Skip to main content
End-to-end recipe for a zone-based intrusion detection pipeline that runs entirely on the edge: raw video stays on-device, anonymised frames feed the WebRTC stream, and the cloud only ever sees alert events.
Stack: Workflows, YOLO, Spatial Filter, Timed Condition, Edge Workers

​
Pipeline

The anonymize node publishes the pixelated frame to the camera driver’s filtered frame channel so the WebRTC stream the operator sees is privacy-preserving by construction: no raw video ever leaves the edge.

​
Prerequisites

  • A twin bound to a camera with a Cyberwave camera driver and CYBERWAVE_METADATA_FRAME_FILTER_ENABLED=true.
  • Edge Core running with at least one healthy edge runtime.
  • The yolov8n model in the workspace catalog (auto-downloaded on first reference).

​
Authoring the workflow

  1. Create a workflow scoped to the camera’s environment, with Run on edge enabled.
  2. Add the trigger: drag a Camera Frame node from the Triggers group, set twin_uuid to the camera twin.
  3. Add detection: drag a Call Model node, pick yolov8n, set target_classes to ["person"]. Wire camera_frame → call_model.
  4. Add privacy: drag an Anonymize node, set mode to pixelate. In the inspector’s Target Classes chip editor, leave it empty to default to person, or add chips (e.g. face, vehicle) if your model emits other class labels you want obscured. Wire call_model → anonymize. The downstream WebRTC stream now shows pixelated people automatically.
  5. Add the zone: drag a Spatial Filter node from the Decision & Control group. In the inspector, click Capture frame to fetch a still from the camera and trace the zone polygon over it. Set point_mode to bbox_bottom_center for floor-anchored subjects. Wire anonymize → spatial_filter.
  6. Add the dwell gate: drag a Timed Condition node. Mode stays on the default sustained. Set min_duration_s to the loitering threshold you care about (5s for a tight intrusion alert, 30s for a “left unattended” scenario). Set cooldown_s to your alert review cadence (30–120s is typical). Wire spatial_filter → timed_condition.
  7. Add the alert: drag a Send Alert node downstream of timed_condition. Set name (e.g. Intrusion), alert_type (e.g. intrusion), severity (warning), and leave category on business. The edge worker compiles this into a cw.publish_alert(...) call wired to the dwell-fired frame.
  8. Activate the workflow. The edge runtime will pull the generated worker (wf_<uuid8>.py) on its next sync, and the polygon will appear as a read-only overlay on the twin’s WebRTC stream so operators can see which zones are armed.

​
Optional: email on alert

The edge worker writes alerts via cw.publish_alert(...) and they land in the Alerts panel. To also receive an email, add a second workflow with an alert trigger: 
alert_trigger(twin = camera-1, alert_type = intrusion)
  → send_email(to_email = …, subject = …, body = …)
This cloud-side workflow wakes on every edge-emitted intrusion alert (no per-frame cost) and uses Django’s email backend to deliver the message. See timed_condition for the full two-workflow shape.

​
Operator experience

  • Frontend overlay: the twin’s video panel shows the zone polygon outlined in blue. Multiple zones (e.g. authorised vs forbidden) appear as separate polygons.
  • Alerts: when the dwell threshold is crossed, a send_alert-created alert appears in the Alerts panel with the polygon and dwell duration embedded in metadata.

​
Privacy boundary

The Zenoh→MQTT bridge does not forward raw frames. Only the alert payload (with optional anonymised snapshot) crosses the WAN. See the Security Pipeline page for the deeper privacy contract.
pixelate is reversible by public depixelation models. Combine with redact or rely on the event-only contract (no frames in the alert payload) for legal de-identification.

​
See also