Detection of road attachment facilities using YOLO models

Recognizing damaged road facilities is essential for safe driving, as they can cause secondary traffic accidents. In this study, YOLO, a CNN-based deep learning object recognition model, was used to identify road facilities. YOLO is more suitable for the purposes of our study because the YOLO object detection that uses a bounding box to determine the approximate location of an object rather than segmentation, which is a cell-by-cell recognition of an image. The model was trained on public data and images to classify five classes: Lava cone, Drum, Fence, Standing Board, and Snow Removal Box, and achieved a good performance of mAP50 of 0.971. The processing efficiency can be improved by lightening the model or upgrading the hardware to increase the processing speed. YOLO has several versions from v1 to v8, and each version performs differently depending on the dataset due to differences in the detailed structure. We conducted experiments with the same dataset for v5, v7, and v8, which have significant structural differences, and YOLOv5 showed the best performance in this study.