Communication-Efficient Multi-Level Decentralized Federated Learning for Trajectory Prediction

Forecasting future trajectories of pedestrians and vehicles is necessary for safety and efficiency maximization in connected and autonomous vehicle (CAV) networks. Existing federated learning (FL) approaches face challenges related to scalability, communication overhead, and privacy preservation. To address these challenges, we introduce a multi-level Hierarchical Decentralized Federated Learning framework tailored for trajectory prediction.Our approach organizes clients into a layered hierarchy, where communication is restricted to parent and child nodes, and synchronization across layers is both controlled and periodic. This design reduces redundant message exchanges while maintaining model consistency. We evaluate our method on two real-world trajectory datasets, Intersection Drone Dataset (inD) and Highway Drone Dataset (highD), and show that it achieves prediction accuracy comparable to Centralized Federated Learning (CFL) while reducing communication costs by approximately 25%. Our results demonstrate that hierarchical structuring in decentralized FL offers a scalable, privacy-preserving, and communication-efficient solution for real-world trajectory forecasting.