EdgeNarrator

Real-time AI scene description for live video streams. Frames are sent to a local Ollama instance running the Moondream vision model, and the resulting description is overlaid on a browser-accessible video dashboard with optional text-to-speech.

Demo(On Nvidia Jetson Orin Nano(8GB))

Features

• Works with local webcams and RTSP streams
• Runs inference locally via Ollama — no cloud dependency
• Live web dashboard: video feed with inference FPS overlay
• Scrolling sidebar showing each result alongside a thumbnail of the analyzed frame
• Browser text-to-speech — speaks the latest result, never interrupts mid-sentence

---

Requirements

• Python 3.13+
• Ollama running locally (ollama serve)
• Moondream model pulled: ollama pull moondream
• OpenCV-compatible camera or RTSP stream

---

Installation

git clone https://github.com/torabshaikh/EdgeNarrator.git
cd EdgeNarrator
python -m venv .venv
source .venv/bin/activate
pip install opencv-python ollama flask

---

Usage

python main.py <source> [--prompt "..."]

Source	Example
Local webcam	python main.py 0
RTSP stream	python main.py rtsp://192.168.1.10:554/stream
Custom prompt	python main.py 0 --prompt "Is there any person in the scene?"

Then open http://localhost:5000 in a browser.

---

Project Structure

EdgeNarrator/
├── main.py           # Entry point: arg parsing, wires modules together
├── state.py          # Shared state: description, analyzed frame, inference FPS
├── video_stream.py   # Threaded VideoStream class (RTSP + webcam)
├── analyzer.py       # Ollama inference loop
└── server.py         # Flask web dashboard (MJPEG feed, sidebar, TTS)

---

Configuration

The prompt can be changed at runtime via the CLI (see Usage above).

To change the model or frame size, edit the constants at the top of analyzer.py:

Constant	Default	Description
MODEL	"moondream"	Ollama model name
FRAME_SIZE	(336, 336)	Resolution sent to the model
PROMPT	"Please describe the scene within 20 words"	Default prompt (overridden by --prompt)

---

License

MIT