Illuminating autonomy: Federated learning for object detection under low-light conditions

Autonomous vehicles (AVs) are transforming modern transportation systems with their potential for enhanced safety, efficiency, and intelligent decision-making. However, their performance often degrades in low-light environments, posing a significant challenge to real-time object detection—an essential component of AV navigation. This study presents Illuminating Autonomy, a novel framework that integrates Federated Learning (FL) with the YOLOv5 object detection model to address the limitations of traditional centralized training. By enabling multiple autonomous units to collaboratively train a global model without sharing raw sensor data, FL ensures data privacy while improving model generalization across varying lighting conditions. The system was tested on a custom dataset containing both high- and low-light scenarios and five object classes: Car, Bus, Truck, Motorbike, and Person. Experimental results demonstrate that the proposed approach achieves up to 95% detection accuracy in challenging low-light conditions, significantly outperforming conventional methods such as Haar Cascades and HOG+SVM. These findings highlight the potential of combining decentralized deep learning and real-time perception to enhance the reliability and safety of autonomous navigation in low-visibility environments.