustness of trajectory prediction models. Adversarial attacks can severely compromise these models, leading to inaccurate predictions and potentially dangerous driving behaviors. While existing heuristic defenses offer some protection, they often fall short against sophisticated, targeted attacks. TrajRS addresses this by extending the Randomized Smoothing framework to provide a certified robust radius for smoothed trajectory predictors. The research formally defines and expands upon robustness in trajectory prediction, tailoring the TrajRS scheme to ensure "robustness for the optimal prediction" and "robustness for all possible predictions." Extensive experiments confirm that TrajRS effectively achieves robustness certification for various smoothed pedestrian trajectory predictors, marking a crucial step towards verifiable safety assurances in autonomous vehicles.