Computer vision · Anonymous counting
We count crossings,not identities.
Computer vision counting uses cameras and AI to find out how many people entered a building, how many vehicles passed through the gate or how many pallets were unloaded. The system counts anonymously, following clear line or zone rules, and differences from what was expected go to a person for review.
Manual counting is slow, and totals arrive too late
When numbers are collected by hand or from estimates, decisions about staff, space or stock are made on incomplete data. And when people are being counted, any solution must also answer the question: how do people stay anonymous?
- Differences between the quantity delivered and the one on the delivery note are only discovered at stocktaking
- The occupancy of a space is estimated by eye, not measured
- Simple sensor counters cannot tell a person from a trolley or a pallet from a box
- Solutions that film people rightly raise privacy questions
Who it is for
- Managers of buildings, event venues or leisure centres who want to know occupancy levels
- Warehouses and distribution centres that compare quantities received with the accompanying documents
- Production units that want to count the parts or packages passing a point
- Organisations that need aggregated statistics on flows, with no data about people
What the system can count: crossings, occupancy, batches
- 01
Crossings over a virtual line
Counts entries and exits through a point, separately by direction, for people, vehicles or objects.
- 02
Zone occupancy
Estimates how many people or objects are in a zone drawn on the image at a given moment.
- 03
Batch totals
Counts the pallets, parcels or parts in an unloading and compares the total with the expected value.
- 04
Flagging differences
When the count does not match the document or exceeds a threshold, the case goes for review.
- 05
Statistics by interval
Counts are aggregated by hour, day or shift, without the need to keep images of people.
- 06
Anonymisation by design
People are treated only as shapes being counted: faces are not analysed, and images can be deleted after processing.
How it works
From frame to verified total
The steps marked in yellow are approved by a person.
- Input
- AI
- Action
- Human in the loop
- Result
- 01
Input
Camera stream
A camera mounted above the crossing point sends the image.
- 02
AI
Detection without identification
The model tracks each person or object only while it crosses the frame.
- 03
Action
Counting rules
Line crossings or presence in the zone are added up in counters.
- 04
Action
Comparison with expectations
The total is compared with the delivery note, the capacity or the set threshold.
- 05
Human in the loop
Review of differences
A responsible person analyses the discrepancies and decides whether anything needs correcting.
- 06
Result
Aggregated report
The result is totals by interval, with no data that identifies people.
What data and camera position are needed
- Video footage from the counting point, at peak and quiet hours
- Labelled examples of the objects counted, especially where they touch or overlap
- An exact definition of the counting line or zone and of the rules: what counts as a crossing and what is ignored
- Reference documents for comparison, for example delivery notes or the maximum capacity of the space
- A suitably positioned camera, ideally looking straight down for people, to reduce overlaps
Possible integrations
- IP cameras with an RTSP stream
- ONVIF-compatible cameras
- NVRs and VMS systems
- WMS systems for goods receiving
- ERPs such as SAP Business One or Microsoft Dynamics 365
- Power BI or Grafana dashboards
- Exports to Google Sheets or Microsoft Excel
- REST API and webhooks
- Email or Microsoft Teams notifications
- IP cameras with an RTSP stream
- ONVIF-compatible cameras
- NVRs and VMS systems
- WMS systems for goods receiving
- ERPs such as SAP Business One or Microsoft Dynamics 365
- Power BI or Grafana dashboards
- Exports to Google Sheets or Microsoft Excel
- REST API and webhooks
- Email or Microsoft Teams notifications
Each integration's compatibility is checked during the analysis phase, for the systems and versions you use.
Control and transparency
What stays with people
- The decision to correct stock or a document based on a counting difference
- Setting the lines, zones and thresholds and reviewing them when the layout changes
- Regular sample-based checks of automatic counts against a manual count
- Confirming that counting people has a legitimate purpose, is proportionate and is communicated to those filmed
Limits and operating conditions
- Counting works best with a camera mounted overhead; in dense crowds, overlapping people or objects may be miscounted
- Stacked or covered objects, for example wrapped in film, are harder to separate from one another
- Side angles and poor lighting increase errors; sometimes a camera needs to be moved or added
- Anonymous counting cannot tell whether the same person entered several times and is not designed to do so
How we measure counting accuracy
We agree the indicators together before launch and track them afterwards. The values depend on the process, the data and the volume.
- Counting error per interval
- The difference between the automatic total and a manual count on the same video footage, for intervals chosen by the team
- Accuracy by direction
- Entries and exits checked separately, to see whether errors occur mainly in one direction
- Behaviour at peak times
- Results compared between busy and quiet hours, at the same counting point
- Confirmed differences
- How many of the discrepancies flagged against documents the responsible person confirmed as real differences
How we implement anonymous counting
- 01
We choose the counting points
We agree where counting happens, what exactly is counted and how the totals will be used.
- 02
We check the cameras and angle
We analyse real images and recommend, if needed, repositioning or an overhead camera.
- 03
We calibrate on real data
We compare counts with manual counts and adjust the lines, zones and model.
- 04
We connect the reporting
Totals reach the chosen dashboard or system, and differences enter the review workflow.
Use cases
The scenarios below are conceptual examples, designed to illustrate realistic applications. They are not implementations delivered by WEB-AZ.
- Logistics
Pallets counted at unloading
At the dock, the camera counts the pallets taken off the lorry, and the total is compared with the delivery note. Differences reach the shift supervisor before the receipt is signed.
Conceptual example - Events
Occupancy of a hall
Cameras at the entrances count entries and exits anonymously, and the organiser sees how close the hall is to its set capacity.
Conceptual example - Food industry
Multipacks counted at the end of the line
Before palletising, the system counts the multipacks passing on the belt and flags when the pace drops below the rate set by the team.
Conceptual example
Useful guides
Before you decide, read
Frequently asked questions
What is people and object counting with AI?
It is a computer vision application that counts, from a camera's images, crossings over a line, how many people or objects are in a zone or how many pallets are in a batch. The result is a total, compared with the expected value, without identifying people. For queues and empty shelves in stores, retail video analytics is the right fit.
Does the system recognise who came in?
No. People are detected only as shapes crossing the frame and are counted. Faces are not analysed and no profiles are built. Facial recognition is not part of this solution.
Is video footage kept?
Not necessarily. The system can process the stream and keep only the totals. If images need to be kept for review, we agree together on the retention period, access and how people are informed, in line with your video surveillance policy and applicable law.
What happens when the system is not sure of a count?
Ambiguous situations, such as tight groups or overlapping objects, can be marked. When the total differs from the expected value, the case goes to a person, who can review the footage and decide.
Can I use an existing security camera?
Sometimes. Cameras mounted at the side or very far from the crossing point often give poor counting results. We analyse real images before recommending whether to use them or add a camera.
How accurate is the counting?
It depends on the angle, density and lighting, so we do not start from a promised figure. We measure the error during the pilot, comparing against manual counts on the same footage, and tell you openly what to expect at your counting point.
Can’t find the answer? Write to us and we’ll reply directly.
Want to know how many, without finding out who?
Tell us what needs counting and where. We review the camera angles, filming conditions and how the results can stay anonymous.
Related solutions
- Video analytics for retailAggregated visitor flows, queue length and empty shelves from in-store cameras, without identifying customers.
- Object detectionModels that recognise the objects that matter to your process in images and pass them on as structured data.
- Operational monitoringVisual tracking of process stages and of vehicle and pallet movement in yards and warehouses, so you can see where the flow gets stuck.