Edge-Optimized Automatic Number Plate Recognition for IoT-Based Smart Parking Using YOLOv8

Automatic Number Plate Recognition (ANPR) is widely used in Intelligent Transportation Systems (ITS) and smart parking applications, but running deep learning-based ANPR directly on low-power edge devices remains difficult because of computation time, memory, and latency limitations. In this study, we develop an edge-oriented ANPR pipeline for an Internet of Things (IoT)-based sensor-triggered stop-and-go smart parking platform, targeting deployment on a resource-constrained edge device. The pipeline combines YOLOv8 for license plate detection, PaddleOCR for text recognition, and a rule-based normalization stage to reduce Optical Character Recognition (OCR) errors caused by spacing inconsistencies and plate-format variations. In the OCR-only ablation study conducted on cropped plate images, PaddleOCR outperformed the other OCR options evaluated, achieving up to 96.0% exact-match accuracy and 98.78% character-level accuracy, with an average OCR-only processing time of 52.55 ms per image. When evaluated as a complete end-to-end pipeline on a Raspberry Pi with ONNX Runtime, the system achieved 83.5% exact-match accuracy and 94.83% character-level accuracy, with an average end-to-end latency of 1713 ms per image, indicating that edge-side operation is feasible for sensor-triggered parking entry and exit events despite CPU-only hardware constraints. In addition to the ANPR module, the proposed platform connects edge devices with Firebase services and Flutter-based user applications for parking status updates, user interaction, reservation matching, and access logs. These results show that a low-cost edge-based ANPR architecture can support practical sensor-triggered smart parking operations without depending on continuous cloud-side inference.