Pothole-YOLO: A Single-Stage Instance Segmentation Method for Pothole Detection

Potholes pose a significant hazard to traffic safety and transportation infrastructure. Traditional methods of human-based pothole detection present challenges in terms of cost and safety. However, instance segmentation technology offers a promising solution by accurately locating potholes and providing high-resolution features. This paper addresses the task of pothole detection by introducing a novel single-stage instance segmentation network called Pothole-YOLO. Our approach incorporates an efficient plug-and-play module called CEA-block to enhance the YOLO backbone and layers. Furthermore, we propose a robust and lightweight segmentation head named Pothole-Protonet to improve the segmentation prediction performance. Additionally, we enhance the pothole boundary features by modifying the Wise-IOU method, instead of using the common IOU method. Experimental results demonstrate that our Pothole-YOLO achieves the highest accuracy in pothole detection compared to other state-of-the-art methods on publicly available pothole datasets.