A self-hosted NVR (Network Video Recorder) for 24/7 recording of RTSP cameras with live view, timeline playback, trickplay thumbnails, motion detection, and AI object detection. Free and open source.
- 24/7 continuous recording -- HEVC passthrough (no transcode, no GPU usage) into 15-minute segments
- Live view -- low-latency direct RTSP via go2rtc with multi-quality streaming (Direct/High/Low/Ultra Low), zoomable video
- Timeline playback -- zoomable 24h timeline, instant seek, forward/reverse speed controls (-4x to 32x), date picker
- Trickplay thumbnails -- hover/scrub preview on timeline
- Motion detection -- per-camera sensitivity, timeline overlay, configurable script triggers
- AI object detection -- RT-DETR for persons, vehicles, and animals with color-coded timeline bars and multi-frame confirmation
- Facial recognition -- enroll known people from past detection thumbnails (SCRFD + ArcFace on DirectML); timeline highlights cyan for known faces, rose for strangers; per-person script triggers
- Clip export -- select a time range and export an MP4
- Retention management -- configurable max age and max storage, oldest recordings purged first
- Native apps -- Windows desktop and Android client
- Runs as a Windows service -- auto-starts on boot
Download the latest RichIris-Setup.exe from Releases.
The installer will:
- Install the RichIris backend and desktop app
- Ask you to choose a data directory for recordings, database, and logs (pick a drive with plenty of space)
- Download required dependencies (FFmpeg, go2rtc, RT-DETR + SCRFD + ArcFace ONNX models) automatically
- Install and start the RichIris Windows service
Open the RichIris app (desktop shortcut or Start Menu). On first launch, enter the server address -- http://localhost:8700 if running on the same machine.
Go to Settings and add your cameras with their RTSP URLs. Cameras start recording immediately once added.
Download RichIris-Android.apk from Releases and install it on your Android device. Enter your server's IP address (e.g., http://192.168.1.100:8700) to connect.
RichIris records via HEVC passthrough (no transcode), so per-camera CPU load is low. The heavier workloads are AI object detection (ONNX inference via DirectML on any DX12 GPU) and on-the-fly transcoding for non-Direct quality tiers (NVENC on NVIDIA GPUs). Sizing below assumes ~4K HEVC cameras at 4-8 Mbps on the main stream.
| Component | Minimum |
|---|---|
| OS | Windows 10/11 64-bit |
| CPU | 4 cores / 8 threads — Intel 8th gen Core i5 / AMD Ryzen 5 2600 or newer |
| RAM | 8 GB |
| GPU | Any DirectX 12 GPU with 2 GB VRAM (integrated Intel UHD 630 / AMD Vega / NVIDIA GTX 1050 all work for RT-DETR via DirectML). NVIDIA required only if you want NVENC transcoded quality tiers. |
| System drive | 256 GB SSD (OS, backend, database) |
| Recording drive | Dedicated HDD or SSD — plan ~60 GB/day per 4K HEVC camera @ 6 Mbps (≈1.7 TB per camera for 30-day retention) |
| Network | Gigabit Ethernet, cameras on the same LAN |
| Cameras | RTSP H.264 or HEVC with sub-stream support (motion/AI read the sub-stream at 2 fps) |
| Component | Recommended |
|---|---|
| CPU | 6+ cores — Intel 12th gen Core i5 / AMD Ryzen 5 5600 or newer |
| RAM | 16 GB |
| GPU | NVIDIA RTX 3050 or better (6 GB+ VRAM) — fast RT-DETR inference and NVENC for Low/Ultra Low live and playback tiers across all cameras |
| System drive | 500 GB NVMe SSD |
| Recording drive | 8+ TB HDD (CMR preferred; SMR is fine for single-writer NVR workloads). Per-camera budget: ~60 GB/day @ 4K HEVC 6 Mbps; scale linearly with resolution, bitrate, and retention. |
| Network | Gigabit LAN, wired cameras (PoE switch recommended) |
- Recording CPU cost is near-zero per camera (copy-mode FFmpeg). 10+ cameras on a modern quad-core is realistic.
- AI detection runs only when motion is detected, and inference is ~11 ms per frame on a modern GPU. The bottleneck for many cameras is the persistent sub-stream decoders (one FFmpeg per camera at 2 fps), not the neural net.
- Transcoded live/playback (High when source is non-HEVC, Low, Ultra Low) use NVENC on NVIDIA GPUs. Without an NVIDIA GPU, stick to Direct quality — Direct is raw passthrough and costs nothing.
- Storage is the usual dominant cost. Use the retention controls in System Settings (max age + max storage) to bound disk usage.
- Memory is mostly FFmpeg + go2rtc + ONNX runtime. Budget ~300-500 MB per camera as a rough ceiling.
All settings are managed through the app's System Settings screen:
| Section | What it controls |
|---|---|
| General | Timezone |
| Storage | Data directory (recordings, database, logs, thumbnails) |
| Retention | Max recording age (days), max storage (GB) |
| Trickplay | Enable/disable timeline thumbnail previews |
| Logging | Log level |
The data directory and port are also stored in bootstrap.yaml next to the application (editable if needed before first launch).
Flutter App (Windows/Android)
|
v
FastAPI backend (:8700) --> go2rtc (:18700/:18554) <-- RTSP cameras
|
v
SQLite DB + Recordings + Thumbnails
- Recording: One FFmpeg process per camera, codec passthrough (no transcode), 15-minute
.tssegments. Reads go2rtc's relay of the main stream, so each camera serves one main-stream session rather than two (falls back to the camera directly if go2rtc is down) - Live view: go2rtc receives camera RTSP streams and re-serves via RTSP (:8554). Flutter app connects directly to go2rtc for smooth HEVC playback. Zoomable video in fullscreen.
- Playback: Direct mode serves raw segments instantly. Other quality tiers transcode on-the-fly via NVENC
- AI detection: Snapshot-based pipeline -- motion pre-filter, then RT-DETR inference with multi-frame confirmation (2 detections in 3 frames + positional movement)
- Facial recognition: Runs only after RT-DETR confirms a person. Pulls a 4K main-stream snapshot from go2rtc, crops to the person bbox, runs SCRFD to find faces, then ArcFace for 512-D embeddings. Cosine-matches against an in-memory cache of enrolled faces. Known faces tint the timeline cyan, unknowns rose. Enroll from the Faces button on the home screen by tagging past detection thumbnails — the filter shows only thumbnails where a face was actually detected.
| Quality | Live View | Playback | Server Load |
|---|---|---|---|
| Direct | Raw passthrough | Raw .ts file |
Zero |
| High | HEVC re-encode, source quality | HEVC NVENC, source quality | GPU |
| Low | HEVC re-encode, 1/8 bitrate | HEVC NVENC, 1/8 bitrate | GPU |
| Ultra Low | 1/16 bitrate, 15fps | 1/16 bitrate, 15fps | GPU (light) |
See DEV-GUIDE.md for development setup, project structure, and build instructions.
MIT