RETAIL
Checkout Queue Monitoring
Checkout Queue Monitoring. What it takes to run it.
An alert when a lane has held above the store's threshold, from the camera already over the tills.
// 01 · The question
“Does a till need opening in the next five minutes?”
// 02 · Why it is hard
A queue is a social structure, not a visual one. People stand in loose clusters, step out and return, and the boundary between queueing and browsing is a judgement the labelling standard has to fix.
// 03 · What the labels have to be
Bounding box + tracking
Decided before the first frame
Queue length is a count inside a region, and wait time needs the same person followed from joining to serving. Tracking is what turns a headcount into the number a manager can act on.
This is the decision that is expensive to reverse. The geometry has to match what the answer contains, and finding out it does not means labelling the set a second time.
// 03a · What it finds
The people in each lane
a box on every person inside the polygon drawn over that lane on the camera above the tills, read on every sampled frame, about every two seconds
The same shopper, frame after frame
one identity kept on each person from joining the back of the queue to reaching the belt, which is what turns a headcount into a wait
The shopper who is not queueing
the person reading the magazine rack beside the lane and the colleague behind the till are in the camera's view and outside the count, because the polygon and the labeling standard for this store fix where queueing starts
The queue that has left the lane
on a Saturday the line runs back past where it sits on a Tuesday, so the zone is drawn to the far end of where the queue goes, not to the end of the belt
// 03b · What it checks
- 01
The count is above the lane's threshold, and has stayed there for the hold the store set. Two shoppers crossing the zone on the way to the exit do not open a till.
- 02
The lane is one the rule covers right now. The rule names the camera, the zone and the label in one sentence and carries a schedule, so the lanes that close at eight stop paging at eight.
- 03
The same queue does not page the same person twice in a row. A cooldown sits on every rule, and a severity, so the lane by the exit can be told from the one at the service desk.
- 04
The rule was approved before it went live. It is written, run against footage, read by a person and only then switched on. Nobody is paged by a rule nobody has seen fire.
- 05
A frame the model is unsure of is not a count. It comes back to a person with the box and the zone drawn on it, the correction is recorded against the frame, and the model is retrained on the hardest frames of the week.
// 03c · What the store provides
- A camera that sees the lanes. The one already over the tills, from above or from the side, at any angle that keeps shoppers apart. Any RTSP feed or recorder; no new camera and no new hardware at the checkout.
- Frames from this store, not a stock dataset. The model learns your lanes, your floor and the light through your front windows, and the polygon is drawn on your camera once.
- The target, in the store's words. How many waiting, or for how long, before a till should open is the store's rule, written as a sentence, with who hears it and how long before it may fire again.
- Someone who can open a till. The alert reaches the duty manager on Slack, email or a webhook into whatever the store already carries, with the frame attached, so the call is made from the picture and not from the number alone.
// 04 · How we run it
Three parts of one loop, on this job.
LexAnnotate
Shopper Tracking
Annotate in-store video for customer detection, employee identification, door events, and product placement. Distinguish between customers, staff (by uniform), and children within each camera's region of interest.
LexInsight
Conversion Analytics
Turn raw foot traffic data into conversion intelligence. Link visual detections to store zones, time slots, and operational events. Surface patterns that static dashboards miss.
LexAlert
Store Alerts
Monitor detection accuracy across store locations. Detect when camera repositioning, seasonal displays, or lighting changes degrade model performance. Route the failures back for retraining.
// 05 · What breaks it after launch
Store layouts are not standardised. Till spacing, camera height and how a queue physically forms differ between formats, so a model tuned on one estate under-reads on another.
It is not the only one that can get this use case, it is the one that usually gets it first. All ten conditions.
// 06 · In retail
Store Intelligence
The Challenge
The models lacked high-quality labeled data, activity alerts came back false, and unmonitored registers left customers waiting at checkout.
Our Execution
Labeled customer and staff positions. Validated alerts and traffic counts for accurate footfall tracking.
The Result
Faster AI training with accurate labeling. Smarter alerts for open doors and empty registers.
4M+
Annotations & validations
// 06a · Questions we get asked
Do we need a new camera over the tills?
No. The ceiling camera that already records the checkout area is the one we use, over RTSP or from the recorder. What it has to see is the whole run of the queue on a busy day, including where it spills out of the lane. If the camera sees only the till, we say so before anything is built.
How does it tell a queue from people walking past?
Three ways. The polygon is drawn over where the queue forms and not over the walkway beside it. A person is counted only once their box has stayed inside the zone for the hold, so a shopper crossing to the exit is never a count. And where queueing starts is written into the labeling standard for this store and learned from your frames, checked by a person, rather than a default we ship.
What does the manager receive?
A message when a lane's count has stayed above the threshold for the window the store set: which camera, which lane, the count, and the frame it was read from, on Slack, email or a webhook. A cooldown keeps the same queue from paging the same person every two seconds, and a severity tells a queue at the service desk from one at the far tills.
Does the count hold when we open the next store?
Not by itself, and this is the condition that reaches queue counting first. Till spacing, camera height and how a queue forms differ between formats, so a model tuned on one estate under-reads on another. The first frames from the new store come back to a person, the corrections retrain the model, and the new version rolls out with no downtime. The store that was already counting keeps the version it was counting with.