Vision Field Notes

Privacy and Ethics in Camera Analytics

Camera analytics turns video of real people into data. Even a simple people counter processes images of individuals, and a badly designed system can collect far more than it needs. Engineers often treat privacy as a legal department problem. In practice, many of the most effective privacy protections are design decisions that only the engineers can make: what to store, what to discard, who can see it, and what the system is allowed to infer.

This article is general engineering guidance, not legal advice. Laws about cameras, biometric data and workplace monitoring differ a great deal by country and region, and they change. Talk to a qualified professional and check the rules that apply to your site before you deploy.

Start with the purpose

Write down, in one or two sentences, what the system is for and what it is not for. "Count visitors entering the shop to plan staffing" is a clear purpose. "Monitor everything on cameras and see what is useful" is not. A clear purpose does three things:

Ask whether video analytics is needed at all. If a door sensor, a ticket system or a weight sensor answers the question, it may collect less sensitive data.

Data minimisation

The best protection for data is not collecting or not keeping it. Concrete design options:

Two versions of a store aisle frame. Left: the live frame with three people boxed. Right: the stored version with the left part of the frame masked grey and the three people pixelated. Below: what the counting purpose needs long-term, people in aisle equals 3, time stamp and camera id.
Caption: The live frame is analysed in memory; anything kept for review is masked and pixelated first, and for a counting purpose the long-term record is just a number. Footage: Intel IoT DevKit sample videos (CC BY 4.0). Detections: YOLOv8n.

Retention

Decide how long each kind of data is kept and enforce it automatically:

Data Typical approach
Live stream Processed in memory, not stored
Alert snapshots and clips Kept for a short, defined period needed to review the alert
Aggregated counts and statistics Kept longer, since they are not personal
Training and test footage Kept only as long as needed, access-restricted, and anonymised where possible
Logs Avoid putting personal data into logs

Do not rely on someone remembering to delete files. Use scheduled deletion and verify that it works. Include backups and copies on developers' laptops in the picture, because those are where old footage tends to survive.

Flow diagram: camera to edge analytics, where raw video is not stored. Events go to alert snapshots (blurred, role-based access, short period) and results go to aggregates (kept longer). Both flow to deletion on a schedule, including backups and dev copies.
Caption: Each kind of data has its own path and retention, and deletion is automatic. No specific periods are shown because they depend on purpose and local rules.

The right retention period depends on your purpose and on local rules, so set it with the site owner and your advisers instead of using a default number.

Transparency and signage

People should know when they are being filmed and why. Typical measures include clear signs at entrances to monitored areas, stating who operates the system and how to contact them, plus a privacy notice describing what is collected, why, how long it is kept and who can access it. In workplaces, communicate with employees before deployment, and involve their representatives where that is customary or required. Requirements for notice and consent differ by jurisdiction and by context (public spaces, workplaces, private property), so confirm locally.

Consent is not always practical or the right legal basis for cameras in shared spaces, and requirements vary. Do not assume a sign alone is sufficient everywhere.

Access control and security

Video is sensitive data, and a leak can harm people. Basic security hygiene applies:

Bias and fairness

Detectors are trained on data, and their performance can differ across groups and conditions. Possible problems include lower recall for people with certain skin tones in poor lighting, for children, for people using wheelchairs, for unusual clothing or for groups under-represented in the training data. Consequences get worse when output is tied to decisions about individuals.

What an engineer can do:

Function creep and misuse

A system built for one thing tends to be asked to do another. A counter becomes a productivity monitor; a safety alert becomes a disciplinary record. Reduce this risk by writing the purpose into the contract and documentation, limiting the data the system produces to what the purpose needs, and refusing, or at least re-assessing, new uses before they happen. Technical limits help: if the system never stores identity, it cannot be turned into an identity tool without visible changes.

A practical assessment

Before deployment, many organisations carry out a data protection impact assessment or similar review. Even where it is not required, a short written review with these questions is valuable:

  1. What is the purpose, and is there a less intrusive way to meet it?
  2. What data is collected, and is each item necessary?
  3. Who is affected, and how will they be informed?
  4. How long is data kept, where, and who can access it?
  5. What could go wrong (leak, misuse, misdetection harming someone)?
  6. How can a person ask questions, object, or request access to data about them, where the law gives them that right?
  7. Who is accountable, and how are incidents handled?

Checklist

Next steps: apply minimisation to a concrete task in People Counting with Line Crossing: A Practical Tutorial, and see how data volume and design interact in Scaling Video Analytics to Hundreds of Camera Streams.