Processing at the Edge or in the Cloud

The deployment architecture is usually presented as a cost and latency question. Its most important consequence is what data exists outside the site.

Edge processing

Analysis on the camera or on a local device.

Only the derived output leaves: a count, a detection, an event.

Frames never transmitted, and in a well-designed system never stored.

Low latency, which matters for real-time alerting and interlocks.

Works without connectivity.

Constrained compute, so models must be smaller and are usually less capable.

Updates are harder across many devices.

Cloud processing

Video transmitted to a central service.

More capable models, unconstrained by device hardware.

Easier updating and central management.

Bandwidth cost, which for many cameras is substantial and continuous.

Latency, which may or may not matter.

And the significant one: your footage of people is now in someone else's infrastructure, with everything that implies for transfers, subprocessors, retention and breach exposure.

The privacy consequence

Edge processing that discards frames removes most of the concern in one architectural decision.

"We count people and transmit only hourly totals" is a genuinely different proposition from "we stream all our cameras to a third party", and it is much easier to explain and to defend.

Ask the vendor what is transmitted, specifically: frames, crops, templates, or derived events only.

Ask what is retained on their side, and for how long, and check the default.

A product that requires video to leave the site for a counting application is answering the question in an unnecessarily intrusive way, and that is worth saying during procurement.

Cross-border transfer

Where the processing happens matters legally, not just physically.

Footage of identifiable people transferred outside the jurisdiction requires a transfer mechanism, and the analysis is more demanding for special-category data.

Ask where the processing occurs and where the data rests, and get it in the contract rather than in a sales conversation.

Subprocessors matter: the vendor's own cloud provider, their support arrangements, their model-training practices.

Ask explicitly whether your footage is used to improve their models, which is common, frequently buried, and is a separate purpose requiring its own basis.

The hybrid arrangement

What most mature deployments end up with.

Edge detection, cloud aggregation. The device does the analysis and sends events; the central system stores and reports on events rather than video.

Video retained locally on the existing recorder, under the existing retention.

Clips exported centrally only on a defined trigger, with the retrieval logged.

This gets the management benefits without the continuous transmission, and it is the arrangement to ask for.

Choosing

If the application is counting or condition detection: edge, and discard frames.

If latency matters for an interlock or alarm: edge.

If connectivity is unreliable: edge.

If the model genuinely cannot run on available hardware: reconsider whether the application needs that model, then consider cloud with a strict transfer and retention arrangement.

Default to edge and justify any departure, which is the reverse of how these projects usually run.

The training-on-your-footage clause

Common, frequently buried, and a separate purpose requiring its own basis.

Ask explicitly whether customer footage is used to improve the vendor's models.

Get the answer in the contract, not in a sales conversation.

Note the consequence for deletion: data absorbed into a trained model cannot meaningfully be removed later, which makes this a decision that cannot be reversed.

Where the answer is yes and non-negotiable, treat it as a reason to decline for anything involving identifiable people.

Asking what is transmitted

A specific question with a specific answer, and it determines most of the privacy analysis.

Frames? Crops? Templates? Derived events only?

At what rate, to where, retained for how long?

On a data flow diagram, which a product that cannot produce one is telling you something about.

Verify it by looking at the network traffic during the pilot rather than trusting the answer.

"Derived events only" is a genuinely strong position and it is worth confirming rather than assuming.