People Counting: Visitor Analytics at 96% Accuracy for Retail and Malls

Camera-based people counting: 96.2% counting accuracy, direction-aware counts, staff filtering, conversion rate from POS, mall occupancy and KVKK-compliant setup.

Translated from the Turkish original · Türkçe aslı

A camera-based people counting system detects each visitor separately in the image from an overhead camera above the door and tracks them along with their direction; it counts entries and exits separately, filters out staff and courier traffic, and calculates the conversion rate by matching net visitors with the POS receipt count. Field-verified counting accuracy is 96.2%; data is processed anonymously and no facial recognition is performed.

Why is people counting the most mismeasured metric in retail?

The most basic question in retail is still the most poorly measured: how many people entered the store today? Revenue alone does not explain performance; "we sold little" and "few people came in" are different problems. The first points to the sales process, staffing or pricing; the second to the shop window, the location or the reach of the campaign. If the count is wrong, every metric derived from it, from conversion rate to revenue per visitor, carries the same error.

Infrared beam counters only produce the information that the beam was broken, in other words "something passed": they get confused by double passes, count a group entering side by side as one, and are limited in telling direction. In the EscherSense comparison table, the typical field accuracy of these counters is in the 80–90% band. That band is enough to follow rough trends, but when comparing two branches or two campaign weeks, the real difference can be lost within the counter's error.

How does people counting with cameras work?

The people counting flow consists of three steps. The first step is counting with direction: footage is analysed at 25 frames per second, each visitor is tracked as a separate object, and depending on the direction in which they cross a virtual line drawn slightly inside the door line, they are recorded as an entry or an exit. Because counting is based on the completed track rather than a single frame, a person who hesitates at the door is not counted again, and each member of a group entering side by side is counted separately. In the second step, staff and delivery traffic is separated out to arrive at net visitors. In the third step, conversion is calculated by matching the receipt count with the visitor count.

One limit should be clearly understood: because the system does not identify people, a customer who leaves the store and returns a while later is counted as a new visit. What is measured is visits, not unique individuals; this definition should be used consistently in reports. In the default setup, footage is processed on an edge server inside the facility and raw video is not sent to the cloud (processing can also be done in the cloud if preferred); the output is not people but numbers and hourly distributions.

How is counting accuracy measured, and what does 96.2% mean?

In the retail case on our References page, counting accuracy was verified at 96.2%; the band expected on site after calibration is 95–97%. Accuracy is a field value, not a laboratory one: during the pilot, real passes in certain time slots are labelled manually and compared with the system's count. A meaningful measurement should include not only quiet hours but also peak hours, group entries, evening light and rainy days.

The direction of the error matters as much as the overall percentage. If the system consistently undercounts, conversion looks higher than it is; if it overcounts, lower. That is why entries and exits are verified separately. A practical consistency test is the end-of-day balance: after closing, entries and exits are expected to be close to each other; if the difference is large, a door may be insufficiently covered or the counting line may be positioned incorrectly.

How is the conversion rate calculated from the visitor count?

The conversion rate is found by dividing the number of receipts in a time slot by the number of net visitors in the same slot. Receipt data comes from the POS and visitor data from the people counting system, matched by hour; the conversion measurement accuracy published on our solution page is 92.8%. Revenue growth sometimes comes from more visitors and sometimes from better selling; the metric that tells the two apart is conversion. Because a group shopping together leaves with a single receipt, it is healthier to track each store's own trend rather than the absolute rates of stores in different formats.

When the hourly traffic curve is overlaid on staff shifts, you can see at what hour queues build up at the checkout and sales are lost. In the sample dashboard on our solution page, weekday hourly conversion is 31% in the 10:00–12:00 slot but falls to 16% in the 17:00–20:00 slot: the heaviest traffic coincides with the lowest conversion, and the first action is to shift checkout and information desk staff to these hours. In the same example, weekly conversion rose by +4.2 points without any campaign change, through staff planning and queue management alone. These values are sample dashboard data; the impact in your store is measured in the pilot.

People counting in store chains, malls and events

In chain structures, traffic from all stores is gathered on a single dashboard; stores are compared not only by revenue but by revenue per visitor and conversion. For the comparison to be fair, the staff filter, door definition and opening hours are kept the same at every branch; otherwise the difference comes from measurement differences rather than store performance. To read the effect of a campaign, the comparison is made with the same days and time slots of previous weeks; comparing weekend traffic with weekdays means attributing a calendar effect to the campaign. People counting is the core layer of the EscherSense platform, which produces 16 metrics from the same camera stream.

In a mall, the question moves from the store door to floors and corridors: area-based traffic distribution shows which floor and which corridor is being fed, and tenant negotiations use measurement instead of estimates. In events and venues with limited capacity, the real need is live occupancy; because entries and exits are counted separately, the number of people inside is calculated continuously and an alert is generated when a defined threshold is exceeded. To see where movement concentrates inside the store, counting is used together with heat map analysis.

Which camera is needed, and are existing cameras sufficient?

What matters is not the camera's brand but its angle. The ideal position is an overhead camera above the door, and a height of 2.6–4 m gives good results; a top-down view reduces people occluding each other in crowds and removes the need for face images. Counting can still be done with security cameras looking at the door from the side, but occlusion increases at peak hours; we recommend positions during the site survey. Wide mall entrances may require more than one camera. Automatic door panels, sunlight reflecting off shiny floors and promotional stands placed in front of the door are checked during the site survey. In EscherSense installations, four to eight cameras are enough for a typical store; for angle and height, see our camera placement guide.

Video analytics or counter sensors?

A camera is not the right choice in every case. If the only need is rough traffic through one door, a simple sensor may be enough; in areas where processing images is not wanted under company policy, thermal or radar comes to the fore. In stores with heavy group entries, where staff filtering is needed, or where conversion and queue data are also required, video analytics answers more questions with a single infrastructure. For a detailed comparison, see the people counting: sensor or camera article.

Common mistakes on site

• Not covering all entrances: when the car park connection or a side door is missed, traffic looks low and the occupancy calculation drifts.
• Placing the counting line outside the door threshold: pedestrians who stop to look at the shop window get mixed into the count.
• Not defining rules for staff, deliveries, children and pushchairs before installation; branches end up measured with different rules.
• Reading the new system as a continuation of the old counter's series: because group entries are now counted separately, traffic appears to have "increased"; what changed is the measurement, not the visitors.
• Reporting data only as monthly averages; staffing and checkout decisions are made by the hour.

Using it together with demographics

Counting answers the question "how many people"; demographic analysis anonymously measures the age band and gender distribution of the visitor audience. No identity records are kept and no face images are stored; hourly and daily distributions are reported, not individual data. Because demographics requires an angle close to face height, it usually needs a camera position separate from the door count.

KVKK and installation

People counting does not use facial recognition or biometric records; analytics outputs are anonymous and event clips are automatically deleted at the end of the defined retention period. Even so, because processing is done with cameras, updating the privacy notice under KVKK (Türkiye's Personal Data Protection Law) is recommended; for the framework, see our KVKK and video analytics article. This is not legal advice. If the same cameras are also used for other purposes such as loss prevention, each purpose should be defined separately; we covered this in the retail loss prevention article.

Installation starts with the existing cameras: doors and angles are inspected on site, counting lines and the staff filter are defined, and the POS connection is set up; data is transferred to POS, ERP, CRM and BI tools via REST API. Installation is completed within a week, followed by measurable results in a 30-day pilot. The success criterion is written down before the pilot; the designing the right pilot article offers a good framework for the measurement method. The typical payback period in retail analytics is 6–14 months depending on store revenue; the decision to expand scope is made on pilot data, not estimates.

Frequently asked questions

Are our existing security cameras sufficient for people counting?

In most cases yes, but position is decisive. The ideal camera is an overhead camera mounted above the door at a height of 2.6–4 m; a top-down view reduces people occluding each other in crowds. With cameras looking at the door from the side, accuracy may drop at peak hours. Before installation we assess your camera inventory free of charge and report suitable points and, if necessary, cameras to be added.

How accurate is counting in the field?

In the retail case on our References page, counting accuracy was verified at 96.2%; the typical band after calibration is 95–97%. Accuracy is a field value: during the pilot, real passes are labelled manually and compared with the system's count. Because door width, camera angle and crowd density affect the result, thresholds are calibrated on your site.

Are staff and couriers counted too?

No, they are filtered out. In the second step of counting, staff and delivery traffic is filtered to arrive at net visitors; thanks to direction information, a person waiting at the door is not counted again either. Which marker is used is determined per store during the site survey. If staff passes are added to the denominator, conversion looks lower than it is and branch comparisons break down.

How is the store conversion rate calculated?

The conversion rate is found by dividing the number of receipts in a time slot by the number of net visitors in the same slot. Receipt data is taken from the POS and visitor data from the people counting system, matched by hour; our published conversion measurement accuracy is 92.8%. Because customers arriving as a group generate a single receipt, it is healthier to track the store's own trend.

How does people counting connect to our POS and reporting systems?

It integrates with POS, ERP, CRM and BI tools via REST API; events can also be published over MQTT. Conversion is calculated by matching the POS receipt count with the visitor count by hour. Reports are produced as hourly and daily traffic, conversion and branch comparisons; in chain structures, all branches are monitored on a single dashboard.

How long does a people counting pilot take?

With existing cameras, installation is completed within a week, followed by a 30-day pilot. During the pilot, the system's count is compared with manually labelled passes, the end-of-day entry–exit balance is checked and the POS matching is verified. Writing down the success criterion before the pilot means the result can be used in the decision to roll out across the chain.

Does people counting cause KVKK issues?

The system does not use facial recognition or biometric records; in the default setup, footage is processed on an edge server inside the facility, no raw video is sent to the cloud (processing can also be done in the cloud if preferred) and the outputs are anonymous numbers. Even so, updating the privacy notice to reflect the purpose of camera-based processing is recommended. This is general information; the final assessment is made by your legal department.