Improving Autonomous Vehicles Safety in Snow Weather Using Federated YOLO CNN Learning.

Accurate object detection (e.g., buildings, vehicles, road signs and pedestrians) is essential to the success of the idea of autonomous and self-driving cars. Various object detection techniques have been proposed to enable Autonomous Vehicles (AVs) to achieve reliable safe driving. Most of these techniques are adequate for normal weather conditions, such as sunny or overcast days, but their effectiveness drops when they are exposed to inclement weather conditions, such as days with heavy snowfall or foggy days. In this paper, we propose an object detection system over AVs that capitalizes on the You Only Look Once (YOLO) emerging convolutional neural network (CNN) approach, together with a Federated Learning (FL) framework with the aim of improving the detection accuracy in adverse weather circumstances in real-time. We validate our system on the Canadian Adverse Driving Conditions (CADC) dataset. Experiments show that our solution achieves better performance than traditional solutions (i.e. Gossip decentralized model, and Centralized model).