Traffic & License Plate Recognition System (ALPR)

This project is an Automatic License Plate Recognition (ALPR) system designed to detect, track, and recognize license plates from traffic video footage.

It combines YOLO26 for object detection and tracking, a custom YOLO-Pose model for precise license plate keypoint estimation, and PaddleOCR for high-accuracy text recognition. The system handles perspective distortion using homography and logs all data into an SQLite database.

Key Features

• Real-time Tracking: Uses YOLO26 + ByteTrack to detect and track vehicles (cars, trucks, buses, motorcycles) with persistent IDs.
• Smart ROI Filtering: Processes vehicles only when they cross a specific virtual line and meet size thresholds.
• Pose Estimation: Utilizes a custom Keypoint Detection model to identify the exact 4 corners of the license plate.
• Perspective Correction: Applies Warp Perspective (Homography) to flatten angled or skewed plates for better readability.
• Heuristic Error Correction: Automatically corrects common OCR errors (e.g., swapping '0'/'O', '1'/'I') based on standard Turkish license plate formats.
• Asynchronous OCR: Offloads heavy text recognition tasks to a background thread to maintain high video FPS.

System Pipeline (Step-by-Step Visualization)

The following pipeline demonstrates how a single vehicle is processed from detection to data logging.

Step 1: Detection & Tracking

The system detects the vehicle in the video stream and assigns a unique Track ID.

Step 2: Vehicle Cropping & ROI

Once the vehicle crosses the virtual line, it is cropped from the main frame.

Step 3: Pose Estimation (Keypoints)

The Pose Model analyzes the crop to find the 4 corners of the license plate.

Step 4: Perspective Warping & Processing

Using the keypoints, the plate is warped into a flat rectangle and sharpened.

Step 5: OCR & Result

PaddleOCR reads the text, applies error correction, and logs the result.

Result: 03AFS621 (Confidence: 0.99)

Step 6: Data Logging

Final results are stored in a local SQLite database for historical analysis.

---

📈 Model Performance & Training

The custom YOLO-Pose model was trained to achieve high precision in license plate localization. Below are the training metrics, including loss curves (box, pose, classification) and mAP (mean Average Precision) scores.

---

📂 Project Structure

📁 Traffic-ALPR/
│
├── main.py                           # Main application script
├── traffic_data.db                   # SQLite Database (Auto-generated)
├── yolo26n.pt                        # YOLO26 Object Detection Model
│
├── 📁 test_videos/
│   └── video-1.mp4                   # Input video
│
└── 📁 pose_results/keypoints_model/
    └── best-2.pt                     # Custom License Plate Pose Model

🛠️ Quick Start

1. Clone the repo: git clone https://github.com/Nislive/license_plate_recognition.git
2. Install requirements: pip install -r requirements.txt
3. Run the system: python main.py

⚖️ License

This project is licensed under the MIT License.