Alarm Filtering

15 min

How it works

Alarm Filtering verifies rule events with a second detector on the core, so fewer false alarms reach operators.

A verified event: a person inside the alarm zone.

Good to know

It verifies three object types: people, vehicles and two-wheeled vehicles.

Events the filter does not confirm are not shown in Monitor.

If verification takes too long, the event is shown anyway, so real incidents still reach operators.

Alarm Filtering needs a GPU on the core. Rules that trigger often, for example in busy scenes, add more GPU load.

It works with color and IR cameras with a normal view.

How it verifies

When a supported rule raises an event, the edge holds it briefly and sends an image of the event object to the core. A detector there verifies that the object is in the image, and that it is the type the rule expects: a person, a vehicle or a two-wheeled vehicle.

flowchart LR
    subgraph EDGE["On the edge"]
        A["Rule raises<br/>an event"] --> B["Event is held<br/>briefly"]
    end
    subgraph CORE["On the core"]
        C["Detector verifies<br/>the object"]
    end
    B -->|"Image of the object"| C
    C -->|"Confirmed"| D["Shown in Monitor"]
    C -->|"Not confirmed"| F["Filtered out"]
    B -.->|"No answer in time"| D

Result	What happens
Object confirmed	The event is shown in Monitor.
Object not confirmed	The event is filtered out.
No answer in time	The event is shown in Monitor.

How to set up

Where to turn it on

In Administrator, select the camera and open Settings.

Under Analytics, turn on Enable alarm filtering and click Save.

Set up at least one supported rule on the camera.

Administrator › Settings for a camera, with Enable alarm filtering under Analytics.

Supported rules and cameras

Supported rules

Rule	Supported
Moving in an Area	✓
Line Crossing	✓
Loitering	✓
Stopped Vehicle	✓
Grouping	✓
Counterflow Traffic	✓

Events from other rules are sent as usual, without alarm filtering.

Supported camera types

Camera type	Supported
Color	✓
IR	✓
Thermal	–

Supported view types

View type	Supported
Normal	✓
Fisheye, overhead	–

What to expect

Use cases

Alarm Filtering helps most where the scene has regular nuisance sources, such as:

Trees and foliage moving in the wind

Flags, plastic and other loose items

Shadows and reflections

Small animals such as dogs and birds

It is tuned to cut false alarms. In rare cases it may also filter out a real event, for example when the object is hard to see in the image.

Bandwidth and load

Filtered events are not uploaded, so Alarm Filtering often reduces upload traffic. A verification image is about 100 KB, while a full event can exceed 1 MB.

Each verification uses a little GPU capacity on the core. Rules that trigger often, for example in busy or crowded scenes, send more verifications and add more load.

Tuning and troubleshooting

What you see	Likely cause	What to do
A false alarm still reaches Monitor	The rule is not supported, the object is very small, or verification took too long.	Check that the rule is in the supported list. For small objects and timeouts, see Technical notes.
A real event is missing in Monitor	The detector could not confirm the object.	Check that the object is clearly visible and large enough in the image. If it happens often, contact support with the event times.
No events are filtered	Alarm filtering is off, or the camera has no supported rule.	Turn on Enable alarm filtering and check the rules on the camera.

Constraints and tradeoffs

Alarm Filtering confirms the object type and location. Rule logic such as dwell time, distance or group size is handled by the rule itself.

A slow or unstable connection between edge and core may cause more events to be shown without verification. See Technical notes.

Very small objects are shown without verification. See Technical notes.

Technical deep dive

How a verification works

A supported rule raises an event on the edge.

The edge holds the event and creates an image around the event object, with some of the surrounding scene for context.

The image is sent to the detector on the core, which looks for people, vehicles and two-wheeled vehicles.

The edge compares the result with the original event. The class must match, and the detection must overlap the event object.

The edge shows the event in Monitor or filters it out. If no answer arrives in time, the event is shown anyway.

Technical notes

For rules with several objects, such as Grouping, each object is verified. The event is confirmed as soon as one object is confirmed.

Objects smaller than the minimum object size, typically 32 × 32 pixels, are not sent for verification and are shown as normal events.

If the core does not answer in time, the event is shown in Monitor without verification. A slow or unstable connection between edge and core makes this more likely.

With alarm filtering on, a few more candidate events reach verification, which can help catch borderline real events.

Verification does not change the rule sensitivity.

Stopped Vehicle runs in the cloud and reaches alarm filtering through its own path. Unattended Object uses its own built-in verification.

Definitions

Verification: validation of a rule event by the detector on the core.

Event object: the person, vehicle or two-wheeled vehicle that triggered the rule.

Confirmed: the detector found the expected object at the event location. The event is shown in Monitor.

Filtered out: the detector did not find the expected object. The event is not shown in Monitor.

Released: no answer arrived in time. The event is shown in Monitor without verification.