TrafficVision is an AI-powered traffic violation detection system inspired by Saher. It uses a sophisticated multi-stage YOLO pipeline to automatically detect seatbelt and mobile phone violations, combined with license plate detection and OCR for violator identification.

• Multi-Stage Detection Pipeline

  • Vehicle detection (cars, buses, trucks)
  • Violation detection (seatbelt, mobile phone usage)
  • License plate detection
  • OCR text extraction from license plates

• Multiple Violation Types

  • Person without seatbelt detection
  • Mobile phone usage detection
  • Helmet violations (for motorcycles)

• Flexible Usage

  • Command-line interface (CLI) for batch processing
  • Web interface powered by Gradio
  • Installable via uvx for one-off usage

• Output Options

  • Save violation images with bounding boxes
  • Save license plate images
  • JSON output with violation details
  • OCR text extraction

Process a single image:

uvx git+https://github.com/AFAskar/trafficvision path/to/image.jpg

Launch the web interface:

uvx --from git+https://github.com/AFAskar/trafficvision trafficvision-web

uv tool install git+https://github.com/AFAskar/trafficvision

git clone https://github.com/AFAskar/trafficvision.git
cd trafficvision
uv pip install -e .

Basic usage:

trafficvision path/to/image.jpg

Process a directory of images:

trafficvision path/to/images/

Save all outputs (violation images and license plates):

trafficvision path/to/image.jpg --save

Save only license plate images:

trafficvision path/to/image.jpg --save-plates

Save only violation images with bounding boxes:

trafficvision path/to/image.jpg --save-violations

Launch the Gradio web interface:

trafficvision-web

Then open your browser to the provided URL (typically http://127.0.0.1:7860)

from pathlib import Path
from trafficvision.pipeline import run_pipeline, get_violation_type

# Run the detection pipeline
images = [Path("path/to/image.jpg")]
violation_results, plate_results, ocr_results = run_pipeline(images)

# Get violation types
for violation in violation_results:
    violation_types = get_violation_type(violation.boxes)
    print(f"Detected violations: {violation_types}")

# Get OCR text from license plates
for ocr_text in ocr_results:
    print(f"License plate text: {ocr_text}")

TrafficVision uses a multi-stage detection pipeline:

1. Vehicle Detection: Uses YOLOv11n to detect vehicles (cars, buses, trucks) in the input image
2. Violation Detection: Analyzes detected vehicles for traffic violations using a custom-trained YOLO model
3. License Plate Detection: Locates license plates on vehicles with detected violations using a fine-tuned YOLO model
4. OCR Processing: Extracts text from license plates using Tesseract OCR

The system uses three pre-trained YOLO models:

• yolo11n.pt - YOLOv11 Nano for vehicle detection
• seatbelt.pt - Custom-trained model for violation detection
• license-plate-finetune-v1s.pt - Fine-tuned model for license plate detection

• Python 3.13+
• PyTorch
• Ultralytics YOLO
• Tesseract OCR
• Gradio (for web interface)
• Other dependencies listed in pyproject.toml

The seatbelt detection model was trained on the Seat Belt Detection dataset from Roboflow (CC BY 4.0 license).

Violation 1:
 - Violation Types: ['person-noseatbelt', 'mobile']
 - OCR Text: ABC 1234

The system can generate JSON output with the following structure:

[
  {
    "ABC 1234": {
      "violation_type": ["person-noseatbelt"],
      "violation_bbox": [...],
      "violation_image": "violation_0.png",
      "license_plate": "plate_0.png"
    }
  }
]

Contributions are welcome! Please feel free to submit a Pull Request.

This project is licensed under the AGPLv3 License - see the LICENSE file for details.

• Inspired by the Saher traffic violation detection system
• Built with Ultralytics YOLO
• Dataset from Roboflow Universe
• License Plate Detection Model from morsetechlab
• OCR powered by Tesseract

AFAskar

Current version: 0.1.13