Think Build Implement Repeat
London, UK +44 7367 067226
WhatsApp FOLLOW f in X
AI & Machine Learning

Computer Vision for Retail Shelf Monitoring

Last updated:

The problem is not stock, it is visibility

Most shops with an empty shelf have the product in the building. It is in the back room, on a cage by the goods-in door, or sitting behind a row of the wrong variant. The till system thinks there are twelve units. The customer sees none and buys elsewhere.

Shelf monitoring with computer vision tries to close that gap by looking at the shelf directly rather than inferring its state from sales and deliveries. A camera captures the aisle, a model finds each product facing, and the system compares what it sees against what should be there. When a gap appears, someone gets a task on their handheld.

It is one of the more mature computer vision use cases in business, which means the techniques are well understood. It also means the vendor pitches have had years to get polished, so it pays to know where the edges are.

What a shelf monitoring system can detect

  • Empty or low facings. The most reliable output by some distance. A gap on a shelf edge is visually obvious and the model does not need to know what should have been there to spot it.
  • Misplaced products. A tin of chopped tomatoes in the passata slot. This needs the model to recognise individual products, which means training on your range.
  • Planogram compliance. Whether the layout matches the agreed plan, facing counts included. Useful for brands paying for placement and for head offices auditing branches.
  • Price label mismatches. Reading the shelf-edge label and checking it against the product above. Harder than it sounds, because labels are small, angled and often partly covered.
  • Promotional display checks. Whether the end-cap went up on Monday as agreed.

Gap detection can be good within weeks. Product-level recognition across thousands of lines takes considerably longer, and the effort scales with how often your range changes.

Fixed cameras, robots or phones

There are three ways to get images of shelves, and the choice drives cost more than the model does.

Capture methodGood forWatch out for
Fixed shelf camerasHigh-value aisles needing near real-time alertsHardware and cabling per bay, power, angles blocked by trolleys
Ceiling cameras (often existing CCTV)Broad gap detection on long runsLow resolution at distance; usually cannot read labels
Staff with a phone appSmall estates, periodic auditsDepends on people doing the walk; images vary wildly
Autonomous scanning robotLarge stores with wide aislesExpensive, slow, and customers find them odd

For an independent grocer or a chain of 20 convenience stores, a phone-based audit twice a day usually beats fixed cameras on return. The model is the same. The difference is you are not paying to wire 300 shelf bays.

How much does shelf monitoring cost?

Illustratively, for a chain of 15 mid-sized stores: a phone-based capture app plus a gap detection model is a project of a few months, with running costs dominated by staff time on the walk and modest cloud inference. Fixed cameras add hardware per aisle, installation, and a maintenance round every time a shelf is refitted.

The model training is rarely the expensive part. The expensive parts are product image collection when you want item-level recognition, integration with your stock and task systems, and the ongoing work of keeping the product library current as lines come and go.

Our rule of thumb when we scope these at SpiderHunts is to price the first phase around gap detection only. If that alone moves availability, item recognition becomes a much easier conversation.

Where it goes wrong

  • Deep shelves. A camera sees the front facing. If staff pull stock forward, the shelf looks full with two units left behind it.
  • Loose produce and bakery. Irregular items in crates are hard to count and harder to judge as low.
  • Constant range changes. Seasonal lines and own-label packaging refreshes quietly break product recognition.
  • Glare and lighting. Chiller doors, shiny packaging and LED strips produce reflections that look like gaps.
  • Nobody acting on the alert. The most common failure. A perfect detection that sits unread for four hours has no value.
The model finds the gap. The rota fills it. Budget for the second part as seriously as the first.

When shelf monitoring is the wrong answer

If your stock records are poor, fix those first. A camera that reports empty shelves in a store where the system also shows zero stock is telling you something you could learn from a report.

It is also a poor fit for very small shops where a manager walks every aisle hourly anyway, and for stores with little range stability. A better replenishment trigger from sales data, sometimes built with ordinary machine learning on your till data, can catch most of the same gaps without a single camera.

A sensible first project

  1. Pick the 30 lines where an empty shelf costs the most in lost sales
  2. Photograph those bays at fixed times for two weeks and label gaps by hand
  3. Train a gap detector and measure how often it agrees with the labels
  4. Push alerts to one store's handhelds and track time from alert to refill
  5. Compare availability against a similar store with no alerts

That last comparison is the honest test. Detection accuracy is a model metric. Recovered sales is the business one, and it is the only number the operations director will care about. Our notes on AI in retail operations cover the stock and rota side in more depth.

Frequently asked questions

Can we use our existing CCTV for shelf monitoring?

Sometimes, for broad gap detection on long runs. Security cameras are usually mounted high and wide, so they rarely have the resolution to recognise individual products or read price labels. A short test on recorded footage will tell you quickly.

How accurate is out-of-stock detection with cameras?

Gap detection on front facings is generally reliable in reasonable lighting. Accuracy drops for deep shelves, chillers with glass doors and loose items. We measure it against hand-labelled photos from your own stores before anyone relies on it.

Does shelf monitoring raise privacy issues?

It can, because shoppers and staff appear in the images. Cameras pointed at shelves, blurring of people before storage and short retention periods keep the risk low, but you still need to cover it in your privacy notice and data protection assessment.

How many product images do we need for item recognition?

It depends on how similar your products look. Distinct packaging needs relatively few examples per line; near-identical variants such as flavours of the same brand need many more, plus shelf-edge label reading to separate them.

Keep reading

Losing sales to empty shelves you only notice at closing?

Send us a few photos of a typical aisle at a busy hour. We will tell you what a camera could reliably detect there, and whether a simpler fix would do the job for less.

Book a free 30-minute call Get a project estimate WhatsApp us

Related services

What we build for problems like this one

AI AgentsMachine LearningAI Integration