Insights

False alarms in DAI: how to classify and calibrate

Published on by the Omnivias team at Omnicog Tecnologia.

A false alarm is an alert from Automatic Incident Detection (DAI, known internationally as AID) that does not correspond to a real incident. Operations classify every alert during the response, and that history shows where and why a rule gets it wrong. With per-camera calibration, unfounded alerts in Omnivias projects in production stay below 10% after site calibration.

True, passive or false

The CFTV-IVA specification from ARTESP, the São Paulo state transport regulator, requires operations to classify each event as true (positive), passive or false. True is a real incident that calls for action; passive is a real event that calls for no intervention; false is an alert with no incident. Classification is what turns alerts into data for improving the system.

Where false alarms come from

  • Light and shadow: long late-afternoon shadows, headlights and reflections on wet pavement.
  • Weather: heavy rain, fog and drops on the lens.
  • Camera out of position: a PTZ that did not return to its preset, or a swaying pole.
  • Expected situations: marked work zones and closures, the concessionaire’s own crews on the road.
  • Generic rule: the same criterion applied to different sites.

How calibration reduces errors

Configuration is done per camera and per PTZ position: lanes, shoulder, right-of-way and direction are drawn in the zone editor. Zones marked with cones are respected by the alert, so the work itself does not trigger it. Camera health warns when the view is blocked or out of position.

After that, the false alarm report by camera and by event type shows where to adjust. Footage from the highway itself strengthens the models throughout operation. The result is measured on the alerts that operations classified: in projects in production, unfounded alerts stay below 10% after site calibration.

The opposite error

Cutting alarms at any cost lets the event that mattered slip through. That is why calibration looks for the balance agreed with operations, not for the lowest number of alerts.

See also: DAI, the CFTV-IVA specification from ARTESP and the camera checklist.

Frequently asked questions

What is the DAI false alarm rate?

In projects in production, unfounded alerts stay below 10% after site calibration, measured on the alerts classified by operations.

What is a passive event?

It is a real event that requires no intervention from operations. In the classification of the ARTESP CFTV-IVA specification, events are true (positive), passive or false.

Do work zones on the road cause false alarms?

Not when the area is marked with cones: DAI recognizes the marked zone and the alert respects it.

How do I know which cameras get it wrong the most?

From the false alarm report by camera and by event type, generated from the classification made by operations.

Start with a pilot on your own cameras

First we assess the cameras the highway already has; then a proof of concept using footage from the road section itself shows the measured result. Depending on the size of the project, the assessment and the proof of concept can be free of charge, and the proof of concept starts running within 10 days of signing the agreement.

  1. Assessment

    Resolution, position, coverage, lighting and network of the existing cameras, against what the operation wants to measure and detect.

  2. Proof of concept

    SAT and DAI running on footage from the road section itself, measured against the agreed success criteria.

  3. Production

    On-premises, cloud or hybrid architecture, control center integration, rules, alerts and reports.

  4. Continuous calibration

    Rules and models tuned camera by camera using the real alert history.

How the pilot worksTest it with my highway’s cameras
WhatsApp