What Computer Vision Monitoring Actually Is

Video analytics is sold as one capability. It is at least three, with different accuracy, different economics and — most importantly — completely different legal positions.

The three categories

Detection. Is there an object of this class in this frame? A person, a vehicle, a pallet, a crack, a spill. No identity, no individual.

Tracking and counting. How many, moving where, dwelling how long. Individuals are distinguished from each other within a scene but not identified.

Identification. Which specific person is this? Matching a face or gait against a known set.

They are not points on one scale. The jump from the second to the third changes the legal analysis completely, and a system that does the first two well tells you nothing about whether it can do the third.

Why the distinction matters immediately

Detection of a person in a hazardous zone needs no identity and is among the strongest justifications in the field.

Counting occupancy needs no identity and can discard person-level data at source.

Identifying who was in the zone is biometric processing of special-category data, engages a body of law, and is rarely necessary for the operational question.

Most deployments ask for the third when the first would answer the question, which is where both the cost and the trouble come from.

What has changed technically

Worth being specific, because the claims and the reality diverge.

Object detection is genuinely good on common classes in reasonable conditions, and has been for some years.

Detection of rare or unusual classes is much weaker, and every real deployment cares about the unusual.

Performance in the field is well below benchmark performance, for reasons covered in their own note: lighting, angle, occlusion, weather, camera quality, and the fact that a benchmark set is curated.

Anything that infers an internal state — intent, emotion, attention — is not a solved problem and is not close to one. It is also prohibited in several contexts.

What it is not

Not a replacement for supervision. A system that flags events still needs someone to act on them.

Not evidence by itself. A detection is a probabilistic output with a threshold behind it.

Not accurate in the way the demonstration suggests. Demonstrations run on footage the system was tuned for.

Not a mind-reading device, whatever the vendor category is called.

The framing that keeps this useful

Ask what decision the output informs, then ask the least capable system that could inform it.

An alarm when someone enters a machine area needs detection. Understanding queue length needs counting. Neither needs to know who anyone is.

Deployments that start from what the technology can do rather than from the decision end up with identity data they did not need, obligations they did not plan for, and an accuracy problem they cannot characterise.

The first question to answer

Before any demonstration, one sentence written down.

What decision will this system inform, and who makes it?

Then: what is the least capable tier that could inform it — detection, counting, or identification?

Then: what happens when the system is wrong, in each direction?

Circulate it. The question frequently changes once someone senior reads it, and that is much cheaper before a procurement than after.

A proposal that cannot answer these three is not ready, whatever the technology can do.

Who this is for

Three audiences with different reasons to be here.

Whoever has been asked to deploy something and needs to know what the technology actually does before committing.

Whoever has to assess it — risk, legal, safety — and needs the technical picture in terms that connect to the obligations.

Whoever is subject to it, who is entitled to know what is being claimed and what is being measured.

These notes are written for the first, with the second's questions answered where they arise and the third's interests treated as a design input rather than an obstacle.