Where Should Federated Learning Run? Design Rules for CCAM Cooperative Perception

Federated learning (FL) promises to keep perception models up to date without shipping terabytes of raw video to the cloud. This is a key enabler for Cooperative, Connected and Automated Mobility (CCAM) where vehicles, roadside units and cloud servers learn jointly. Yet latency, bandwidth, energy draw and trust depend as much on where computation runs as on the algorithm itself. This paper provides a comprehensive mapping of the FL deployment landscape in the vehicular environment. We define the performance, security and privacy requirements for real-time perception. We analyse six concrete deployments (two vehicle-centric, one infrastructure-centric and three hybrid) detailing their training and inference pipelines. A qualitative comparative analysis shows that each deployment excels in one area while conceding ground in another. Lessons learned and identified research avenues (context-aware mode switching, incremental snapshots, slice-aware compression and auditable unlearning) outline a roadmap toward deployable, energy-conscious and privacy-preserving CCAM perception.