This project turns a Seeed reComputer RK3576 Dev Kit into a local AI NVR. fnOS provides storage and application management, while Frigate reads RTSP camera streams and uses the RK3576 NPU through the RKNN detector.
The upstream project documents a verified system with fnOS 1.2.0302, Linux
6.18.18.c1202-trim, RKNPU driver 0.9.8, Docker 28.5.2, and Frigate
0.18.0-rk. Its acceptance baseline uses four low-resolution substreams at
5 FPS for detection. Treat these as the tested versions, not as a compatibility
claim for other fnOS images, kernels, or boards.
Requirements
- Seeed reComputer RK3576 Dev Kit.
- The developer image for reComputer RK3576 Dev Kit: download the compressed image.
- A microSD card, SSD, or eMMC setup supported by that image, plus recovery media and a backup of every target disk.
- One or more RTSP cameras. Keep usernames and passwords out of Markdown, screenshots, and source control.
- SSH access and administrator privileges on fnOS.
Install and check fnOS
Download the reComputer RK3576 developer image, verify its checksum against the release provider's published value, and write the decompressed image using a disk-imaging tool that supports compressed disk images. Imaging overwrites the selected target, so confirm the device before starting.
Boot the board, complete fnOS initialization in a browser, and enable SSH. Record the actual architecture, kernel, addresses, and storage layout before changing any system file:
uname -a
uname -m
ip -br addr
lsblk -o NAME,SIZE,TYPE,MOUNTPOINTSThe verified board uses
/boot/dtb/rockchip/rk3576-recomputer-rk3576-devkit.dtb. Do not replace a DTB
unless it matches the current board and kernel. Back up the existing file first,
and keep bootable recovery media available.
Check that the kernel recognizes the NPU:
uname -r
sudo dmesg | grep -i rknpu | tail -n 15
sudo cat /sys/kernel/debug/rknpu/version
sudo cat /sys/kernel/debug/rknpu/load
readlink -f /sys/class/drm/renderD128/device/driverAn initialized driver only confirms that Linux sees the NPU. Frigate must still load an RKNN model and complete inference successfully.
Install Frigate from App Center
Open App Center on the fnOS desktop and install the Rockchip/RK build of Frigate. The upstream project verified this image:
docker.cnb.cool/frigate-cn/frigate:0.18.0-rkIf the installer appears to stop at 80%, inspect the container and the fnOS gateway before uninstalling it:
sudo docker ps -a --filter name=frigate
sudo docker logs --tail 80 frigate
sudo curl --unix-socket /vol1/@appcenter/frigate/app.sock \
-sS -o /dev/null -w 'HTTP %{http_code}\n' http://localhost/A working installation reports the container as Up (healthy) and the gateway
as HTTP 200. fnOS may expose Frigate only through its installed-applications
page instead of publishing port 8971 directly.
Map the RKNN device
On the tested system, the NPU is /dev/dri/renderD128 on the host, while the
Frigate RKNN runtime checks renderD129 in the container. Back up and edit the
active file at /vol1/@appcenter/frigate/docker/docker-compose.yaml, then add
an alias for the same device:
devices:
- /dev/dri/:/dev/dri/${wizard_coral}
- /dev/dri/renderD128:/dev/dri/renderD129Do not edit the similarly named
/vol1/@appcenter/frigate/docker-compose.yaml. Validate the Compose file and
recreate only the Frigate service as described in the upstream
device-mapping guide.
Configure cameras and detection
Back up the active Frigate configuration before editing it:
sudo cp -a /vol1/@appconf/frigate/config.yaml \
/vol1/@appconf/frigate/config.yaml.bak.manualStart from the upstream
credential-free example,
but merge only the fields you need into the fnOS-generated configuration.
Preserve proxy authentication, TLS, version, and storage settings created by
the package. Supply {FRIGATE_RTSP_USER} and {FRIGATE_RTSP_PASSWORD} through
container environment variables instead of committing camera credentials.
Use one to four low-resolution substreams and 5 FPS detection for initial
acceptance. This detection rate does not limit the live-view frame rate. Begin
with software decoding; test preset-rkmpp one stream at a time after RKNN,
networking, and recording are stable. Do not use preset-vaapi on RK3576.
Restart Frigate and check RKNN initialization:
sudo docker restart frigate
sudo docker ps --filter name=frigate
sudo docker logs --since 3m frigate 2>&1 | \
grep -Ei 'Successfully loaded RKNN|Failed to initialize|error'Validate the deployment
Clone the source repository and run its read-only verification script:
git clone https://github.com/xinjiechen727/fnos_with_rk3576.git
cd fnos_with_rk3576
chmod +x scripts/verify.sh
sudo ./scripts/verify.shAlso confirm that:
- Frigate is healthy and its logs show a successfully loaded RKNN model.
renderD128andrenderD129in the container refer to the same host device.- Live camera video and detection boxes appear in the Frigate interface.
/sys/kernel/debug/rknpu/loadshows activity during detection.
For cameras on another private network, follow the upstream Tailscale subnet-routing guide. Do not expose RTSP port 554, Frigate ports 8971/5000, or the fnOS management interface directly to the public internet.
Recovery and further details
Keep separate backups of the DTB, Compose file, and Frigate configuration. To
roll back, restore only the affected backup and restart or recreate the Frigate
service. Do not run docker compose down -v during a normal rollback because
it can remove data volumes.
See the upstream deployment guide for image-format guidance, independent RKNN inference validation, and detailed rollback steps. The repository also includes troubleshooting notes and the tested four-camera configuration.