Computer vision in retail: use cases, limits and the EU rules
Computer vision in retail is the use of cameras and image recognition to see what happens on the shop floor and in the back room: empty shelves, wrong price labels, long queues, how many people came in, which product is on the scale, and what was missed at a self-checkout. This guide explains the use cases that work today, the limits that decide whether a project succeeds, what the GDPR and the EU AI Act allow, and how to run a pilot that gives you a real answer before you roll anything out.
It is written for retail operations, loss prevention and IT managers. It names products neutrally and does not quote prices.
The short version
- The use cases that pay back most often look at products and shelves, not at people: out-of-stocks, planogram and price label checks, and product recognition at checkout.
- Use cases that look at people can be done anonymously, such as footfall and queue length, if the system is designed for it from the start.
- Recognising individuals or inferring sensitive traits is prohibited or tightly restricted in the EU. Treat it as off the table unless your lawyers say otherwise.
- Before buying anything, check whether data you already have, such as sales or stock records, answers the same question.
Shelf monitoring
An empty shelf loses a sale the customer may never come back for, and the stock record often says the item is there. Computer vision checks the shelf itself. Images come from fixed cameras facing the shelves, from cameras on a shelf-scanning robot, or from staff taking photos on a phone during a walk-through. The software recognises products and gaps and turns them into tasks.
- Out-of-stock and low stock. A gap on the shelf becomes a refill task for the right aisle, with a check against the stock record: if the system says there are units in the back room, the task is to refill; if not, it is to reorder.
- Planogram compliance. The shelf is compared with the planned layout, and products in the wrong place, facing the wrong way or missing a facing are flagged.
- Price and label checks. Shelf labels are read and compared with the price in the system, so a label that was not changed after a price update is found before a customer finds it.
The hard part is not the model, it is the catalogue. Packaging changes, new products arrive every week and private-label items look alike. A shelf system needs a routine for adding new products and checking the ones it is unsure about, or its accuracy falls month by month.
Checkout and loss prevention
- Product recognition at the scale. A camera above the scale in the fresh produce section suggests the product, so the customer or cashier picks from three options instead of searching a list.
- Self-checkout assistance. The camera compares what was placed in the bagging area with what was scanned and asks the customer to check an item that was missed, before a staff member is needed.
- Back-room and receiving checks. Incoming pallets and cases are recognised and counted against the delivery note, and visible damage is photographed for the supplier claim.
These use cases look at products and actions, not at who the customer is. Designed well, they never need to store an identifiable image of a person beyond the seconds it takes to process it.
Footfall, queues and store layout
Counting people at the entrance gives you conversion: transactions divided by visitors. Measuring queue length lets the system ask for a second checkout to open before the queue gets long. Anonymous movement heatmaps show which parts of the store people pass and which they skip. All three can be done without identifying anyone: the camera or sensor processes images on the device, keeps only counts and positions, and discards the image. Overhead sensors built for counting do this by design. That design choice is what keeps these use cases on the right side of the GDPR.
What the EU rules allow
Two sets of rules shape computer vision in European stores.
The GDPR. Video of identifiable people is personal data, and every use needs a lawful basis, a clear purpose, a retention period and signs that tell people what is happening. When images are processed to uniquely identify a person, for example through facial templates, the data is biometric data, a special category under Article 9 of the GDPR, which can only be processed under narrow conditions. The European Data Protection Board's Guidelines 3/2019 on processing of personal data through video devices explain how these rules apply to cameras. A data protection impact assessment is the first step of any project that analyses people on camera.
The EU AI Act. According to the European Commission's summary of the AI Act, the prohibitions that took effect in February 2025 include emotion recognition in workplaces and education institutions, biometric categorisation to infer certain protected characteristics, and untargeted scraping of the internet or CCTV footage to build or expand facial recognition databases. AI systems used for remote biometric identification, emotion recognition and biometric categorisation are listed as high-risk use cases. The Commission states that the AI Act became applicable on 2 August 2026, with some exceptions, and that high-risk systems in the sensitive areas of Annex III follow on 2 December 2027 under the AI Omnibus amendments. (Commission page read on 11 November 2026.)
In practice, for a retailer this means:
- Reading the emotions of staff on camera is prohibited. Analysing staff performance on camera at all raises employment law questions and usually needs agreement with employee representatives.
- Estimating protected characteristics of customers from their images is off limits.
- Facial recognition watchlists sit in the most restricted category under both laws.
- Shelf, product, counting and queue systems that never identify a person are the safe ground, and that is where most of the business value is.
We do not give legal advice. Your data protection officer or lawyer decides what your stores may do; we design the system so the privacy choices are built into how it works, not added in a policy afterwards.
Privacy by design, in concrete terms
- Process at the edge. Images are analysed on a device in the store, and only the result, a count or an alert, leaves the building.
- Look at shelves, not faces. Cameras for shelf monitoring point at the shelf, at a height and angle that keeps faces out of the frame.
- Blur before storing. When an image must be kept, for a refill task or a supplier claim, people in it are blurred automatically first.
- Short retention. Images used only to produce a count are deleted within seconds. Counts and alerts are kept for as long as the business needs them.
- EU hosting. Any data that leaves the store is processed in an EU data region, under a written data processing agreement.
When computer vision is the wrong tool
Some questions are cheaper to answer with data you already have. Forecasting demand from sales history is one of them; see predictive analytics for retail. A product that has sold every day and suddenly sells nothing for a day is probably out of stock, and sales data shows that without a camera. Electronic shelf labels remove the label-checking problem. RFID tags count stock in fashion retail more reliably than images of folded clothes. Before a computer vision project, ask which of these already answers the question; if one does, start there.
Ready-made products or a custom build?
Ready-made products exist for most use cases: people counting sensors, shelf-scanning services such as those from Trax and Focal Systems, self-checkout vision from checkout system vendors and scale recognition from scale manufacturers. For a standard need in a standard store layout, a product is usually faster. A custom build makes sense when the use case is specific to your business, when you need the results in your own systems rather than in a vendor's dashboard, or when you want to own the models and the data. Many projects combine both: a product for counting, and custom software that brings the counts, sales and staffing data together.
How to run a pilot
- One question, one store. Pick the single question you most want answered, such as how often a top-selling product is missing from the shelf, and one store to test it in.
- Privacy assessment first. Write down what is filmed, what is kept, for how long and who sees it, and have your data protection officer approve it.
- Measure the ground truth. For two weeks, staff record the answer by hand at set times. That is what the system will be measured against.
- Install and compare. Run the system for the same period and compare its answers with the manual record, item by item.
- Decide on numbers. If accuracy and the value of the answer justify it, plan the rollout. If not, you have learned that for the price of one store.
How DT Nova helps
We help retailers decide whether computer vision is the right tool for a question, design pilots with privacy built in, and build the software that turns camera output into tasks and reports in the systems you already use: refill tasks to staff phones, alerts to store managers, and footfall next to sales and staffing in one report. Hosting runs in an EU data region and the code is in your company's own account. See AI consulting, data analytics consulting and inventory management software. The same techniques track players and the ball in sport; see artificial intelligence in sports.
Questions people ask
What is computer vision in retail?
It is software that reads images from cameras in stores, warehouses or product photos and turns them into data: an empty shelf, a missing price label, a queue that is getting long, the number of people who entered, or the product on a scale. The output is a count, an alert or a task, not a recording someone watches.
Is it legal to use cameras that analyse customers in the EU?
It depends on what the system does. Counting people anonymously, detecting empty shelves or measuring queue length can be designed so no one is identified. Recognising individual faces, inferring sensitive traits or reading staff emotions is either prohibited or tightly restricted under the GDPR and the EU AI Act. A data protection impact assessment comes before any pilot.
Can we use facial recognition to stop known shoplifters?
In the EU this is one of the hardest cases. Facial templates used to identify people are biometric data, a special category under the GDPR, and remote biometric identification is listed as high-risk under the AI Act. Most retailers who look at it closely choose other loss prevention measures. Take legal advice before considering it.
Do we need new cameras?
Sometimes. Existing security cameras are often mounted for overview shots and too far from the shelf to read labels. Shelf monitoring usually needs cameras pointed at the shelf or images from staff phones on a walk-through, while footfall counting often works with a dedicated overhead sensor.
How accurate is computer vision in a store?
It depends on lighting, camera position, how similar the products look and how often packaging changes. Accuracy should be measured in your own stores against counts made by people, before anyone relies on it. A pilot in one store answers this better than any brochure.
Sources
- European Commission: AI Act, regulatory framework for AI (read 11 November 2026)
- Regulation (EU) 2024/1689, the Artificial Intelligence Act (read 11 November 2026)
- European Data Protection Board: Guidelines 3/2019 on processing of personal data through video devices (read 11 November 2026)