ANPR - Automatic Number Plate Recognition

A real-time vehicle detection, tracking, and license plate recognition system based on RF-DETR with PaddleOCR for Arabic and English text extraction.

Features

• Vehicle Detection: RF-DETR based detection exported to ONNX for cars, motorcycles, buses, and trucks
• Multi-Object Tracking: ByteTrack for robust vehicle tracking across frames
• License Plate Detection: Fine-tuned RF-DETR model and exported to ONNX for plate localization
• Dual-Language OCR: PaddleOCR with Arabic and English support
• Parallel Processing: Non-blocking OCR runs in background threads
• Zone-Based Detection: Plates detected only in configurable regions (bottom half by default)
• CSV Export: Automatic export of OCR results with timestamps

Installation

# Clone the repository
git clone https://github.com/yourusername/ANPR.git
cd ANPR

# Create virtual environment
conda create -n anpr python=3.10
conda activate anpr

# Install dependencies (CPU)
pip install -r requirements.txt

# For GPU support
pip uninstall onnxruntime paddlepaddle
pip install onnxruntime-gpu paddlepaddle-gpu

Usage

Basic Usage

# Process video (CPU, no display)
python main.py --source video.mp4 --output output.mp4

# Process video with display
python main.py --source video.mp4 --output output.mp4 --show true

# Process video with GPU acceleration
python main.py --source video.mp4 --output output.mp4 --show true --gpu true

# Process webcam
python main.py --source webcam --show true

Command Line Arguments

Argument	Default	Description
source	required	Video file path or webcam
output	None	Output video path
show	false	Display output window (true/false)
gpu	false	Use GPU acceleration (true/false)
model	output/inference_model.onnx	Vehicle detection model
plate-model	output/licence_plate_inference_model.onnx	Plate detection model
confidence	0.5	Vehicle detection confidence
plate-confidence	0.2	Plate detection confidence
input-size	384 384	Model input dimensions
track-thresh	0.25	Tracking activation threshold
track-buffer	30	Lost track buffer (frames)
match-thresh	0.8	Tracking match threshold
no-trace	false	Disable trajectory visualization
no-plates	false	Disable plate detection
cropped-folder	cropped/	Folder for plate crops
max-plates	3	Max plate detections per vehicle

Process Existing Cropped Plates

python plate_ocr.py

This processes all images in the cropped/ folder and saves results to ocr_results.csv.

Project Structure

ANPR/
├── main.py              # CLI entry point
├── processor.py         # Main video processing pipeline
├── vehicle_tracker.py   # Vehicle detection + ByteTrack tracking
├── plate_detector.py    # License plate detection
├── plate_ocr.py         # PaddleOCR wrapper (Arabic + English)
├── ocr_worker.py        # Parallel OCR processing
├── base_detector.py     # Shared ONNX inference logic
├── visualizer.py        # Annotation and display utilities
├── config.py            # Configuration constants
├── requirements.txt     # Dependencies
├── output/              # ONNX models
│   ├── inference_model.onnx
│   └── licence_plate_inference_model.onnx
└── cropped/             # Saved plate crops + OCR results
    └── ocr_results.csv

Output

Cropped Plate Images

Saved to cropped/ folder with naming convention:

{tracker_id}_{detection_count}_f{frame_number}.jpg

Example: 51_2_f153.jpg (Tracker #51, 2nd detection, frame 153)

OCR Results CSV

cropped/ocr_results.csv contains:

Column	Description
timestamp	ISO format timestamp
frame_num	Video frame number
tracker_id	Vehicle tracking ID
plate_text	Combined OCR result
plate_text_en	English OCR result
plate_text_ar	Arabic OCR result
confidence	OCR confidence score
vehicle_box	Vehicle bounding box [x1,y1,x2,y2]
image_path	Path to cropped plate image

Configuration

Edit config.py to modify:

# Vehicle classes to detect (COCO IDs)
VEHICLE_CLASS_IDS = [3, 4, 6, 8]  # car, motorcycle, bus, truck

# Model paths
DEFAULT_MODEL_PATH = "output/inference_model.onnx"
DEFAULT_PLATE_MODEL_PATH = "output/licence_plate_inference_model.onnx"

# Detection settings
MAX_PLATE_DETECTIONS_PER_VEHICLE = 3
CROPPED_FOLDER = "cropped"

Requirements

• Python 3.10+
• OpenCV
• PyTorch
• ONNX Runtime
• PaddlePaddle + PaddleOCR
• Supervision (ByteTrack)
• RF-DETR

See requirements.txt for full dependencies.

Workflow

    Video Frame
        │
        ▼
┌─────────────────┐
│Vehicle Detection│  (RF-DETR ONNX)
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  ByteTrack      │  (Multi-object tracking)
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Zone Filter     │  (Bottom half of frame)
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Plate Detection │  (ONNX, max 3 per vehicle)
└────────┬────────┘
         │
         ├──► Save cropped image
         │
         ▼
┌─────────────────┐
│ Parallel OCR    │  (PaddleOCR: Arabic + English)
└────────┬────────┘
         │
         ▼
    CSV Results

License

MIT License

Acknowledgments

• RF-DETR - Vehicle detection
• Supervision - ByteTrack tracking
• PaddleOCR - Text recognition