NeuralEye

A real-time computer vision application with two switchable detection modes: Human Pose Estimation and Animal Detection — powered by YOLO26 and streamed live to a browser via WebSocket.

Modes

Mode	Description	Output
Human Pose	Tracks 17 human body keypoints in real time. Triggers a "Hands Raised" alert when both wrists are above the shoulders.	Skeleton overlay + Hands Raised: YES / NO
Animal Detection	Detects 10 COCO animal classes: bird, cat, dog, horse, sheep, cow, elephant, bear, zebra, giraffe. Draws bounding boxes with class labels and confidence scores.	Annotated bounding boxes + detection list

Switch between modes instantly using the toggle in the browser UI — only one mode runs at a time.

Tech Stack

Layer	Technology
Language	Python 3.11+
Pose Model	Ultralytics YOLO26 — yolo26n-pose.pt
Detection Model	Ultralytics YOLO26 — yolo26n.pt (COCO)
Image Processing	OpenCV
Web Server	Flask + Flask-SocketIO
Containerisation	Docker

Repository Structure

neuraleye/
├── src/
│   ├── detector.py          # PoseDetector — YOLO26 pose model + hands-raised logic
│   ├── animal_detector.py   # AnimalDetector — YOLO26 COCO model, animal classes only
│   └── vision.py            # VideoStream — OpenCV capture + FPS helper
├── templates/
│   └── index.html           # Web UI — mode toggle, webcam feed, live status
├── app.py                   # Flask + SocketIO server (dual-mode frame handler)
├── main.py                  # CLI runner (human pose, OpenCV window)
├── requirements.txt         # Python dependencies
├── Dockerfile               # Python 3.11-slim, pre-bakes both YOLO models
└── docker-compose.yml       # Single-service compose on port 5001

Quick Start

Local

pip install -r requirements.txt
python app.py

Open http://localhost:5001 in your browser, allow camera access, and use the toggle to switch modes.

CLI (Human Pose only — OpenCV window)

python main.py

Docker

docker compose up --build

Open http://localhost:5001.

The Docker image pre-downloads both YOLO model weights (~6 MB each) at build time, so the first inference is instant. Expect a ~2–3 GB image due to PyTorch.

How It Works

1. The browser captures webcam frames and sends them to the server over WebSocket along with the current mode.
2. The server decodes the frame, runs the appropriate YOLO model, and returns an annotated JPEG.
3. The browser displays the annotated feed and updates the status line in real time.

Model weights (*.pt) are not tracked in git — Ultralytics downloads them automatically on first use.

License

Built with Ultralytics YOLO.