Skip to content

Live Detection

The Live Detection page provides a real-time view of the AI detection engine in action, overlaying detection bounding boxes and labels directly on camera streams. This page is separate from the standard Live View and is designed for monitoring detection performance and testing sensitivity settings.

Live Detection serves two primary purposes:

  1. Monitoring — Observe detection activity across cameras in real time. See what the AI engine is detecting, at what confidence levels, and in which zones.
  2. Testing — Adjust detection sensitivity settings and immediately observe the effect on live camera streams. This is invaluable for tuning confidence thresholds and class filters before committing changes to the recorded detection configuration.

The Live Detection page has its own independent confidence threshold and class selection settings, separate from the recorded detection configuration. This separation is intentional:

SettingRecorded DetectionLive Detection
Confidence thresholdConfigured per server in Settings > SystemConfigured on the Live Detection page
Class selectionConfigured per server in Settings > SystemConfigured on the Live Detection page
ScopeGenerates events stored in the databaseDisplay only — no events generated

This design allows operators and administrators to:

  • Test lower confidence thresholds to see what the model detects at lower certainty without flooding the event log.
  • Preview additional classes to evaluate whether enabling a new class would be useful before adding it to the server configuration.
  • Verify zone effectiveness by watching live detections and confirming they align with configured zones.

The Live Detection page displays cameras from servers that have detection enabled. Select one or more cameras from the location tree to view their detection overlays.

Each camera stream displays:

  • Bounding boxes around detected objects, colour-coded by class
  • Class labels with confidence percentage above each box
  • Targeting reticle — corner brackets and centre crosshair on each detection for precise localisation

The overlay updates in real time as the detection engine processes frames.

Use the controls on the Live Detection page to adjust:

  • Confidence threshold — Slide to lower values to see more detections (including lower-confidence ones) or higher values to see only high-certainty detections.
  • Class filter — Toggle individual classes on and off to focus on specific detection types.

Changes take effect immediately on the displayed overlays.

When a new camera is added to the system:

  1. Open the camera on the Live Detection page.
  2. Set the confidence threshold to a low value (e.g. 20%) to see everything the model detects.
  3. Observe the detections over several minutes to understand the false positive rate.
  4. Gradually increase the threshold until false positives are minimised while real detections are retained.
  5. Apply the optimal threshold to the server-level recorded detection configuration.

After creating or modifying detection zones:

  1. Open the camera on the Live Detection page.
  2. Verify that detections within the zone boundaries are being captured.
  3. Confirm that detections outside zone boundaries are being filtered out.
  4. Adjust zone boundaries if needed and re-verify.

After a model change or training run:

  1. Open several cameras on the Live Detection page.
  2. Compare detection accuracy and false positive rates against the previous model.
  3. Confirm that the new model performs as expected before relying on it for event generation and alarms.

The Live Detection page is subject to role-based access control. Users must have the appropriate page permission to access Live Detection. Camera visibility is restricted based on the user’s assigned locations.