Computer Vision for Cold-Chain and Freezer Aisles

The hardest shelf to watch

Open shelves are easy for a camera compared with a freezer aisle. Between the camera and the product sit a glass door, reflections of the aisle lights, condensation, and frost, and the product packaging itself is often frosted or iced over. Computer vision for a freezer aisle is a case study in the failure modes that make shelf detection unreliable, and in what it takes to work around them.

Glass doors and reflections

A closed glass door reflects the aisle — lights, shelving opposite, sometimes the shopper. A detection model trained on open shelves sees those reflections as clutter and its accuracy drops. The model has to be trained specifically on through-glass imagery, and camera placement and angle have to be chosen to minimise direct reflections.

Condensation and frost

Doors fog when they are opened and closed in a warm aisle, and product packaging accumulates frost in the case. Both obscure the visual features a detector relies on. There is no clean fix — the practical response is a lower expectation of per-frame accuracy, more frames aggregated over time, and confidence thresholds set so the system stays quiet when the view is degraded rather than guessing.

Door state matters

A freezer door standing open is itself worth detecting — it is an energy cost and a cold-chain risk. That is a useful signal the aisle offers that an open shelf does not, and it is often more reliably detectable than the stock behind the glass.

Hardware in the cold

The camera and edge device may be mounted in or near a chilled environment, which affects component life and can cause lens fogging on the camera side too. Enclosure choice, heating, and placement are part of the install, not details.

Adjusting the routing

The confidence-based routing from our computer vision project does more work here. When the view is poor — a fogged door, heavy glare — the right behaviour is to suppress alerts for that bay until the view clears, rather than flood staff with low-confidence detections that are mostly the glass, not the stock.

What still works

Despite all this, gross stockouts — a whole facing empty behind clear glass — are detectable, and door-open events are reliable. The realistic scope is catching the large, high-value gaps and the open-door risks, not fine-grained facing counts.

A worked example

A freezer bay's most popular ice cream sells through on a hot afternoon. The facing is clearly empty behind a clean section of glass, the detection is high-confidence, and staff refill it. An hour earlier, the same door was fogged from repeated openings; during that period the system suppressed alerts for the bay rather than firing repeatedly on the condensation. Separately, a door left ajar after a restock triggers an energy-and-cold-chain alert within a minute.

Aggregating over time beats any single frame

Because any one through-glass frame is unreliable, the working approach is temporal: combine many observations of the same bay over several minutes and only raise an alert when the empty state is persistent across them. A single bad frame during a door opening, a moment of glare, or a passing shopper never triggers anything on its own. This costs a little latency — the alert comes a few minutes after the shelf actually empties rather than instantly — which is an easy trade in an aisle where the alternative is a stream of false positives that trains staff to ignore the freezer feed entirely.

Where this stops being right

• Fine-grained facing counts behind glass are not realistic; the reliable signals are gross gaps and door state.
• Heavily iced cases may be beyond useful detection until defrosted, and the system should say so rather than guess.
• Chest freezers with a top-down view are a different problem again and need their own model and camera approach.

FAQ

Why are freezer aisles so hard for computer vision? Glass doors add reflections, condensation, and frost between the camera and the product, and packaging is often iced over. All of it removes the visual detail a detector depends on.

What can it reliably detect there? Gross stockouts — a whole facing empty behind clear glass — and doors left open, which are an energy and cold-chain risk. Fine facing counts are not reliable.

How should it handle a fogged or glaring view? Suppress alerts for that bay until the view clears. Firing low-confidence detections that are mostly the glass, not the stock, is how staff learn to ignore the freezer alerts.

ISTRALLEN builds computer vision for freezer aisles that targets gross gaps and door-open risks and stays quiet when the glass defeats it — see AI for Retail.