Class-Incremental Motion Forecasting
Researchers have introduced a novel approach to motion forecasting for autonomous vehicles called class-incremental motion forecasting. This method addresses the challenge of new object classes emerging over time and imperfect perception by predicting future trajectories directly from camera images. The proposed framework adapts to new classes while preventing the loss of previously learned information, utilizing pseudo-labels and an open-vocabulary segmentation model to filter predictions and a replay strategy to retain prior knowledge. AI
IMPACT This research could improve the adaptability and robustness of autonomous driving systems in real-world, dynamic environments.