Intelligent License Plate Recognition (ANPR) for Traffic Lights
- Date : 2025
- Computer Vision
- Completed
Offline ANPR system for Tunisian and Libyan plates: YOLOv8 detection with an OpenCV fallback, Arabic/Latin OCR, format validation and real-time video processing.
Overview
Summer internship at SOTECA Electric × Istithmar Equipement, Sfax, from July to August 2025.
I designed and built an offline automatic number plate recognition (ANPR) system tailored to Tunisian and Libyan license plates. It detects plates in images, video files and live camera feeds, reads mixed Arabic/Latin text, validates it against national formats, and logs every detection for traffic monitoring and access control.
Key figures
- plate detection recall on the test set
- 90%+
- plate detection recall on the test set
- countries’ plate formats supported (Tunisia, Libya)
- 2
- countries’ plate formats supported (Tunisia, Libya)
- input sources: image, video file, live camera
- 3
- input sources: image, video file, live camera
Technologies
Python
YOLOv8
OpenCV
EasyOCR
SQLite
TkinterProblem
Commercial solutions (Hikvision, Dahua, OpenALPR) handle Tunisian and Libyan plates poorly because of the Arabic script and regional plate formats.
Solution
A hybrid pipeline: YOLOv8 localizes the plate, with an OpenCV contour-based fallback; multi-pass EasyOCR (Arabic + English) reads it after preprocessing; the text is normalized, corrected and validated against national formats; detections are stored in SQLite, all inside a multithreaded Tkinter desktop app.
My contribution
Designed and built the whole system: detection pipeline, OCR and text correction, format validation, desktop application, persistence layer, and UML modeling (class and use-case diagrams) of a modular three-class architecture.
Results
The detection pipeline reaches over 90% recall on the test set and works fully offline on images, videos and live cameras. Next steps: embedded deployment on a Raspberry Pi 4 with model quantization, field testing at real intersections, and API integration into the company’s traffic management infrastructure.
Features
- Plate detection with YOLOv8 and an OpenCV contour-based fallback
- Multi-pass OCR with EasyOCR (Arabic + English), image preprocessing, Arabic text normalization and OCR error correction
- Regex validation of Tunisian (standard, RS, special series) and Libyan plate formats
- Region-to-text binding using overlap, distance and confidence scores
- Multithreaded Tkinter desktop app for images, video files and live camera feeds
- SQLite logging with CSV export, plate and vehicle images archived by day, automatic deduplication