Federated Learning for Human-Centric ITS: Integrating Multimodal Interaction and Cross- Regional Adaptation

Intelligent Transportation Systems (ITS) struggle with privacy risks, cross-regional heterogeneity, and inadequate human-centric design—gaps that fragmented research has yet to address. This paper proposes a federated human-centric ITS framework that unifies privacy-preserving collaboration, multimodal interaction, and standardized evaluation. The framework uses dynamic sparse federated learning to reduce communication overhead while enabling few-shot cross-regional adaptation. It integrates context-gated spoken language understanding for natural driver interaction, zero-shot recommendation for personalized travel services, and VR-informed interface design to minimize cognitive load. Performance is assessed via super-efficiency SBM-DEA, ensuring comprehensive measurement of technical efficiency and user experience. Experiments on METR-LA and PEMS-BAY show 18.7% higher traffic prediction accuracy and 62.3% lower communication cost than baseline federated models, with 89% user satisfaction in VR evaluations. This work provides a scalable solution for next-generation ITS by balancing privacy, adaptability, and human-centricity