Camera and YOLO Experiments¶

Camera detection is experimental and is not part of the production kiosk. The standard setup omits camera dependencies, and scripts/start-kiosk.sh starts the server with --no-camera.

Use this guide only to evaluate a CSI camera such as the optional InnoMaker OV9281 global-shutter module. OpenFlight shot measurements still come from the radars.

Prerequisites¶

The experiment script requires these modules in the Pi's Python environment:

picamera2 for CSI camera capture;
ultralytics for YOLO inference; and
opencv-python for image processing and display.

They are intentionally absent from OpenFlight's normal install. Install them in a separate experimental environment appropriate for your Raspberry Pi OS image; do not add them to the production kiosk unless camera support is being restored.

Check the camera¶

Confirm Raspberry Pi OS sees the module before debugging OpenFlight:

rpicam-hello --list-cameras

Then run a short headless capture with an existing YOLO model:

uv run python scripts/vision/test_yolo_detection.py \
  --model models/golf_ball_yolo11n_new_256.onnx \
  --headless --num-frames 10

For a desktop preview on the Pi:

DISPLAY=:0 uv run python scripts/vision/test_yolo_detection.py \
  --model models/golf_ball_yolo11n_new_256.onnx \
  --imgsz 256 --threaded

Useful options¶

Option	Purpose	Default
`--imgsz`	YOLO inference size; smaller is faster	256
`--width` / `--height`	Camera capture resolution	640 × 480
`--fps`	Requested camera frame rate	60
`--confidence`	Minimum detection confidence	0.3
`--threaded`	Separate capture and inference threads	off
`--no-display`	Skip overlay display while benchmarking	off
`--buffer-count`	Camera buffers; fewer reduces latency	2
`--image PATH`	Test one saved image instead of the camera	unset

Start at --imgsz 256. Reduce it if inference is too slow, or increase it when the ball is too small to detect reliably. Measure performance on the actual Pi; frame rate depends on the model, runtime, resolution, and thermal state.

Model export¶

Export a PyTorch model to ONNX:

uv run python scripts/vision/test_yolo_detection.py \
  --model models/golf_ball_yolo11n.pt \
  --imgsz 256 --export-onnx

OpenVINO export is also supported with --export-openvino; add --int8 only after checking the accuracy loss on representative ball images.

Production status¶

The server still contains an optional camera tracker, but the kiosk disables it and the camera dependency extra is empty. Restoring camera-assisted measurement requires dependency packaging, startup integration, hardware validation, and tests; this benchmark script alone does not enable it.