What is store camera analytics? Footfall counting, conversion rate and checkout queues

What is store camera analytics, and which numbers can existing cameras produce?

Camera analytics means processing footage from in-store cameras with computer vision so that it produces numbers and event records instead of images. Four measurements matter to a grocer: visitors entering the store, density in areas inside the store, the queue in front of the tills and events at the till. This guide focuses on the first three. How footage is turned into events, where AI gets it wrong and how a till event is matched with the POS record are covered in the guide to AI camera analytics.

No measurement produces a decision on its own. The value of the visitor count appears when it is divided into the receipt count for the same hourly slot: the conversion rate. The value of the queue appears when it is read together with the number of tills open at the same moment: is till capacity keeping up with demand. The value of density appears when it is compared with the area's sales. That is why camera data is always kept at the same store and hourly grain as sales data; a daily total hides the hour where the problem is.

Existing security cameras can often be used for this, but not every camera suits every measurement. A camera that sees the entrance together with the street at a wide angle may be good enough for rewinding to an incident, yet not good enough for counting people one by one in a crowd. Which camera is assigned to which measurement is chosen with the rules in the second section. The general framework linking footfall, conversion and basket size is in the store sales guide; here the focus is how the camera produces these numbers and how far they can be trusted.

Conversion rate

Receipts ÷ Visitors × 100Numerator and denominator come from the same store and the same hourly slot. Two sources where one is hourly and the other daily are not divided into each other.

Area density

People in the zone ÷ Zone area (m²)When following the same zone over time the head count is enough; when comparing zones of different sizes, divide by square metres.

Hourly share of footfall

Visitors in the slot ÷ Daily visitors × 100

How are counting lines and zones defined, and which camera is used?

Computer vision detects a person as a box and follows it from frame to frame. Counting is done by looking at which direction this track crosses a virtual line drawn on the image: crossing inwards counts as an entry, crossing outwards as an exit. A zone is an area drawn on the image. How many people are in the zone at a given moment gives density; the number of people in the waiting zone in front of the till, and the time spent there, give the queue.

Where the line sits is half of the accuracy. On a line drawn right on the door threshold, a person who pauses in front of the automatic door, turns back or stands on the line can be counted more than once. The line is drawn a few steps inside the door, perpendicular to the flow and where nobody stops; the trolley and basket area stays outside it. If entry and exit use the same door, one line separates the two directions; if there are separate doors, each door is defined separately and the total is the store's entries.

Camera choice follows the same logic. People occluding each other is one of the main difficulties of video tracking named in the literature: a camera that sees two people walking side by side from the side may take them for one. A camera looking down from high up, close to vertical, reduces this overlap. The same angle is also the opposite of the face-focused recording the KVKK announcement asks businesses to avoid: for counting, what is visible is the head and shoulders, not the face.

• Entrance camera: inside the door, high up and close to vertical; it sees the whole doorway, not the street beyond it
• Light: daylight through the door creates backlight; evening reflections on the shop window and night lighting are checked separately
• Checkout line: each till's waiting lane is a separate zone; the till belt and the payment point stay outside the zone
• Aisles and walkways: a wide angle is enough for density; zones are named as in the store layout plan
• Clock: the camera recorder's clock is synchronised with the POS clock from the same source; a clock offset breaks hourly matching

1. Write down the question to measureConversion, queue or area density. Each question determines which camera and which line or zone it needs.
2. Take a camera inventoryEach camera's position, angle, resolution and field of view are listed with a screenshot. For a camera unsuitable for counting, a change of position or angle is noted.
3. Draw the line and zones on the imageCounting lines, till waiting zones and aisle zones are defined with the same naming in every store and documented with a screenshot.
4. Keep the rule the same in every storeThe rule chosen for staff entries, children and groups is applied the same way in every store. The conversion rates of two stores counted under different rules cannot be compared.
5. Redefine when the camera changesWhen a camera's position or angle, or the store layout, changes, lines and zones are redrawn and accuracy is tested again.

How staff and group entries are handled in counting is covered in the store sales guide. Using a method that recognises people, such as facial recognition, to remove staff from the count means processing biometric data; ways that do not recognise people, such as a separate staff door or separately flagged opening and closing hours, are preferred.

How accurate is camera counting, and how is it tested?

No counting method is error-free; the problem is not the error itself but not knowing how large it is and in which direction. The most common error in camera counting is undercounting in crowds: people entering side by side look like one. A rarer but more misleading error is overcounting: someone who pauses on the line or turns back appears to have crossed twice.

Accuracy is tested with a manual count sample. One quiet and one busy hour are chosen; over the same period one person counts entries at the door with a hand tally, or the recording for the same time window is reviewed and counted. The difference between the camera count and the manual count is written down as a percentage. Quiet and busy hours are measured separately, because the error usually grows with the crowd and a single average error hides that difference.

The second check is the entry-exit balance. By the end of the day everyone who entered the store has left; entries and exits should be close to each other. If the gap grows day by day or is always in the same direction, the line position, the door's swing area or the lighting may have changed. This check needs no manual count and can run automatically every day.

There is no universal limit for acceptable error; the chain writes its own tolerance. What matters more is that the error is consistent from store to store and from hour to hour. A count with a stable error shows trends and differences between stores correctly; a count whose error varies by store makes two stores' conversion rates incomparable.

Counting error

(Camera count − Manual count) ÷ Manual count × 100A negative value means undercounting, a positive value overcounting.

Entry-exit balance

(Daily entries − Daily exits) ÷ Daily entries × 100

Corrected visitors

Camera count ÷ (1 + Counting error rate)The error rate is written as a decimal; a 6.1 per cent undercount gives 1 + (−0.061) = 0.939.

Reviewing recordings for validation is also processing of camera data. Review is limited to the counting purpose, only the selected time window is looked at, and the task is given to an authorised person.

How is the checkout queue measured and linked to the number of open tills?

The queue is measured in two ways. Queue length is the number of people in the waiting zone in front of the till at a given moment. Queue time is the time from when a customer enters the waiting zone until they reach the till. The two are linked by the basic relationship of queueing theory: the average number of customers in the queue equals the arrival rate at the tills multiplied by the average waiting time. If the camera counts length and the POS shows how many receipts reach the tills per hour, average waiting time can be estimated from this relationship.

The camera counts people; the till counts receipts. A couple shopping together are two people in the queue and one receipt at the till. So when queue length is converted to receipts, either a group rule is applied or the threshold is defined directly in people; the two units are not mixed in one calculation.

The measure linking the queue to the number of open tills is till utilisation: receipts arriving per hour divided by the receipts the open tills can process per hour. In the simplest queueing model, the average number of customers waiting in the queue is utilisation squared divided by one minus utilisation. Because the denominator is one minus utilisation, the queue grows disproportionately as utilisation approaches 100 per cent: going from 80 to 90 per cent more than two and a half times the queue. Real tills behave differently from this model, but the shape of the curve is the same: at a busy hour a small rise in traffic produces a big queue.

How customers react to the queue has also been measured with cameras. A study at the deli counter of a supermarket, in which the queue was recorded at regular intervals with video recognition and combined with sales records, found that the effect of waiting on purchasing is non-linear and that customers look more at how long the queue appears than at how fast it moves. The same study notes that pooling several queues into a single queue can lengthen the queue customers see and lower revenue. That is why the till threshold is defined not only in time but also in the length the customer sees.

Average customers in the queue

Arrival rate at the tills (receipts/min) × Average waiting time (min)

Hourly capacity per till

60 ÷ Average transaction time (min)Transaction time runs from scanning the first item to the end of payment; it is read from the POS record.

Till utilisation

Receipts per hour ÷ (Open tills × Hourly capacity per till) × 100

Average customers in the queue (single till, simple model)

Utilisation² ÷ (1 − Utilisation)The simplest queueing model, assuming random arrivals and random service times; it is for understanding the shape of the curve, not for precise prediction.

1. Write down the thresholdIn minutes for queue time and in people per till for queue length; they are two separate thresholds. The value is the chain's service target, not an outside standard.
2. Decide who receives the alertA threshold breach goes to the shift supervisor; the person who will open the till and the list of backup cashiers are known in advance.
3. Tie the action to a recordAfter the alert, how many minutes it took to open the till and when the queue fell back below the threshold are recorded; otherwise the number of alerts rises and nothing changes.
4. Move a recurring pattern into the rotaA threshold breach that repeats on the same day and hour is a shift planning problem, not a matter for on-the-spot response; the cashier rota is rebuilt around that hour.

Worked example: an evening conversion drop, the queue threshold and a fourth till

Below are the camera and POS data for one weekday in a five-till neighbourhood supermarket; the figures are illustrative. Average transaction time is read from the POS record as 1.5 minutes, so one till can process 40 receipts an hour. The queue time threshold is 4 minutes.

The table points at a single slot. Between 18:00 and 20:00 footfall is high, but conversion is about 10 points lower than the rest of the day; in the same slot till utilisation is 96.7 per cent and average queue time is above the threshold. The number of open tills does not change between the afternoon and the evening: three tills are open in both slots. At 18:40 queue time rises to 6 minutes and the threshold alert reaches the shift supervisor.

There are two traps in estimating lost sales. The first is the nature of evening shopping: customers on their way home from work may browse more and buy less, so not all of the conversion drop is attributed to the queue. The comparison base is therefore not the other hours of the day but similar weekday evenings in the same store and the same slot when four tills were open. The second is that some customers who give up on the queue come back to the same store another day; that sale is not lost but deferred. In the example the deferred share is conservatively taken as half.

The cashier cost per hour comes from the business's own payroll. If the extra till is opened not with a new person but by moving an employee from shelf replenishment at that hour, the cost is the postponed shelf work, and that goes into the calculation too.

How is camera data read together with sales, and what does it not show?

Camera data tells you where and when something changed, not why. An evening conversion drop may come from the queue, from a product missing from the shelf, from price or from the customer mix. The cause is found by reading the same slot's sales, stock and till data together: if conversion falls without the queue threshold being breached, the problem is not at the tills; if particular categories' sales fall along with conversion, look at the shelf side.

The second rule is the direction of causality. At busy hours both sales and the queue rise; in raw data the two climb together and the queue can look as if it is increasing sales. The practical way to separate the effect is comparison: same store, same slot, a different number of tills; or similar stores where the same change was and was not made. That is why the comparison base in the worked example was chosen as it was.

The third rule is the unit. If conversion is to be compared across stores, the counting rule, the hourly grain and the validation method must be the same. If the rules differ, only each store's own trend is compared, not their conversion levels. In a store where a camera has been replaced or rearranged, the date of the change is flagged in the report; otherwise a counting difference is mistaken for a change in performance.

Pre-rollout checklist: cameras, clocks and data protection

Even when camera analytics works technically, three gaps make it unusable: a camera unsuitable for counting, a clock not synchronised with the POS, and a privacy notice that has not been updated. Retention periods, warning signs, audio recording and access rights are covered in detail in the store camera recording and data protection guide; here are only the points that change when moving to analytics.

Using a camera installed for security to count visitors or measure queues is a new processing purpose. Under Article 5 of the Communiqué on the obligation to inform, when the purpose of processing personal data changes, the obligation to inform is fulfilled separately for that purpose before the processing activity starts. Article 10 of Law No. 6698 lists what the notice must contain: the identity of the data controller, the purpose of processing, to whom and for what purpose the data may be transferred, the method of collection and legal basis, and the rights of the data subject.

The announcement of 8 June 2026 by the Turkish Personal Data Protection Authority (KVKK) gives direct guidance on two points. Using cameras to see whether employees work efficiently or to increase discipline is not regarded as a legitimate purpose, and cameras should not be used outside the stated purpose to monitor employees' attendance or performance; queue and density output is therefore read at area and hour level, not at cashier level. The announcement also calls for data minimisation and asks that no wide-angle or face-focused recording covering every area of the workplace be made.

1. Write down the purpose and the metricsFootfall counting, queue measurement, area density; for each, which camera will be used and at what level (store, area, hour) the output will be kept are written down.
2. Choose cameras and check anglesEntrance, checkout line and main walkway cameras; check for a near-vertical angle, backlight and night reflections. A change is decided for unsuitable cameras.
3. Synchronise the clocksThe camera recorder and the POS are connected to the same time source; hourly conversion and till matching rely on this.
4. Document lines and zonesEach store's counting line, till waiting zones and aisle zones are archived with screenshots.
5. Test accuracyA manual count sample at a quiet and a busy hour, plus the entry-exit balance. The margin of error is stated in a footnote to the report.
6. Update the privacy noticeThe new purpose is added to the privacy notice, and to the sign if needed, before processing starts; employees are informed separately.
7. Define thresholds and ownersThe queue and density thresholds, who receives the alert and the expected action are written down.

This section is general information, not legal advice. The purpose of the cameras, the legal basis, the privacy notice and the retention period are the business's decisions as data controller and are assessed with a lawyer who works in data protection law.

How does camera analytics work in Mindpera?

Mind Cam processes footage from cameras in grocery stores with computer vision and produces event and visitor count data. Visitors entering the store are counted by store and by hour; because visitors and sales sit in the same hourly slot, the conversion rate is measured rather than estimated. The density map shows how much each area inside the store is used. An alert is raised when queue time or area density exceeds its threshold; every event is recorded with store, area, event type and time.

On the Mind Ops side, the same camera data comes together with POS and ERP records in one data model. Sales, conversion and basket size are calculated with defined formulas and appear on the same page of the store daily scorecard; an item opened on a threshold breach goes to its owner's task list.

The purpose of the cameras, the privacy notice, the retention period and access to recordings are the business's decisions as data controller; an analytics product does not remove those obligations. If camera, till and sales data sit in different systems, our team handles the installation and brings the sources into a single data model.

• Visitor counting: entries by store and by hour
• Density map: how much each area inside the store is used
• Threshold alert: when queue time or area density exceeds its threshold
• Mind Ops store scorecard: sales, conversion and basket on the same page