A Comparative Analysis of YOLO-Based Traffic Sign Detections With a Novel Turkish Traffic Sign Dataset

Traffic signs are one of the most critical components that regulate traffic on roads. It informs drivers about the priorities and restrictions on the road they are on. This study aims to provide a solution for detecting 37 frequently encountered traffic signs on Turkish highways through deep-learning-based Yolo models. Detecting signs using computer vision is challenging for classical methods due to environmental conditions, and deep-learning-based methods promise successful results. Worldwide general standards determine traffic signs, but each country’s specific differences also influence their design. We have prepared a new dataset, TraffiSign-Turk, to compare the trained models in this study. The dataset contains 25,978 different labeled objects across 10,561 distinct images. Besides traffic signs, it also includes labels for vehicles and pedestrians in the photos. Using Yolo-v5 and Yolo-v8, we achieved an acceptable level of successful object detection accuracy that operates at online speed on the dataset. These findings have proven that Yolo-based models can be used to detect environmental objects necessary for autonomous driving. An online driver assistance system is developed based on trained models. We have introduced a comprehensive new dataset to the literature for autonomous vehicle studies.