Deep Learning-Based Pothole Detection System with Aerial Image

A pothole is typical road damage with a pot shape of 15mm or more depth. A large pothole requires rapid maintenance because it causes a traffic accident. In this paper, we propose an aerial image-based pothole detection system using drones and deep learning. Since drones are not affected by traffic conditions, they can shoot large areas faster than shooting while driving on the road. The proposed system uses deep learning to detect potholes in drone images. For pothole detection, a deep learning classification model based on the Inception v3 model learns a self-constructed aerial image-based pothole data set. To evaluate the pothole detection performance of the model, we measure recall and precision using about 5,496 image datasets. The proposed system is implemented as a web service system that can be used in the real environment by utilizing the developed deep learning model. When the implemented service system is applied to actual road maintenance, it is shown that it is possible to detect a pothole in the broader area in a shorter time than the existing pothole detection method and provide information to the road repairer.