Reconhecimento de imagens: Uso do método Yolo no reconhecimento de placas de trânsito

Abstract

The traffic sign recognition has been an essential part of driver assistance systems, which are able to help drivers avoiding a large number of potential hazards and improving the driving experience. However, recognizing traffic signs is a task that is full of challenges, such as visual environment, physical damage and partial occlusions, among others. To deal with the constraints, convolutional neural networks (CNN) are accommodated to extract the visual characteristics of traffic signs and classify them into corresponding classes. This graduate work, initially intends to recognize traffic signs in Brazil, having as main framework the use of the You Only Look Once (YOLO) method, in its fifth version. Using a single convolutional network that simultaneously predicts multiple bounding boxes and class probabilities for these boxes, YOLO trains on complete images and directly optimizes detection performance. YOLOv5 brings better accessibility for real-time object detection and prediction performance based on training speed or accuracy compared to its previous version. Tests were carried out with three YOLO architectures, small, large and extra large, obtaining accuracies of 92.9%, 92.6% and 92.0%, respectively, in addition to creating a dataset of the Brazilian boards and an user’s interface.