A Raspberry Pi with a camera. Real-time ROS 2 object detection on the device, releases built from git tags, over-the-air updates and automatic rollback when a release breaks.
git tag vX.Y.Z
│
│ publish a GitHub Release
▼
GitHub Release
│
│ CI builds and pushes the image
▼
container image (on GitHub's registry)
│
│ polled every 2 minutes
▼
agent/updater.py
│
│ pulls the new image, restarts the service
▼
vision-stand.service (the perception/ ROS graph)
│
│ health check
▼
├─ pass → stays on it
└─ fail → rolls back to the previous version
perception/: the ROS 2 workspace. Camera capture, motion detection, on-device object detection and the version overlay.agent/:updater.py, the device-side process that pulls new releases and rolls back on a failed health check.scripts/:env.shsets up the ROS environment.export_model.pyis a one-time dev-machine script that exports YOLOv8n to ONNX at 320 and 640. It needs PyTorch, so it never runs on the Pi.deploy/: provisioning script, systemd units and the per-device env template.
| Node | What it does |
|---|---|
camera_node |
Reads a video file or camera and publishes frames. |
detector_node |
Frame-differencing motion detector. |
object_detector_node |
Runs YOLOv8n on each frame, publishes detections plus an annotated video stream and overlays the running version. It also exposes a /detector/take_snapshot service that freezes the current frame and its detections to disk. |
git clone https://github.com/maxriazantsev/vision-stand && cd vision-stand
source scripts/env.sh # ROS environment + this project's ROS_DOMAIN_ID
cd perception
colcon build
source install/setup.bash
ros2 launch detector pipeline.launch.pyBy default this runs against the bundled test_clip.mp4. To use a real
camera instead:
ros2 launch detector pipeline.launch.py source:=0A git tag is the version. There's nothing to hand-edit:
- Cut a GitHub Release with tag
vX.Y.Z. - CI builds the image and publishes
ghcr.io/maxriazantsev/vision-stand:vX.Y.Z. - Each device polls the GitHub Releases API every two minutes and pulls whatever's new.
Two ways, depending on who notices the problem first:
- Manual: mark the bad release as a pre-release on GitHub. The Releases API's "latest" skips pre-releases, so every device still following latest falls back to the one before it on its next poll.
- Automatic: after every swap, the device runs a health check against the freshly started container. If it fails, the device reverts to whatever it was running before, on its own, without waiting on a human. The release that failed is remembered and skipped until a newer one is published.
The health check only proves camera frames arrive after a swap. It catches a crash, a missing camera or ROS not starting, but not wrong detections.
A device can also be pinned to one version instead of following latest, by
setting PINNED_VERSION in its /etc/vision-stand/device.env.
deploy/provision_device.sh <ssh-host> <vision-stand.service|vision-stand-sim.service>Needs only Docker on the target; installs it if missing, detects the
camera's stable /dev/v4l/by-id/... path and lays down the systemd units
and the updater. vision-stand-sim.service runs against the bundled test
clip instead of a camera, for a device-free dry run of the same update path.
Every device also gets vision-stand-foxglove.service, a bridge that
exposes the live ROS topics on port 8765. Connect to it with
Lichtblick, to
see the annotated feed in real time.