Collaborative autonomous driving utilizing Federated Learning (FL) is frequently constrained by long-tailed data distributions and perceptual imbalances. Implementing federated learning for environmental scene classification in autonomous driving faces severe challenges due to the long-tailed and non-IID distribution of real-world climatic data. To address this, this paper proposes FedRare, a client selection framework based on a multi-dimensional utility function, integrating scenario criticality (\(W_C\)), distribution rarity (\(W_R\)), and local update quality (\(W_L\)). To address the safety concerns in edge cases, we implement a granular evaluation protocol by categorizing 18 driving scenarios into standard, diverse, and safety-critical groups. Experiments on the BDD100K dataset show that FedRare achieves a mean recall of 72.82% during the late training stage (averaged over the final five communication rounds across five random seeds) while the baseline is 63.42%. Results from the trained model validation demonstrate that FedRare achieves a 74.13% recall in safety-critical groups (e.g., rainy/snowy night), while maintaining high perceptual reliability where traditional loss-based methods often fail. Furthermore, the \(W_L\) constraint stabilizes the training process, keeping the performance fluctuation (standard deviation) at approximately 0.02. This work proposes a framework to enhance perceptual robustness for autonomous driving within complex, long-tailed environments.
