Motion Prediction for Autonomous Driving with Drone Traffic Monitoring

Abstract

Motion prediction is essential for safe and reliable decision-making in autonomous driving and robotic systems. Although severe performance degradation is usually expected when a model is deployed to a new environment with different road structures and traffic patterns, collecting onboard driving logs and annotating High-Definition (HD) maps for every new city takes tremendous effort and time. Therefore, this work investigates the potential of using drone-based traffic monitoring as an efficient and scalable data source for transferring motion forecasting models to unseen environments. With occlusion-free drone observations, any traffic participant can be selected as the ego vehicle for training motion forecasting models, yielding high-quality data at much larger volume and with a more diverse distribution. We introduce Songdo Drive, a motion forecasting dataset built from 137 hours of drone traffic monitoring over 20 complex intersections in Songdo, South Korea. Experiments show that a zero-shot model trained on an existing autonomous-driving benchmark degrades significantly when evaluated on Songdo Drive, yet the performance can be substantially improved by leveraging the advantage of drone-based trajectory data.