Publication

Abstract : Intelligent transportation systems (ITS) have become a critical and useful tool to execute contemporary traffic management and planning strategies for the governments and municipalities. Traditional automatic number plate recognition (ANPR) systems often struggle with different conditions such as varying lighting, angles, and occlusions, expecting the need for more robust and affordable deep learning supported solutions. This paper proposes a highly effective ANPR system for ITS that relies on a Transformer based approach of deep learning. Our study highlights the global feature extraction of vision transformers (ViTs) to develop an ANPR system that has three main operational components; license plate detection using You only look once V8 (YOLO), character segmentation through morphological techniques, and alphanumeric recognition with an optimized ViT model. We evaluate our system on established datasets such as UFPR-ALPR and OpenALPR benchmark under various vehicle types, illumination, and angle. The results indicate our system performs exceptionally well, achieving a license plate detection rate of 98.72%, a character recognition rate of 97.84%, and an end-to-end system accuracy of 96.91%. Further, we achieve strong detection results in terms of the average precision measures of detection, a mAP@0.5 of 98.65%, precision of 98.50% recall of 97.92% and F1 of 98.21% at realtime processing speed of 28 FPS on an NVIDIA RTX 3060 GPU. The proposed model outperforms existing techniques and offers highly effective and scalable ANPR solution for ITS.