Local camera detection, face recognition and AI notifications using an Imou Rex 3D, Frigate and Intel integrated graphics.
Experimental community project. Includes the actual OpenVINO app source developed for the original installation, sanitized Frigate settings, and a new portable notification example. It is not affiliated with Imou, Frigate, Intel or Home Assistant.
- Frigate 0.18 configuration: main-stream detection, go2rtc restream, VAAPI decoding, OpenVINO object detection, CPU face recognition, two-day event recording retention.
- HAOS amd64 app: Qwen3-VL-2B INT4 on OpenVINO GPU, with local Ollama CPU fallback.
- Android notification example with full snapshot, a stable event tag, face-name updates and an optional clip action.
- Installation, AI integration, troubleshooting and privacy documentation.
The original deployment used an Intel i3-13100T / UHD 730 and Imou Rex 3D 3MP. The owner reported a successful full GPU description in approximately 48 seconds and successful face recognition. This is one observation, not a general benchmark.
The public 0.1.1 package makes service URLs and camera name configurable. These portability edits have local automated checks; they have not been rebuilt or retested on HAOS hardware. The notification YAML is a newly prepared example, not an export of the owner's latest AI automation.
- Follow installation.
- Adapt Frigate configuration to your camera and MQTT credentials.
- Confirm detection, recordings and face recognition before adding AI.
- Install the OpenVINO app, run
verify, inspect its description, then switch toserve. - Adapt notifications and read AI integration.
The app requires HAOS amd64, Intel GPU access, Ollama and a CIFS/NFS mount at /share/ollama. This storage guard is inherited from the original setup and intentionally retained. It is not a generic install-and-run app for every HA host.
flowchart TD
camera[Imou RTSP] --> frigate[Frigate and go2rtc]
frigate --> events[MQTT events and face labels]
events --> ha[Home Assistant notifications]
ha --> vision[LLM Vision optional]
vision --> gpu[OpenVINO app]
gpu -->|failure| cpu[Ollama CPU]
Face identity comes from Frigate's enrolled face library; the language model describes visible content. Neither is suitable for deciding access permissions.
- Installation
- Face recognition
- Notifications and AI
- Troubleshooting
- Privacy and reporting
- Italian quick start
- Changelog
Run python3 -m unittest discover -s tests -v. GPU inference and image builds need the target hardware and are outside these unit checks.
Project code is MIT licensed. Frigate, OpenVINO, Ollama and model weights retain their respective licenses. Model weights and camera media are not distributed here.