This work presents the detection and recognition of vehicle identification plates using computer vision techniques applied to traffic enforcement. A proprietary dataset with Brazilian license plates was created, including the steps of recording, selection and annotation of images, combined with an international dataset for training variants of the YOLO model, followed by an analysis of the overall performance of these models. In addition to license plate detection, optical character recognition (OCR) was performed with the EasyOCR and PaddleOCR models, the latter being the most efficient. The experiments showed that the YOLOv8s-gb model outperformed YOLOv5su-g in average confidence, average sensitivity
and processing time. The combination of PaddleOCR with YOLOv8s-gb and YOLOv5su-g applied to a total of 460 license plates resulted in the recognition of 244 and 208, respectively, while EasyOCR recognized 118 and 89 license plates in the same scenarios. The study highlights the importance of specific datasets to improve computer vision models in local contexts, contributing to the advancement of automatic license plate recognition in Brazil.
|
