Computer vision is software that analyses images or video to recognise objects, states and events: how many products have passed along the conveyor, whether a part has a defect, whether an aisle is blocked. For these tasks you don't need to know who a person is, only what is happening.
When people talk about cameras and artificial intelligence, many immediately think of facial recognition. In practice, many of the most useful applications for a company work without it. This article explains where computer vision is actually used, which technical conditions matter and why biometric identification stays out of ordinary projects.
In short
Common applications are anonymous counting, visual inspection, event detection and operational monitoring. The result depends heavily on real images and filming conditions (light, angle, resolution). Local (edge) processing limits the images that leave the building, and decisions with impact stay with a person.
Practical computer vision applications without identifying people
Most computer vision projects in manufacturing, logistics and retail fall into one of the categories below. All of them can work without the system learning anyone's identity.
- Anonymous counting: products on a conveyor, pallets at a loading dock, vehicles at a barrier or the flow of visitors in a shop, as simple aggregate numbers, without tracking specific people.
- Visual inspection: checking parts, labels or packaging for visible defects, as support for quality control.
- Event detection: someone entering a restricted area, an object fallen in an aisle, a door left open, an empty shelf.
- Operational monitoring: how full an area is, queues at checkouts or loading docks, adherence to a workflow visible in the image.
In manufacturing, the focus is usually on inspection and counting. In logistics, on events at the loading docks and in the warehouse. In retail, on aggregate flows and the state of the shelves. The difference between them is not the technology, but the question the system has to answer.
What data and images are needed for training
A computer vision system learns from examples. To detect a defect, it needs images of good parts and defective parts, in the real conditions it will work in. To count pallets, it needs images from the angle of the camera that will be used.
- Real images from the place of use, not catalogue photos.
- Sufficiently varied examples: day and night shifts, different seasons, different products.
- Examples of the situation being looked for, including rare ones, such as less frequent defects.
- A written definition of what “defect”, “event” or “occupied” means, agreed with the people who normally make the call.
- Agreement on how the images used for training are stored, who has access and when they are deleted.
Labelling the images, that is, manually marking what appears in them, is often the part that takes the most effort. It is worth planning from the start, together with the company's specialists who know what a good case and a bad one look like.
Cameras and filming conditions: light, angle, resolution
A good model can't make up for a poor image. That's why the discussion about cameras comes before the discussion about algorithms.
- Light: steady and sufficient. Reflections, backlighting and sudden changes between day and night reduce accuracy.
- Angle: the camera must see the area of interest clearly, without frequent obstructions. A camera mounted for general surveillance rarely has the right angle for inspection.
- Resolution: enough for the detail being looked for. A small scratch needs a different camera and distance than counting boxes does.
- Speed: for moving objects, the frame rate and exposure time matter, so the image isn't blurred.
- Stability: a camera knocked out of position can ruin a setup that was working well.
Practical tip
Before any project, record a few days of footage from the exact spot where the system would work. You can review it with the technical team and quickly see whether the problem can be solved with your existing cameras or whether you need different ones.
Edge or cloud: where the analysis runs
Image analysis can run on a local device close to the camera, known as edge processing, or on servers in the cloud. The choice affects costs, reaction speed and privacy.
Local processing has the advantage that images don't need to leave the building: the device passes on only the result, for example a number or an alert. Reaction is fast and doesn't depend on the internet connection. On the other hand, local hardware has to be maintained and has limited computing power.
The cloud offers more flexibility and is easier to update, but it involves transmitting images or parts of them, with implications for bandwidth and data protection. Many projects use a combination: routine analysis locally, and in the cloud only aggregate data and images selected to improve the model.
Privacy by design and why we exclude biometric identification
Privacy by design means that protecting people is considered from the start, not added at the end. In concrete terms: the camera sees only the area needed, images are processed locally where possible, only the results are kept, and faces or other details that aren't needed can be blurred automatically.
Facial recognition and other forms of biometric identification are not part of our ordinary projects. The reason is simple: biometric data is considered especially sensitive, and using it raises serious issues of lawfulness, necessity and proportionality. In almost all operational cases, the objective can be achieved without knowing who the person in the image is. Any request of this kind would require a separate, careful assessment with legal specialists.
For information, not legal advice
Using video cameras in the workplace or in spaces open to the public comes with personal data protection obligations: informing people, a legal basis, purpose limitation, retention periods and, in some cases, an impact assessment. The rules may apply even if the system doesn't identify anyone. This article is for information only; for your company's specific situation, consult a data protection specialist or a lawyer.
The limits of computer vision and human confirmation
No computer vision system is perfect. It will sometimes miss a defect or flag a problem that doesn't exist. That's why its proper role is to flag, while the decision stays with a person: the operator who checks the rejected part, the shift supervisor who confirms the alert, the manager who interprets the flow data.
- Accuracy drops when conditions change compared with the training images: a new product, different lighting, a moved camera.
- Very rare situations are hard to learn, because there are few examples.
- Covered or overlapping objects are hard to count correctly.
- The system needs monitoring and periodic adjustments, not a one-off setup.
A sound project establishes from the start how performance is measured: how many false alerts occur, how many real cases are missed, how long human confirmation takes. These measurements are tracked in operation, on real data.
It is worth starting with a pilot phase, on a single line, dock or area, in which the system runs in parallel with the current process. The team compares the flags with what people observe, and only then is expansion decided. This way the limitations come to light early, while the cost of a correction is still small.
How we approach a computer vision project
At WEB-AZ we start from the operational question, not the technology: what needs to be observed, how often and what is done with the result. We check the filming conditions, propose local processing where it makes sense, keep only the data needed and build the flow so that important decisions stay with people. If you want to find out whether your problem can be solved with your existing cameras, an initial assessment is a good starting point.
Next steps