Abstract
Intelligent Transportation Systems (ITS) face synergistic challenges of data privacy leakage, cross-regional adaptation barriers, insufficient human-centric design, and unstandardized
performance assessment. To address these issues, this paper proposes FedHTS, a
Federated Human-Centric Traffic System integrating multimodal intelligence and
standardized evaluation. FedHTS leverages federated learning (FL) for privacy-preserving
cross-regional collaboration, incorporates human-centric perception and interaction modules, and adopts standardized performance metrics. Specifically, we design a dynamic sparse
federated learning framework to reduce communication overhead while supporting few-shot
cross-regional adaptation. A multimodal decision engine fuses Neuro-VAE-Symbolic traffic
dynamics modeling with context-gated spoken language understanding and zero-shot
personalized recommendation. Human factors are integrated via VR-based control strategy
simulation and driver state inference (informed by neural imaging insights). Performance is
evaluated using super-efficiency SBM-DEA combined with neural network regression. Experiments on METR-LA and PEMS-BAY datasets demonstrate FedHTS outperforms
baseline FL models by 18.7% in traffic prediction accuracy and reduces communication cost
by 62.3%. Subjective evaluations via VR simulation show 89% user satisfaction with human- centric services. FedHTS provides a comprehensive solution for next-generation ITS by
unifying privacy protection, cross-regional adaptation, and human-centric design.
