Choose and tune a detection profile

The AI profiles

AI profiles recognise what they see, so shadows and sunlight cause far fewer false counts. They need an AVX2 processor.

• AI vehicles and people: people, cars, vans, trucks, buses and bikes, each with its own threshold. For entrances, car parks, roads and mixed traffic.
• AI people (overhead): people seen from above, for doors, corridors and gates.
• AI people (tilted): people seen at an angle, for halls, shops, footpaths and conventional security cameras.
• AI faces: estimated age band and gender for faces in view. See Age and gender estimates.

Basic motion profiles

Overhead and tilted motion profiles track anything that moves without telling a person from a car. They are about ten times lighter on the computer, so one PC carries far more of them, and they are the right choice for a simple in-and-out count on a modest machine.

Each comes in higher and lower sensitivity. Go higher if real objects are missed; go lower if shadows, reflections, plants or changing light are being counted.

Tune the detection threshold (AI profiles)

The threshold is how sure the AI must be before it accepts a detection.

1. 1

   Lower it if real objects are missed

   Typical causes: small or distant objects, objects partly hidden, poor light or a low-resolution stream.

2. 2

   Raise it if things are counted that should not be

   Shadows, reflections, signs or furniture taken for people or vehicles.

3. 3

   Move in small steps

   Change it by about 0.05 at a time and check against what really happened, at quiet and busy times, by day and by night.

4. 4

   Reset if you get lost

   The Default button restores the recommended settings for that profile.

Calibrate (basic motion profiles only)

AI profiles need no calibration. Basic motion profiles do: a ruler drawn over the picture tells the app how big a typical object is.

• Size the ruler to an average person (from above, for overhead cameras).
• If sizes vary a lot across the scene, size it to the smaller ones.
• Done well, the green tracking boxes hug each moving object. Boxes far too big or too small mean the ruler is wrong.
• On tilted cameras, set the marker as well: it tells the app the smallest and largest sizes to expect.

The area of interest

The area of interest is the part of the picture the app analyses. It starts as the whole frame; narrow it to where the counting happens. On AI profiles this cuts noise and processing; on basic motion profiles it is essential, to shut out automatic doors, lifts, trees and reflections. Double-click the boundary to add or remove corner points.

Custom objects

Need to count something that is not a person or vehicle, such as trolleys, bins or livestock? That is a model trained on your own footage, set up with you. See Custom object counting and talk to us.