Automatic License Plate Recognition

The application detects and tracks vehicles in video footage, identifies license plates, and uses OCR to read the plate text. It is designed for Swedish license plates, validating them against standard formats such as LLLNNL or LLLNNN (L = letter, N = number).

Models

• YOLOv11m for vehicle detection.
• Fine-tuned YOLOv11n trained on a custom license plate dataset for license plate detection.
• PaddleOCR for text recognition on detected plates.
• OpenCV for video handling and annotations.

Installation

• Python 3.13+
• Install dependencies with uv:

  # For cuda
  uv sync --group cuda

  # For cpu
  uv sync --group cpu

Usage

Run the main script with a video file as input:

uv run app/main.py path/to/video.mp4 [options]

Command-Line Arguments

• VIDEO_SOURCE: Path to the video file. Required.
• --save-video: Save annotated video output (default: False).
• --vehicle-conf: Confidence threshold for vehicle detection (default: 0.6).
• --plate-conf: Confidence threshold for plate detection (default: 0.7).
• --ocr-conf: Confidence threshold for OCR (default: 0.85).
• --ocr-conf-retry-threshold: Max confidence to retry OCR (default: 0.95).

Type --help for more information.

Example

Process a video and save annotated output:

uv run app/main.py input.mp4 --save-video

Output

Results are saved in a timestamped directory in results/ directory:

• results.csv: CSV with columns: Timestamp (s), Plate Text, Confidence, Car Image, Plate Image.
• Cropped vehicle and plate images.
• config.json: Dump of the used configuration.
• output.mp4: Annotated video (if --save-video is used).

Configuration

All defaults are defined in config.py. You can override them via command-line arguments or modify the Config class directly.

• Logging: Set environment variable LOGLEVEL to DEBUG, INFO, etc. (default: INFO).
• Debugging: Enable DEBUG logging to save each prediction attempt in debug/.

License

MIT License