RT-DETR-Pothole: Lightweight Real-Time Detection Transformers for Improved Road Pothole Detection

Abstract Assessment of on-time road condition is crucial for ensuring the safety of the motorist. One of the recent approaches to detecting road potholes is to analyze images captured from an unmanned aerial vehicle (UAV). Although the traditional deep learning model could perform accurate detection during offline analysis, there is still a limitation of the available algorithms that could perform real-time evaluation. Therefore, this study proposes a lightweight transformer algorithm, the real-time detection transformer (RT-DETR), for online evaluation of road pothole images. The models were tested in practical deployment scenarios and compared with several other object detection models, such as Faster RCNN- SqueezeNet, YOLOv8x, YOLOv9e, YOLOv10x, and YOLO11x. The results show that the RT-DETR- Pothole outperformed all other models in detection accuracy, achieving the highest mAP0.50 (0.834) and mAP0.50-0.95 (0.565), along with a high F1-Score (0.809), indicating superior precision and recall, and at the same time it could maintain low inference time. Overall, RT-DETR-Pothole is the most suitable model for real-time pothole detection, especially for detecting smaller, less visible potholes, with a reasonable inference time for pavement engineering applications.