Pothole Segmentation and Area Estimation with Deep Neural Networks and Unmanned Aerial Vehicles

In this research, we explore the problems of pothole detection, segmentation, and area estimation using deep neural networks and unmanned aerial vehicles (drones). We start by compiling two datasets, one that contains ground-level and aerial images of potholes, and another that only contains ground-level images, and we train a total of six deep neural network models for pothole detection; we do this to determine whether aerial images are necessary for training UAV-based object detection models. We then determine which pothole detection model is the most accurate and we also determine which combinations of camera angle and UAV altitude are best for detecting potholes. Furthermore, we take the strongest pothole segmentation model and apply it to area estimation using a combination of homography, the intrinsic and extrinsic parameters of the UAV camera, and novel methods. Our method for pothole area estimation using YOLOv8 has an average area estimation error of 9.71%.