Real-Time Road Sign Detection with YOLOv8 on a Custom Dataset: Exploratory Grounding DINO Comparison

Road sign detection is critical for high-level automated driving (Advanced Driver Assistance) and fully autonomous vehicle development as it is necessary that vehicles can detect the relevant traffic signals/road signs rapidly and accurately under variable traffic flow and illumination conditions. In this study we present a YOLOv8 based road sign detector which was trained on a custom data set consisting of 1,000+ images representing four safety-relevant road sign types: speed limit, stop, traffic light and cross walk. These images were obtained from heterogenous sources of web-based imagery, hand annotated using CVAT, resized to 640x640 pixels, and processed through a data preparation pipeline that involved transfer learning, the identification of a best combination of augmentations for the specific task of road sign classification, and hyper parameter tuning. Before settling on a final configuration for the augmentations (mosaic, mixup, hsv perturbation, shear, and translation), the evaluation of candidate augmentations was performed on a separate validation set of 88 images containing 133 annotated road sign instances. For the validation set, the road sign detector achieved an average precision at 50% confidence threshold (mAP50) of .935, and an average precision at 95% confidence threshold (mAP50-95) of .760, with the highest precision being achieved by the road sign classifier for speed limit and cross walk signs. The archived run record report indicated that the total processing time per image was approximately 17.3 milliseconds, indicating potential for real-time operation at this level of complexity. However, traffic light classification continued to be the weakest category for both recall and localization accuracy. In addition, a qualitative comparison of our results to those of the Grounding DINO demonstrated accurate localization of the sign region but poor class separation, particularly for categories that are visually similar.