Skip to main content
Version: V12

How Live AI Detection Works

A camera that only streams is passive. It shows you the scene, but it's on you to notice anything. Detection is what makes a camera watchful. As the feed comes in, AI Live Insight analyzes it frame by frame and recognizes the objects you care about, a person, a vehicle, a weapon, so the system, not just the operator, is paying attention every second the camera is live.

This article explains what detection does with a live feed, the kinds of objects it can spot, and the two dials, confidence and severity, that decide what's worth flagging.

What Detection Does​

For each camera, you choose which objects to watch for. While the camera is live, the surveillance service checks every frame, and when it recognizes one of your chosen objects with enough certainty, it creates a detection. A single detection is the engine behind three things you actually use:

  • Alerts. When a detection matches the camera's alert settings, it raises an alert so someone knows.
  • Events. Each detection becomes an event on the timeline and in the event feed, so you can scan a camera's history and jump to the moment it happened.
  • Recordings. A clip is captured around the detection, giving you the video evidence to review later.

So detection isn't a feature off to the side. It's the thread that ties the live feed, the alerts, the timeline, and the recordings together.

What Detection Can Spot​

AI Live Insight can recognize a broad set of objects on a live stream, and they group into a few object classes:

  • People: person, head, and face.
  • Vehicles: car, bus, truck, bike, boat, and airplane.
  • Weapons: guns and knives.
  • Safety and PPE: items such as head, eye, ear, hand, and foot protection, and high-visibility gear, so you can flag where protective equipment is or isn't present.
  • Text on objects: license plate, house number, and street sign.
  • Environment and hazards: objects such as fire, fire extinguisher, fire hydrant, and graffiti.

The trick isn't turning everything on. It's matching each camera to what it's there for. A parking entrance watches vehicles and license plates. A lobby watches people and faces. A loading dock might watch for safety gear. Pick the classes that fit the camera's job, and skip the rest so the camera isn't flagging noise.

How Confidence Works​

Every detection comes with a confidence score, a percentage that captures how sure it is. The Confidence Threshold you set is the bar a detection has to clear before it counts: set it to 35, and only detections the AI is at least 35 percent sure about get through.

That makes confidence a sensitivity dial:

  • A higher threshold means fewer, more certain detections. Good when false alarms are costly and you only want strong hits.
  • A lower threshold catches more, including faint or partial views, at the price of more false positives.

The recommended starting point is 35. Tune it per camera from there: a busy scene with lots of partial views may need a different setting than a clean, head-on entrance.

How Severity Sorts What Matters​

Not every detection deserves the same reaction. Severity Level lets you rank classes as Low, Medium, or High, so the system knows the difference between "worth logging" and "act now." A gun at High and a parked car at Low both get detected, but severity is what pushes the serious one to the top of the Operator Dashboard and the alert lists. Set it to match the real-world stakes of each class on that camera.

Where You Configure Detection​

Detection is set in two places, and they work together:

Start broad at the portal level, then sharpen each camera to its job.

Detection Versus Indexing​

It's worth being clear about what detection here is not. Live detection runs on the surveillance service and works on the camera feed as it happens. That's separate from the VIDIZMO indexing service, which detects objects in uploaded video and documents after they're processed. The two share many of the same object classes, but live surveillance acts in the moment, on the live stream, rather than on content already in the library.

To read about the underlying object models in depth, see the object detection articles in Content Management, linked below.

Did you find this article helpful?