Robust Vehicle License Plate Detection in Real-World Traffic Scenarios

Abstract

Intelligent transportation systems rely on Automatic Number Plate Recognition (ANPR) to support applications such as traffic surveillance, law enforcement, and automated parking management. However, in Pakistan, the lack of conventional plate sizes, the popularity of elaborate or personalized designs, the variety of fonts, and the range of color schemes make ANPR a difficult process. The accuracy of traditional recognition methods is seriously affected by these variables.

This study proposes a robust ANPR outline precisely de- signed for Pakistani license plates. Convolutional neural networks (CNNs) are used to extract features, optical character recognition (OCR) is used to identify characters, and YOLOv8 and YOLOv3 are used to recognize license plates. The work was tried on 200 images of license plates from motorcycles and cars, and it attained a recognition accuracy of 97.5%. With an Detection accuracy of 96.7%, the experimental results presented that the proposed system was effective in handling the very difficult variations of Pakistani license plates, as well as the improved detecting capacity of up-to-date YOLO versions, such as YOLOv8

Keywords: Convolutional Neural Networks (CNNs); Optical Character Recognition (OCR); Automatic Number Plate Recog- nition (ANPR); YOLOv3; YOLOv8; Intelligent Transportation Systems; License Plate Detection; Pakistani Vehicles.