<![CDATA[Abstract – Road damage and road-control objects, particularly potholes and speed bumps, can affect driving comfort and traffic safety. Conventional inspection depends on direct observation and can require substantial time and human resources. This study evaluates object detection approaches based on one-stage and two-stage detectors for automatically identifying potholes and speed bumps. YOLOv11 and Single Shot Detector (SSD) are evaluated as one-stage detectors, while Faster R-CNN is evaluated using ResNet50, ResNet101, MobileNetV2, and MobileNetV3 backbones. A conventional image-processing approach is also included as a baseline. The dataset consists of 790 images containing two object classes, pothole and speedbump, and is divided into training, validation, and testing subsets. Evaluation uses mAP@0.5:0.95, mAP@0.5, mAP@0.75, recall, confusion matrix, and qualitative testing on images and video. The results show that YOLOv11 achieves the highest overall performance with mAP@0.5:0.95 of 49.30%, mAP@0.5 of 85.90%, mAP@0.75 of 47.40%, and recall of 84.30%. Among Faster R-CNN backbones, MobileNetV3 provides the best performance with mAP@0.5:0.95 of 45.37%, mAP@0.5 of 83.74%, mAP@0.75 of 40.85%, and recall of 53.34%. The conventional image-processing approach obtains substantially lower results. Overall, YOLOv11 provides the best balance of detection performance and real-time capability for the dataset used in this study. Keywords – pothole, speedbump, object detection, YOLOv11, SSD, Faster R-CNN, deep learning.]]>
Copyrights © 2026
