Enhancing Generalization of Car-Following Models Via Domain-Incremental Learning

Prediction of car-following behavior plays an important role in the development of autonomous driving. Often, the assumption is that in the traditional learning-based models, both training and test distributions come from the same domain, but the model's performance drastically reduces when exposed to new domains with unseen traffic scenarios at inference time. In this study, we explore continual learning for the prediction of car-following acceleration across multiple real-world driving datasets such as HighD, SPMD, Lyft, Waymo, and NGSIM, where each represents distinct traffic scenes. We implemented Elastic Weight Consolidation (EWC) and Memory Aware Synapses (MAS) techniques as a regularization-based continual learning method and compared their performance with standard training. Our framework includes models with differing prediction targets: NN and Follownet for acceleration prediction, and TransFollower and Trajectron for speed prediction. Each model is sequentially trained on multiple domains and evaluated across all previously encountered domains. Our results show that normal training shows severe catastrophic forgetting, with test errors increasing significantly on previous domains. In contrast, EWC and MAS effectively preserve performance on earlier learned datasets, achieving substantial reductions in forgetting without sacrificing overall accuracy. These findings highlights the importance of continual learning strategies in achieving robust, transferable car-following models across different driving scenarios.