Cameras already exist in most facilities for security. Applying automated analysis to them changes what they are, legally and practically.
What vision can do
Count and track objects through a space.
Detect occupancy and dwell in zones.
Read labels and codes where scanning is impractical.
Detect damage and defects on a line.
Detect safety conditions: a person in a vehicle aisle, missing protective equipment, a blocked exit, a spill.
Estimate posture, which has ergonomic applications and serious sensitivity.
The change from security to analysis
A recording camera is passive. Someone reviews it after an incident.
An analysing camera is continuous automated observation of everyone in view, producing structured data about their behaviour.
These are different activities with different obligations, and a facility that adds analytics to existing cameras has materially changed what it is doing, whatever the original notice said.
Re-do the notice and the assessment. The existing security signage does not cover it.
Where it is defensible
Object and pallet counting, which need not identify people.
Line inspection, where the subject is a product.
Safety detection with immediate output — an alarm when someone enters a vehicle aisle — where the value is preventing harm and no record of individuals need be retained.
Zone occupancy for congestion analysis, aggregated and not attributed.
In each case the design should discard person-level data as soon as the operational output is produced.
Where it should not be used
Individual productivity assessment from video. Intrusive, unreliable, and the fastest available way to destroy the workforce cooperation the whole programme depends on.
Behavioural or emotional inference, which rests on contested science and should not be deployed on employees at all.
Facial recognition of employees, which carries the strongest legal restrictions in most jurisdictions and has almost no operational justification in a facility where people already carry credentials.
Covert deployment, which is unlawful in many places and indefensible in the rest.
Break and welfare areas, which should be excluded technically rather than by policy.
Designing for minimum exposure
Process at the edge, so images never leave the camera and only counts or events are transmitted.
Discard frames immediately once the output is derived.
Blur or mask people where the subject is the product or the space.
Zone counts rather than tracks.
Exclude areas in the camera configuration, not in the reviewing software.
Retain the derived data, not the video, with a short retention for the video that remains.
The consultation point
Vision analytics on people should be consulted on before deployment, which is a requirement in several jurisdictions and good practice everywhere.
Explain what is detected and what is not, specifically. Most objections come from an assumption of broader capability than exists.
Publish the exclusions, which is what makes the assurance credible.
Expect to narrow the scope as a result of the consultation, and treat that as the process working rather than as an obstacle.
Processing at the edge
The design decision that determines how sensitive a vision deployment is.
Analyse on the camera, transmitting counts and events rather than images.
Discard frames immediately once the output is derived.
Store the derived data, not the video.
Where video must be retained, keep it short and separate from the analytics stream.
This reduces both the legal exposure and the trust cost and costs almost none of the operational value, since the value is in the counts rather than in the footage.
Explaining what it does not do
Most objection to camera analytics comes from assumed capability rather than actual capability.
Say specifically what is detected: a count, a zone entry, an object class.
Say what is not: identity, expression, conversation, attitude.
Say where the exclusions are and that they are configured in the camera rather than in the reviewing software.
Offer to show the output, which is usually a number rather than a picture and settles the matter faster than any assurance.
Publish the configuration summary, which is what makes the explanation credible.