Enhanced Pothole Detection Using YOLOv5 and Federated Learning

Ensuring road safety is a vital concern that impacts lives, communities, and economic stability. However, prevailing challenges in road safety pose threats and disparities in justice. Addressing road hazards, particularly potholes, is crucial to averting accidents and damages. It causes vehicle damage and safety risks due to sudden road depressions, especially in the case of water-filled or submerged potholes, leading to skidding, hydroplaning, and potential accidents. This study introduces an artificial intelligence (AI)-driven architecture for robust pothole detection. We detected potholes through the You Only Look Once (YOLOv5) model with a remarkable 83% precision. Then, we deploy a specialized federated learning-based convolutional neural network (CNN) model for accurate pothole identification. Federated learning is used for privacy-preserving reasons, as the image data may contain location-sensitive data. The utilized model achieves an accuracy of 85.71% for the pothole classification task into two categories, i.e., submerged and dry. Detected pothole data is securely cataloged, and based on the conditions and graveness, effective evasive maneuvers are conducted to ensure driver and vehicle safety. The proposed model not only bolsters road safety but also expedites repair processes, minimizing accidents and vehicular damage.