Federated Learning for Robust Vehicular Object Detection with Heterogeneous Data Handling

Federated Learning (FL) can enable continuous online training of perception models for autonomous driving while preserving data privacy. However, data heterogeneity among vehicles poses challenges to FL performance. This work introduces FedProx+LA, a novel FL method that combines proximal terms and label-aware aggregation to address data heterogeneity, leading to superior convergence rates and object detection performance compared to baseline methods.