Website: skynode-si.netlify.app (source in web/)

A low-cost visual sensing platform, starting with one passive sky-tracking node.

A webcam on a pan-tilt mount watches the sky, a YOLO model finds aircraft and birds in each frame, the mount turns to follow the target, and every sighting is logged. This is Phase 1: one node, mostly tested in simulation so far. The physical unit is not built yet.

Scope: passive sensing and tracking only. No payloads, no effectors, and nothing that interacts with or interferes with aircraft. No jamming, no spoofing, and no transmitting on aviation or drone-control frequencies. Ordinary Wi-Fi and USB networking between Skynode's own parts is fine.

flowchart LR
    CAM[Webcam] -->|frames| DET[Detection<br/>YOLOv8n · ONNX]
    DET -->|boxes| TRK[Tracking<br/>pick one target]
    TRK -->|pixel error| CTL[Pan-tilt control<br/>proportional]
    CTL -->|P92.5 T47.0<br/>Wi-Fi or USB| PICO[Pico WH<br/>servo firmware]
    PICO -->|PWM| SERVOS[Pan + tilt<br/>servos]
    SERVOS -.->|camera moves| CAM
    DET --> LOG[(Sighting logger)]
    TRK --> LOG
    LOG --> DASH[Live dashboard]

1. Detection. The brain (a laptop for now, a Raspberry Pi 4 later) grabs webcam frames and runs a YOLO model through ONNX Runtime. It runs a custom two-class model (drone and aircraft). See Model.
2. Tracking. It picks one target and measures how far that target sits from the center of the frame, in pixels.
3. Pan-tilt control. A proportional controller turns the pixel error into a small correction ("the target is 40 px right, so pan +2°").
4. Actuation. The brain sends a plain-text command like P92.5 T47.0 to the Pico over Wi-Fi (UDP) or USB serial. The Pico moves both servos smoothly and keeps them inside safe angle limits.
5. Logging. Each sighting is recorded with time, class, confidence, pan and tilt angle, and a snapshot. A live dashboard shows the camera, the boxes, and the log as it grows. See Live dashboard.

Detection, tracking, and control are separate modules, so each layer can be reused on future platforms.

Software only. You can run detection on a webcam or video file without the Pico.

pip install -r brain/requirements.txt opencv-python-headless onnx pyyaml
python -m unittest discover tests -v      # hardware-dependent tests skip with a stated reason

Then follow brain/README.md to export the YOLOv8n model once and run the brain.

No hardware? python -m brain.demo <folder of test videos> --model <model.onnx> runs detection, the logger and the dashboard on your own videos. See Demo mode. For the Pico, see pico/README.md.

All numbers here come from the repo. Nothing has been measured on the physical node.

Detection model v2, tested on 43 real-world videos of planes, military jets, and drones:

• On real aircraft, v2 wrongly called 7.5% of frames a drone (798 of 10,657 frames).
• That was 0.8% for civilian planes and 10.8% for military jets (mostly distant F-35s).

Model v3 (merged classes drone and aircraft, 960 px input, resumed from v2) is in training. Next step: score it against the same 43 clips.

Training used YOLOv8n on a 28,526-image aircraft and drone dataset. See Training.

• Range (estimate, not yet measured). A wide-lens webcam detects small drones only at short range, likely tens of meters. Aircraft are detectable much farther.
• Conditions. Visual sensing is weaker in darkness, fog, and rain.
• Not a replacement for Remote ID or RF sensors. It can see drones that broadcast nothing, which Remote ID can't, but it doesn't replace either.
• Hardware not built. Nothing has run on the physical servos, camera mount, or Pico yet. Anything marked "simulated" has only been tested in software.

skynode/
├── pico/       MicroPython servo firmware (runs on the Pico WH)
├── brain/      detection, tracking, control, logging (runs on laptop / Pi 4)
├── hardware/   3D-print files for the pan-tilt mount
├── training/   Colab training notebook and dataset scripts
├── web/        project website
├── docs/       wiring diagram, banner, renders
└── tests/      laptop-side tests

Model files are not in git (*.onnx and *.pt are gitignored).

• Current model: a custom two-class YOLOv8n, with classes drone and aircraft. It was built by merging the labels of a source dataset of 28,526 labeled images into those two classes. Its first version is v2 (results above).
• v3 (drone and aircraft, 960 px input, resumed from v2) is in training. It has not been scored yet.
• The model path and target classes are settings in brain/config.toml, so a new model drops in without code changes.
• See Training for how models are made. Finished models go on GitHub Releases.

Also in bom.csv. Costs are rough USD estimates for the whole line (both servos in the servo row), before shipping.

About $113 in new parts for the sensing hardware (camera, servos, mount, power), before the computer that runs the model. The total leaves out the Pico WH and breadboard (already owned) and the computer.

SG90 wire colors: brown = GND, red = +5 V, orange = signal.

Power notes

• The servos get their own 5 V 2 A supply, so a stalling servo can't brown out the Pico. Don't also connect Pico VBUS (pin 40) to the + rail, or two supplies will fight.
• The common ground wire is required: servo signals are measured against GND.

hardware/pantilt.py is a parametric CadQuery design: servo size, horn, webcam size, and wall thickness are parameters at the top of the file. Each part has an editable STEP file and a print-ready STL. It prints as three parts without supports:

Measure your servo and camera and adjust the parameters before printing. See hardware/README.md.

Vision. Phase 1 is one passive sky-tracking node. The long-term goal is autonomy for missile and drone detection and defense systems. Open work is tracked as GitHub issues under the Phase 1: one working node milestone.

• Pico servo firmware: smooth motion, calibration, command protocol (tested in software)
• Detection and tracking loop, working in simulation, covered by automated tests
• Pan-tilt mount designed in CAD, with STEP and STL files
• Wiring diagram and bill of materials
• Detection model training: v3 with drone and aircraft classes (in training, not yet scored)
• Sighting logger (built and tested in simulation; not yet run on real hardware)
• Live dashboard (built and tested in simulation; not yet run on real hardware)
• Build the physical hardware
• Outdoor test
• Test against real flight data (ADS-B) and publish accuracy results
• Pi 4 + solar deployment

Nothing has run on real hardware yet. Every item marked as simulated or software-tested has only been checked on a computer.

The website's Status section lists these same items. Update both together.

Claude Code (Anthropic's coding agent) helped write much of the code, tests, and documentation here. Every change is in the git history.

MIT, see LICENSE.