Active Federated YOLOR Model for Enhancing Autonomous Vehicles Safety

A precise and real-time object detection system is crucial to ensuring the safety, smoothness, and trust of Autonomous Vehicles (AVs). Several machine learning techniques have been designed to improve vehicle detection capabilities and reduce the shortcomings caused by limited data and by transferring these data to a central server, which has shown poor performance under different conditions. In this paper, we propose an active federated learning-integrated solution over AVs that capitalizes on the You Only Learn One Representation (YOLOR) approach, a Convolutional Neural Network (CNN) specifically designed for real-time object detection. Our approach combines implicit and explicit knowledge, together with active learning and federated learning with the aim of improving the detection accuracy. Experiments show that our solution achieves better performance than traditional solutions (i.e., Gossip decentralized model and Centralized model).