Pothole Detection and Dimension Estimation Using Deep Learning and Image Processing

Abstract

Road infrastructure directly impacts transportation efficiency, safety, and connectivity, yet potholes—caused by heavy traffic loads, poor construction quality, water infiltration, and extreme weather—remain a persistent problem that leads to vehicle damage, higher maintenance costs, congestion, and accidents. Traditional pothole detection is manual, labour-intensive, error-prone, and infeasible for large road networks, while sensor-based methods suffer from false detections and image-processing methods are sensitive to lighting and texture variation. This paper presents an automated system for pothole detection and dimension estimation that combines a deep-learning object detector with classical imageprocessing measurement. A convolutional-neural-network detector identifies potholes in images and video streams, after which contour detection, pixel-to-real-world scaling, and perspective transformation are used to estimate pothole width, length, and approximate depth so that repair priority can be assessed by severity. GPS integration geo-tags detected potholes for mapping, and detected data can be stored centrally for monitoring through dashboards. The prototype is implemented in Python using the Ultralytics YOLO framework, OpenCV, NumPy, and a desktop interface. By unifying artificial intelligence, image processing, and geospatial tagging, the system reduces dependence on manual inspection, improves road safety, and provides a scalable solution for infrastructure management.