Automatic detection and classification of Brazilian traffic signs using YOLO

Poor traffic sign conditions compromise road safety and due to the huge dimensions of the road network, its manual inspection is an expensive and time consuming task. This work demonstrates the use of the YOLO (You Look Only Once) algorithm for automated detection and classification of Brazilian traffic signs. An initial model was trained on a small dataset of manually analyzed images. The trained model was then used to process a larger set of images extracted from public videos and from the Internet. The resulting enlarged dataset was subjected to manual verification and correction of the initial detections/classifications, generating a final dataset with over 2,000 images (over 6,000 after data augmentation). A final model was then trained on the refined dataset, achieving significant accuracy, recall, and mAP@50 values for the 10 most well represented traffic signs. The simple approach presented herein shows a strong potential for cost-effective, real-world applications in intelligent traffic systems and urban monitoring.