Skip to content

Use the least capable system that answers the question

An alarm when someone enters a machine zone does not need to know who entered. Fifty notes on detection, counting and identification — what each can do, where the errors fall, and why most questions posed as “who” are answerable as “how many”.

All notesStart at the beginning

All 50 notes

Grouped by section, each appearing once. The label above each note says what kind it is.

Foundations

6 notes

Three capabilities share the name and differ enormously in what they do and what the law permits. Separating them is the first task, and it decides everything after.

What it cannot see

5 notes

An average accuracy figure conceals that the errors are not evenly distributed. Plus the physical conditions that no software fixes, and the drift that happens quietly.

Where it works

5 notes

The deployments with a real record, most of which point cameras at objects rather than people, and the ones where a supplement is substituted for the control that should have been installed.

The boundary

6 notes

The section this subject requires. What should not be built, why the objections converge rather than trading off, and what to offer instead when the request is reasonable underneath.

Regulation

5 notes

The prohibitions are already in force. What each tier requires, why consent does not work in a workplace, and the records that cannot be reconstructed afterwards.

Building and evaluating

9 notes

Camera placement is the highest-return investment and appears in no vendor proposal. Plus the evaluation that means something and the threshold decision that is a business call, not a setting.

Running it

9 notes

Where the analysis runs determines what leaves the building. Plus the alert arithmetic that decides whether anyone reads the output, and the systems nobody remembered to turn off.

Reference

5 notes

The boundary stated plainly: the questions a camera cannot answer, the asks that should be declined, and a description of the end state to measure a proposal against.

What this is

Fifty notes on camera-based monitoring, written for the people who have to specify, assess and operate it.

No vendor material, no product recommendations, no sponsored content.

Four things that hold across deployments

Three capabilities share the name. Detection asks whether something is present, counting asks how many, identification asks who. The jump to the third changes the legal position completely, and most deployments ask for it when the first would answer the question.

An accuracy figure conceals more than it reports. It hides the threshold, the base rate, the conditions, and — most importantly — that the errors are not evenly distributed across people.

Camera placement decides more than the model. Most disappointing deployments are disappointing for physical reasons, and no amount of tuning fixes a camera pointed at the wrong thing.

The alert arithmetic is the deployment. A false alarm rate that seems small per camera becomes an alert every few minutes across an estate, and a channel nobody reads is not a control.

Where to start

New to this: the foundations, particularly the note separating detection from identification.

A demonstration was impressive: the limits section.

Something involving identification or inference has been proposed: the boundary section, before anything technical.

You have cameras and a question: the note on where to start, which is ninety days without a purchase.

What this is not

This is not legal advice. Descriptions of the AI Act, data protection and biometric statutes are general and vary by jurisdiction.

It does not describe how to deploy covertly, how to infer emotion or intent from video, or how to categorise people by sensitive characteristics. Those are covered only to explain why they should not be built.