YOLOv5-Pothole: An Improved Pothole Perception Method Based on YOLOv5-Seg

Road pothole damage represents a significant challenge in transportation infrastructure. Numerous studies have aimed to automate pothole detection by employing diverse computer vision techniques. There is a critical need for an automated road pothole detection process that offers sufficient accuracy and speed, while also being cost-effective and straightforward to implement. We introduce an enhanced road pothole detection algorithm, YOLOv5-pothole, based on the YOLOv5-seg model. Inputs for this algorithm include RGB images, depth images, and camera position data, all captured by a D455 depth camera. Concurrently, LiDAR sensors are used to acquire 3D point cloud data. The RGB images are semantically segmented by the self-trained YOLOv5-seg network, and corresponding virtual point clouds are generated through coordinate transformation. Then we register the 3D point clouds with the virtual point clouds generated from 2D pothole images to supplement the semantic information of the pothole. Finally, the pothole point clouds are individually segmented using Euclidean clustering, providing the information of the pothole. The GRDDC2020 experimental dataset was used to train the YOLOv5-seg network. We tested the YOLOv5-pothole algorithm on five randomly selected road potholes, with experimental results indicating a mean detection error of 7.0%. The algorithm demonstrated effective detection across potholes of varying sizes. This study offers valuable insights for future real-time road pothole detection research.