safesight-AI

Covid Safety Monitor

Comment

Python PyQt6 OpenCV YOLOv8 ONNX ONNX Runtime License

SafeSight AI is a real-time safety intelligence desktop app for detecting people, masks, and distancing violations from images, videos, or live webcams — powered by YOLOv8 ONNX and OpenCV.

Features • Screenshots • Demo • Installation • Usage • How It Works • Architecture • Tech Stack

Features

  • 🛡️ Multi-mode input: Image, Video File, or Live Webcam
  • ⚡ Real-time YOLOv8n ONNX person detection on CPU (no PyTorch required)
  • 😷 Face mask classification with OpenCV DNN + MobileNetV2 ONNX
  • 📏 Social distancing enforcement with homography + metre calibration
  • 🛰️ Bird's-eye view map with violation lines and labeled dots
  • 📊 Live metrics bar (people, compliance %, mask violations, distance violations)
  • 📈 Streaming violations chart for video/webcam sessions
  • 🎯 Adjustable confidence threshold and minimum safe distance
  • 🧵 QThread-powered pipeline keeps the UI fully responsive
  • 💾 Save annotated images or rendered videos
  • 🧠 Smart fallbacks when a model is missing (no blank screens)
  • 🧩 Clean PyQt6 UI with custom frameless chrome

Screenshots


Main UI overview (current available screenshot)
Detection overlays + metrics (same screenshot until more are added)

Demo / How to Run

python3 main.py

Installation

  1. git clone https://github.com/Vihaan-Singhal1/safesight-AI.git
  2. cd safesight
  3. python3 -m venv venv && source venv/bin/activate (Mac/Linux) OR venv\Scripts\activate (Windows)
  4. pip install -r requirements.txt
  5. mkdir -p models && curl -L "https://huggingface.co/webml/yolov8n/resolve/main/onnx/yolov8n.onnx" -o models/yolov8n.onnx
  6. python3 main.py

Usage

Image Mode

  1. Select Image in the sidebar.
  2. Click Browse… and choose a .jpg/.png/.webp file.
  3. Adjust Confidence Threshold and Min. Safe Distance if needed.
  4. Click Analyse to run detection.
  5. Click Save Output to export the annotated image.

Video Mode

  1. Select Video File.
  2. Click Browse… and choose a .mp4/.mov/.avi/.mkv file.
  3. Click Analyse to start playback with overlays.
  4. Use Stop anytime; click Save Output to export the annotated video.

Webcam Mode

  1. Select Webcam (Live).
  2. Choose the Camera index (0–9).
  3. Click Analyse to start live detection.
  4. Click Stop to end the session; Save Output writes the latest frame/video.

Settings Reference

Setting Type Range Default Effect
Input Source Radio Image / Video / Webcam Image Chooses pipeline mode and source type
Camera Index Spinbox 0–9 0 Selects webcam device ID
Confidence Threshold Slider 0.10–1.00 0.50 Filters low-confidence person detections
Min. Safe Distance Slider 0.50–5.00 m 1.50 m Triggers distance violations below this threshold
Show Distance Lines Checkbox On/Off On Draws violation lines between people
Show Bird's-Eye View Checkbox On/Off On Displays top-down map of positions

How It Works

1. Person Detection

YOLOv8n is exported to ONNX and executed with onnxruntime on CPU. Frames are letterboxed to 640, run through the model, and filtered with NMS to keep clean person boxes. No GPU, PyTorch, or Ultralytics runtime required.

2. Face Mask Detection

Each person crop is scanned using OpenCV's Caffe face detector, then classified by a MobileNetV2 ONNX head. The UI surfaces a three-class outcome: Mask On, No Mask, or Unknown (when no face/model is available).

3. Social Distancing

Bottom-centre foot-points of each person box are projected through a perspective homography into a bird's-eye plane. Pairwise distances are computed in real-world metres using calibrated pixels-per-metre scaling, then rendered as violation lines and colored dots.

Architecture

.
├── __main__.py                 — python -m safesight launcher
├── main.py                     — Qt application entry point
├── convert_model.py            — builds mask_detector.onnx from Keras weights
├── requirements.txt            — Python dependencies
├── assets/
│   ├── banner.svg              — README hero banner
│   ├── screenshots/
│   │   └── image.png           — UI screenshot
│   └── icons/
│       └── .gitkeep            — keeps icons directory in git
├── core/
│   ├── __init__.py             — core package init
│   ├── annotator.py            — supervision-based drawing + labels
│   ├── detector.py             — YOLOv8 ONNX + mask pipeline
│   ├── distancing.py           — homography + bird's-eye map
│   └── video_thread.py         — threaded processing pipeline
├── ui/
│   ├── __init__.py             — ui package init
│   ├── main_window.py          — frameless window + signal wiring
│   ├── sidebar.py              — input controls + settings
│   ├── display_panel.py        — main preview, bird's-eye, charts
│   ├── metrics_bar.py          — realtime metrics cards
│   └── styles.qss              — Qt stylesheet
├── models/
│   ├── yolov8n.onnx             — person detector (auto-download)
│   ├── mask_detector.onnx       — mask classifier (MobileNetV2)
│   ├── _mobilenetv2_base.onnx   — base ONNX for conversion script
│   └── face_detector/
│       ├── deploy.prototxt      — Caffe face detector config
│       └── res10_300x300_ssd_iter_140000.caffemodel — Caffe weights
├── output_videos/
│   └── IMG_0872.mp4             — sample output video
└── venv/                        — local virtualenv (not committed)

Tech Stack

Component Technology Why This Choice
GUI PyQt6 Native desktop UI with strong threading support (QThread)
Person Detection YOLOv8n ONNX Fast inference without PyTorch
Mask Detection OpenCV Caffe DNN Lightweight face detection pipeline
Distancing NumPy + OpenCV homography Accurate real-world coordinate mapping
Video Processing OpenCV VideoCapture Reliable cross-platform capture
Annotation Supervision (Roboflow) Clean, consistent overlays
Graphing pyqtgraph Real-time charts inside Qt
Runtime onnxruntime Cross-platform CPU inference

Models

Model Format Size Purpose Auto-download
yolov8n.onnx ONNX ~12MB Person detection Yes (on first run)
mask_detector Caffe (face) + ONNX (mask) ~4MB + ~3–5MB Mask classification Bundled

Datasets & Training (Separate Branches)

Model training assets are stored on separate branches to keep this repo clean:

  • datasets branch — raw training data and annotations
  • model-training branch — training scripts, Jupyter notebooks, experiment logs To explore the training pipeline: git checkout datasets / git checkout model-training

Project Background

Originally built during COVID-19 (2020) using YOLOv3 + basic Tkinter UI. Rebuilt from scratch in 2026 with:

  • YOLOv3 → YOLOv8n (3x faster, better accuracy)
  • Tkinter → PyQt6 (professional desktop app)
  • Pixel distance → Perspective homography (real metre estimates)
  • Basic CNN → MobileNetV2 DNN (lighter, more accurate)
  • No threading → QThread pipeline (fully responsive UI)
  • Partially implemented → Fully working including live webcam

Contributing

Issues and pull requests are welcome. Please keep changes focused, document new behavior, and include screenshots for UI changes.

License

MIT

Built With

Share this project:

Updates