Personalized Federated FlowChain for Human Trajectory Prediction

We introduce Personalized Federated FlowChain (PFF), a method for pedestrian trajectory prediction using federated learning (FL) in realistic settings where data is unevenly distributed across clients. A normalizing flow-based model called FlowChain performs well in centralized training but struggles to generalize across clients with diverse environments and motion patterns. We observe that averaging the full model across clients, particularly the generative flow head, results in unstable training and reduced accuracy. PFF addresses this by sharing only the encoder across clients, while each client keeps its own prediction head. After training, clients further improve their heads through local fine-tuning without any additional communication. This setup allows clients to learn from shared patterns while adapting to their own data. Our evaluation focuses on client performance, which reflects how federated models are actually used in practice, with each client relying on its own predictions. Results show that PFF improves both stability and accuracy compared to standard aggregation and regularization-based approaches.