An Embedded Real-Time Automatic License Plate Recognition System Using YOLO Algorithm

Automatic License Plate Recognition (ALPR) is crucial in Intelligent Transportation System but faces challenges like weather and light conditions, camera angles, and license plate distortion. With advances in deep learning, as well as computing platforms, particularly GPUs, these algorithms have found major applications. The task becomes even more complex with Iranian license plates due to the strong similarity of some Persian characters, and the need for real-time processing is often overlooked. Consequently, this work proposes a two-stage deep learning-based algorithm for ALPR, with impressive precision and real-time applications. The methodology involves License Plate Detection (LPD) and Character Recognition (CR) using separate fine-tuned YOLOv5 networks, extracting characters in two sequential steps. The model shows robustness under challenging scenarios such as uneven lighting, low-quality images, and noise. Experimental results show an end-to-end mean Average Precision (mAP) of 95.5% and an inference speed of 23 Frames Per Second (FPS), meeting real-time requirements. Specifically, a mAP of 98.2% is achieved in the CR stage, effectively addressing character similarity issues. The developed model is implemented on the Jetson Nano, an embedded device, using DeepStream and demonstrates strong performance. For real-time detection, the TensorRT-based model deployed on the Jetson Nano achieved 6 FPS inference speed.