Abstract
Automatic Number Plate Recognition (ANPR) supports traffic monitoring, access control, parking management and intelligent transportation analytics. A reliable ANPR system must localize plates, transcribe characters, associate repeated detections across frames and store audit-ready records. This paper presents a compact and reproducible ANPR framework that integrates YOLOv8n for plate detection, a Convolutional Recurrent Neural Network with Connectionist Temporal Classification (CRNN-CTC) for sequence recognition, EasyOCR for pseudo-labeling and fallback recognition, and DeepSORT for multi-frame association. YOLOv8n was trained as a one-class detector using 640-pixel images. Detected plate crops were passed to the CRNN recognizer. EasyOCR generated pseudo-labels where manually verified crop text was unavailable and was retained as a fallback recognizer for short CRNN outputs. DeepSORT linked detections across frames for vehicle-level logging. YOLOv8n validation produced precision of 0.980, recall of 0.943, mAP@0.50 of 0.971 and mAP@0.50:0.95 of 0.686 on 2,048 validation images containing 2,195 plate instances. A separate IoU-matched inference evaluation produced 92.32% detection accuracy/ACR, 97.46% precision, 94.58% recall and 96.00% F1 score. The CRNN recognizer achieved 86.61% exact-match training accuracy and 97.58% character-level F1 against EasyOCR pseudo-labels. The proposed workflow demonstrates a practical ANPR pipeline for research-grade intelligent transportation studies. Tracking accuracy is not claimed because temporal identity ground truth was not available.