Redes neurais convolucionais de um e dois estágios para detecção e classificação de placas pare e de velocidade

Abstract

This work investigates the application of Convolutional Neural Networks (CNNs) for traffic sign detection and classification, using one-stage and two-stage approaches. The one-stage models, based on YOLOv8, integrate detection and classification into a single flow, prioritizing speed and simplicity. The two-stage models, which incorporate Keras-OCR, add an extra classification stage, increasing accuracy. A customized dataset was developed with images of speed and stop signs, using pre-processing and annotation techniques to improve quality. Metrics such as mAP, precision and recall were used to evaluate the models, demonstrating that one-stage approaches are ideal for real-time applications, while two-stage approaches excel in more complex scenarios. The research highlights the importance of choosing suitable CNN architectures for Intelligent Transportation Systems (ITS) and Advanced Driver Assistance Systems (ADAS), contributing to safer and more efficient autonomous technologies. The final two-stage model has a mAP of 0.95 in the first stage and an accuracy of 0.77 for 8 classes. It also achieves an accuracy of 100% in two of the eight classes: Stop and v20. It also has accuracies greater than 0.93 for two others: v30 and v60.