Detection and Recognition of Road Signs by a Computer Vision System Using the YOLO Convolutional Neural Network

Abstract A new version of a neural network model for detecting and recognizing road signs in images and video files using the YOLO convolutional neural network is presented. Based on an analysis of literature sources, the choice was made in favor of using the YOLOv11m convolutional neural network, developed by Ultralytics, in the created system. The article provides a comparative analysis of current YOLO architectures for object detection based on performance and detection accuracy criteria. The YOLOv11m architecture is 2.5 times faster than the most accurate YOLOv11x version, while losing in quality by only 3%, which is an insignificant difference. Based on the analysis, it is shown that YOLOv11m provides a balance between high performance and accuracy of object detection in video. The convolutional neural network, trained for one hundred epochs on data from the Russian Traffic Sign Database (RTSD), achieved a value of 0.927 according to the mAP50 quality metric (mean Average Precision calculated at an Intersection over Union (IoU) threshold of 0.50). The value of 100 epochs was chosen as a balance between training speed and accuracy, as well as the quality of road sign detection in video. The trained neural network model for road sign detection allows detecting images for six unique classes (“mandatory” (blue circles), “danger” (red triangles), “restrictive” (circles with a red border), “main road” (yellow diamond), “service” (rectangles with a blue border), “special regulations” (blue rectangles)) under various weather conditions and at different times of day. Since the obtained recognition system does not provide 100% accuracy, it can be used for driver assistance, but not on autonomous vehicles.