Adaptive trajectory prediction without catastrophic forgetting

摘要

Pedestrian trajectory prediction is a necessary component of autonomous driving technology. However, current methods face two troubles when utilized to the actual world, one is the distribution difference between training and testing environments, and the other is catastrophic forgetting. These two issues will lead to an inevitable drop in the overall performance of the model in real-world scenarios. To tackle these two issues, we propose a framework that consists of modules for domain adaptation and continual learning. Specifically, a pedestrian interplay modeling method based totally on pedestrian social habits is proposed. Moreover, we add a domain adaptation module to analyze the data distribution difference between the source domain and the target domain, so as to alleviate the domain difference problem. Finally, a continual learning module is introduced to retain the information which is learned to limit the change of model parameters to deal with the catastrophic forgetting. We design trajectory prediction experiments that conform to real-world activities, and the experimental results verify the superiority of our proposed model. To the best of our knowledge, we are the first work that attempts to apply domain adaptation and continual learning methods to remedy real-world trajectory prediction problems.