Computer Vision for Forecourt and Petrol-Station Shops

A shop that runs on one person

The shop attached to a petrol station is often staffed by a single person who is also taking fuel payments, and it is rarely anyone's operational priority. Fast lines — drinks, snacks, tobacco, coffee-to-go supplies — empty between the checks that never happen. Computer vision for a forecourt shop fits that reality: a small, cheap deployment that tells the one person on site what needs attention without adding to their load.

The staffing constraint shapes everything

With one person behind the till, an alert has to be worth interrupting a fuel transaction for. That pushes hard toward high-confidence, low-volume alerting — the confidence-based routing from our computer vision project, tuned so only a clear, high-value gap becomes a notification, and everything ambiguous waits for a quiet moment or a review queue.

A small camera budget

A forecourt shop is small and the investment case is modest, so the design has to work with a handful of cameras — often one per key bay rather than full coverage. Prioritising the two or three lines that lose the most revenue when empty, and the coffee station, covers most of the value.

High-shrink categories

Forecourt shops carry a concentration of high-shrink lines. Shelf monitoring is not loss prevention, but a facing that empties far faster than sales explain is a signal worth surfacing to the operator or area manager as an operational flag, distinct from any people-facing security system.

On-device, low-connectivity

Forecourt sites frequently run on a basic broadband or mobile link shared with the payment terminals. On-device inference and a small structured event rather than an image upload keeps the monitoring off that constrained connection entirely.

Opening hours and staffing patterns

A 24-hour site with a lone overnight worker has different needs from a daytime-only shop. Alert timing and volume should adapt to the shift — quieter overnight, with a prioritised catch-up list for the morning team rather than a night of notifications one person cannot act on.

Area-manager visibility

One operator cannot see trends across a week. Rolling the detection events up to an area manager — which sites run out of what, how fast alerts are actioned — turns a single-site tool into an estate signal, which is where a forecourt network gets the real return.

A worked example

Mid-morning, the chilled energy-drink bay empties while the sole staff member is serving a fuel queue. A high-confidence alert lands on their device. When the queue clears, they refill from the back chiller instead of finding it empty at the next delivery. Overnight, a lower-value gap on a slow line is logged for the morning shift rather than pinging the lone night worker.

Riding the existing site infrastructure

A forecourt already has power, mounting points, network, and often CCTV cabling in the shop area. A shelf-monitoring add-on that reuses that infrastructure — and reports into the same area-management portal the fuel and payment data already feed — is far cheaper to roll out across a network than a standalone install per site. The rollout becomes a small kit and a configuration step rather than an electrician and a survey at every location, which is what makes the numbers work for a category of shop that was never a big investment target.

Where this stops being right

• A single site rarely justifies the build alone; the economics work across a network of forecourts.
• Behind-the-till and locked tobacco stock is not camera-visible and needs stock controls.
• Loss prevention is a separate system with people in frame and its own legal handling — do not fold it into shelf monitoring.

FAQ

Is a forecourt shop big enough to benefit from computer vision? A single site on its own usually is not. Across a network of forecourts, where each site is lightly staffed and rarely audited, the aggregate benefit is real.

How does it work with only one person on site? Alerting is tuned to high-confidence, high-value gaps only, with ambiguous detections queued. The one metric that matters is whether those few alerts get actioned.

Does it need good connectivity? No. Inference runs on-device and only a small event is sent, so it does not compete with the payment terminals for the site's link.

ISTRALLEN builds computer vision for forecourt networks that fits a single-staff site and rolls up to area management — see AI for Retail.