Skip to main content
STUB DOCUMENT: This page is intentionally minimal and will be expanded with deeper technical details in a future update.
The generic cyberwave-edge-camera-driver subscribes to detection results published by any ML worker on the same Zenoh router and renders bounding boxes + labels on the WebRTC stream before it reaches the frontend. No client-side changes are required.

​
How it works

  • Frames captured from the camera are published clean on cw/<twin_uuid>/data/frames/<sensor> — ML workers always receive unannotated images.
  • The driver also subscribes to cw/<twin_uuid>/data/detections/**. Any worker that publishes detections on that channel (e.g. detections/ultralytics, detections/onnxruntime) drives the overlay.
  • Bounding boxes are drawn in-place on the capture ndarray before it is handed to the WebRTC encoder, so the browser sees annotated frames without any extra infrastructure.
  • Workers publish every inference — including empty {"detections": []} heartbeats — so the driver’s freshness timer is kept alive between non-empty frames. Batches older than 2 seconds are discarded to avoid stale overlays.

​
Detection payload schema

{
  "detections": [
    {"label": "person", "confidence": 0.92, "x1": 120, "y1": 80, "x2": 340, "y2": 620}
  ],
  "frame_width": 1920,
  "frame_height": 1080
}
Coordinates are in pixel space of the detection frame. The driver rescales them to the capture resolution.

​
Configuration

VariableDefaultDescription
CYBERWAVE_DETECTION_OVERLAYStrueSet to false to disable overlays on RGB cameras.
Overlays are disabled automatically for twins that declare a depth sensor in their capabilities.

​
Interop with native drivers

The same channel, payload schema, and staleness policy are used by the C++ OBSBOT driver, so an ML worker only needs to publish once — both driver families render the results identically.