Adaptive Object Detection in Autonomous Vehicles: Federated Learning for Low-Light Environments

This research examines the potential for surfacing through Federated Learning (FL)-based YOLOv8 functionality toward better object detection in autonomous vehicles (AVs) in low-light conditions. Native challenges to reliable object detection in dimly lit scenarios such as nighttime or tunnels affect the performance and safety of an AV. In this paper, a decentralized training strategy is proposed by FL to allow AVs to collaboratively train the object detection models without compromise on data privacy. The integrations of the proposed FL-based YOLOv8 along with active image enhancement techniques is recommended to improve visibility in dark environments, including contrast adjustment and gamma correction. Extensive experiments conducted allowed the testing of model detection accuracy, computational efficiency, and communication overhead. Compared to centralized learning models, the FL-based YOLOv8 already ‘with-object’ achieves 82.4% mean Average Precision, improved Intersection over Union scores, and a faster average inference time. It also achieves a larger than 68% reduction in communication overhead, hence ensuring scalability in real-world AV networks. The study emphasizes the significance of combining FL with adaptive image processing to further develop and augment the object detection systems under degraded illumination conditions contributing to safer and more reliable autonomous driving systems.