We are pleased to announce that our paper, A Federated Perception Architecture for Robust Safety-Relevant Object Detection in Autonomous Vehicles, was presented by our PhD student Aiman Faiz at CARS@EDCC 2026 (10th International Workshop on Critical Automotive Applications: Robustness & Safety), held in Canterbury, UK, on April 7, 2026. The paper was featured in the session on Architectures for Cooperative and Safe Autonomous Systems. (iks.fraunhofer.de)
About the Work
Reliable object detection is a core requirement for autonomous vehicles, especially in dynamic and safety-critical traffic environments. In this work, we present a federated perception architecture designed to support robust, safety-relevant object detection across multiple vehicles without requiring the exchange of raw sensor data. The system was implemented and evaluated in the CARLA simulator, where ten autonomous vehicles collaboratively adapted YOLOv8-based detectors under non-IID local data distributions.
Our approach combines a general pretrained detector for common road objects with a specialized pole detector tailored to safety-relevant roadside infrastructure. To improve stability during distributed training, we integrated a FedProx-inspired aggregation strategy, helping reduce model drift and improve consistency across vehicles during federated updates.
Key Contributions
- Federated Object Detection for AVs: We designed a collaborative perception pipeline in which multiple autonomous vehicles train shared detection models while keeping raw data local.
- Robustness under Heterogeneous Data: The study explicitly addresses non-IID data distributions, a realistic challenge in multi-vehicle environments where each vehicle experiences different routes, viewpoints, and weather conditions.
- CARLA-Based System Evaluation: The architecture was deployed in a realistic simulation setting using ten autonomous vehicles equipped with RGB cameras, enabling system-level evaluation of training stability and detection consistency.
- FedProx-Inspired Stabilization: Experimental results showed that federated aggregation improved the stability and consistency of detection performance across vehicles operating in the same environment.
Why It Matters
As autonomous driving systems become more connected and data-driven, perception models must remain accurate, adaptive, and privacy-aware. This work shows how federated learning can support cooperative perception while reducing the need for centralized data collection. It also highlights the importance of studying safety-relevant object detection not only in terms of accuracy, but also in terms of robustness, consistency, and deployability in realistic automotive settings.
For more information or to discuss potential collaborations, feel free to contact Aiman Faiz.