Distributed Framework for Pothole Detection and Monitoring Using Federated Learning: A Privacy-Preserving Edge Computing Approach

Abstract

The research focuses on tackling challenges in increasing pothole leading to deteriorating road conditions. In reality, object detection faces various computational challenges at edge devices that includes model training, global server aggregation and data privacy. By leveraging federated learning within the FedCV framework, the system monitors, detects, and reports potholes. The novel aspect of research is the adaptation of the YOLOv8 object detection model, integrated with the FedCV framework, to ensure data privacy and security at the central server, while simultaneously boosting performance at the edge device during pothole detection training with local data. This integration addresses major computational challenges at the edge camera device level. As the research combines deep learning and federated learning within an edge-based framework, it improves real-time pothole detection and ensures efficient privacy-preserving road maintenance and advancing smart city initiatives. The model has demonstrated outstanding performance in classification, achieving an average accuracy of 96%. With a precision of 96.4%, recall of 96.7%, and an F1 score of 96.6%, its evaluation underscores its superiority compared to previous studies. These results establish the model as a reliable approach for pothole detection.