Combining Shelf Cameras With POS Data for Better Stockout Signals

A quiet shelf has two explanations

A product that is not selling could be out of stock, or it could just be unpopular. A camera that sees a gap could be looking at a genuine stockout, or at a temporary hole a shopper made that will fill from behind. On its own, neither the shelf camera nor the point-of-sale data is conclusive. Shelf vision plus POS data is about using each to resolve the other.

What each source knows

The camera knows the current physical state of the shelf — full, gapped, empty — but not why, and not what the system thinks is in stock. POS knows what has sold and how fast, and the inventory system knows the on-hand count, but none of them can see the shelf. Fused, they tell a fuller story than any one alone.

The high-value case: phantom inventory

The combination that matters most is: the camera sees an empty facing, the system shows stock on hand, and POS shows no sales for hours. That pattern is phantom inventory — the item is lost, misplaced, or miscounted, so it is neither on the shelf nor selling, while the system believes it is available and will not reorder it. This is a loss that pure inventory data cannot surface because the count looks fine.

Disambiguating an ordinary gap

When the camera sees a gap but POS shows the item still selling steadily, it is probably a transient hole with stock behind it — lower priority. When the camera sees a gap and sales have stopped dead, it is a real stockout to act on now. The sales signal turns a raw detection into a prioritised task.

The fusion sits above the edge

The detection still runs on-device as in our computer vision project, sending a structured shelf-state event. The fusion with POS and inventory data happens centrally, where those systems already live, and produces a smaller, better-qualified list of actions than the camera alone would.

Expected sales rate as the tuning knob

The comparison needs an expected sales rate per product per store — roughly how often this line normally sells here. A gap on a line that sells every ten minutes with no sales for an hour is urgent; a gap on a line that sells twice a day is not. That rate is derived from POS history and is what makes the prioritisation sensible.

A worked example

A camera flags an empty facing on a mid-range cereal. The fusion layer checks: the inventory system shows fourteen units on hand, and POS shows no sale of that SKU since the store opened four hours ago. That triad — empty shelf, stock on hand, no sales — is raised as a phantom-inventory investigation, not a routine refill, and a supervisor finds a case of it left on a bottom pallet in the wrong aisle.

Keeping the two clocks aligned

The camera's view of the shelf and the POS or inventory snapshot have to refer to roughly the same moment for the fusion to mean anything. A gap detected now, compared against a sales total from a batch that last updated this morning, can point the wrong way — the item may have sold out in the last hour, which the stale figure hides. The fusion layer needs reasonably fresh data on both sides, and where a feed is only updated in batches, the logic has to account for that lag rather than treat every number as current.

Where this stops being right

• Without a reliable expected-sales-rate per line, the prioritisation degrades to raw detections again.
• Stores with poor inventory accuracy get noisy phantom-inventory flags; the technique assumes the on-hand count means something.
• Very low-velocity lines produce weak sales signals and are better handled by periodic manual checks.

FAQ

What does adding POS data to shelf vision achieve? It explains the gap. Sales still flowing means a transient hole; sales stopped means a real stockout; stock on hand plus no sales plus an empty shelf means phantom inventory.

What is phantom inventory and why does it matter? It is stock the system believes is available but that is not on the shelf and not selling — lost or misplaced. Inventory data alone cannot see it because the count looks correct, so it never triggers a reorder.

Where does the two data sources get combined? Centrally. The camera sends a structured shelf-state event from the edge, and the fusion with POS and inventory happens where those systems already run.

ISTRALLEN builds retail computer vision that fuses shelf detections with POS and inventory data to surface real stockouts and phantom inventory — see AI for Retail.