Smart Pothole Net: Automated Pothole Detection, Severity Estimation, and Repair Verification System

Abstract

Road safety and vehicle longevity are greatly impacted by potholes, especially in unfavorable environmental circumstances like rain and dim lighting. This work describes a real-time, deep learning-based system that uses dashcam video feeds to detect and map potholes. In order to improve robustness under a variety of real-world settings, the suggested method uses an optimized YOLOv8l object detection model that was trained on an augmented dataset that combined the Road Damage Dataset (RDD), the BharatPotHole dataset (7,074 pictures), and additional custom-collected samples. To the best of our knowledge, our effort is among the first endto-end frameworks that combine lifecycle tracking, geo-location, and pothole detection into a single pipeline. For precise spatial localization, the system uses Optical Character Recognition (OCR) to extract timestamps from video frames and synchronize them with GPS data. Real-time road condition monitoring is made possible by the storage and visualization of detected potholes on a web-based mapping platform. To differentiate potholes from visually similar phenomena like shadows, oil stains, and road patches, a false-positive reduction method is created. Pothole lifecycle events, such as detection, verification, and repair updates, are also tracked using a temporal analysis module.