A Federated Perception Architecture for Robust Safety-Relevant Object Detection in Autonomous Vehicles

Reliable object detection is fundamental for Autonomous Vehicles (AVs), yet achieving adaptive and stable perception in dynamic traffic environments remains an open challenge. Federated learning offers a privacy preserving alternative to centralized training, but its robustness in multi-vehicle perception systems is not yet well understood. This paper reports on the design and experimental evaluation of a Federated Object Detection (FOD) system implemented in the CARLA simulator. Multiple autonomous vehicles collaboratively adapt YOLOv8 based detectors under non-IID local data distributions using a two step detection strategy that combines a pretrained model for common objects with a locally adapted pole detector. Ten autonomous vehicles are deployed in CARLA, each equipped with cameras. Vehicles periodically retrain local models and aggregate updates using a FedProx enabled federated averaging. Systemlevel experiments show that federated aggregation improves training stability and consistency across vehicles. These results highlight the potential of federated learning for cooperative perception and motivate future work toward more robust deployment in autonomous driving systems.