Vehicular Federated Learning for Pedestrian Detection Under Adverse Image Conditions

Pedestrian detection is a critical task for ensuring safety in autonomous driving and advanced driver assistance systems (ADAS), as accurate and rapid detection can prevent accidents and protect vulnerable road users. This paper evaluates the performance of Federated Transfer Learning for pedestrian detection using fine-tuning on a pre-trained model. Federated learning enables decentralized training on user devices. We compare its performance with centralized learning under challenging conditions, including small datasets, rainy images, high-speed vehicle captures, and noisy annotations. Experimental results show that while both models suffer under these conditions, the use of client data in federated learning can enhance the size of the total data used for training across clients. Consequently, this results in federated learning having a more consistent and overall better performance than the centralized approach, with a maximum performance gap of 36.42% observed in high-speed scenarios.