Implementation of Deep Learning for Pothole Detection With Smartphones

Road potholes pose serious safety risks, contribute to vehicle damage, and increase maintenance costs for road infrastructure. This research develops a smartphone-based pothole detection system using deep learning with the YOLOv8 model, implemented in TensorFlow Lite to enable real-time, on-device inference through an Android application. Images of road surfaces are captured via smartphone cameras, processed locally to detect potholes, and classified into three size categories. Detection results, including bounding boxes, GPS coordinates, and timestamps, are automatically stored in Firebase Firestore and Firebase Storage, and can be monitored through a web dashboard designed for government agencies. Model training used a combined dataset from Roboflow and field-collected images, processed through cleaning, annotation, augmentation, and split for evaluation. Testing across multiple Android devices demonstrated high detection accuracy, achieving 90.91% precision, 96.77% recall, and an F1-score of 93.75%. The system provides a practical and scalable solution for infrastructure monitoring, enabling real-time detection, structured reporting, and cloud integration to support datadriven road maintenance and enhance public safety.