An adaptive client selection method for long tailed scene classification in autonomous driving

Area of Science:

• Autonomous Driving Systems
• Machine Learning
• Computer Vision

Background:

• Federated Learning (FL) in autonomous driving struggles with imbalanced, non-IID data, particularly for environmental scene classification.
• Long-tailed data distributions and perceptual imbalances limit the effectiveness of current FL approaches for real-world driving scenarios.

Purpose of the Study:

• To propose FedRare, a novel client selection framework for federated learning in autonomous driving.
• To enhance perceptual robustness and safety in autonomous driving systems operating in complex, long-tailed environments.

Main Methods:

• Developed FedRare, a client selection framework utilizing a multi-dimensional utility function.
• Integrated scenario criticality, distribution rarity, and local update quality into the utility function.
• Implemented a granular evaluation protocol categorizing 18 driving scenarios into standard, diverse, and safety-critical groups.

Main Results:

• FedRare achieved a mean recall of 72.82% in late training stages, surpassing the baseline's 63.42% on the BDD100K dataset.
• Achieved 74.13% recall in safety-critical scenarios (e.g., rainy/snowy nights), demonstrating improved performance in edge cases.
• The framework stabilized training, maintaining performance fluctuation (standard deviation) around 0.02.

Conclusions:

• FedRare effectively addresses challenges posed by long-tailed data and perceptual imbalances in federated learning for autonomous driving.
• The proposed framework enhances perceptual robustness and safety, particularly in critical and rare driving scenarios.
• FedRare offers a promising solution for developing more reliable autonomous driving systems in diverse real-world conditions.