People Counting Accuracy Tests and Calibration

A lit indicator, numbers on a screen or a sensor connected to the network does not prove that a people counter is counting accurately. It may record someone in the wrong direction, merge two people walking side by side, or incorrectly count someone who waits at the entrance and turns back. Even if the sensor counts correctly, the reporting platform may still have missing data.

Checks should therefore address two questions: Are physical crossings measured correctly? and are the measured values reported correctly? Understanding the differences between 3D sensors and camera-based people counting helps set installation expectations; a technology label alone is no guarantee of accuracy.

Key factors affecting counting accuracy

Installation geometry: Height, tilt, field of view and entrance width may fall outside the device’s supported operating range.
Counting line: A line crossing an area where people wait or turn back may create incorrect records; entry and exit directions may also be reversed.
Obstructions: Signs, lighting, security antennas or temporary decorations may block the monitored area.
Crowd behaviour: Side-by-side crossings, pushchairs and sudden changes of direction should be tested separately.
Device and reporting settings: Time zones, store opening hours, device-to-entrance mappings and data processing rules affect the results.

Calibration and validation are different processes

Calibration or configuration adjustment means adapting counting lines, sensitivity or appropriate product parameters to site conditions. Validation compares system results with an independent reference using a defined test method and documents performance. Adjusting settings does not validate a system: it must be checked again afterwards.

How do you run a practical accuracy test?

Record the test date, entrance, device ID, software version and counting line.
Use manual reference counts to monitor entries and exits separately during both quiet and busy periods.
Compare the sensor’s entry and exit totals with manual records for the same period.
Do not conceal counting errors within a daily net total; evaluate them by direction and challenging scene type.
Adjust settings if necessary, then retest using the same method.
Finally, compare the device’s own counts with the corresponding hours in the central report.

For example, suppose the manual reference records 200 entries and the device records 192. The absolute difference is 8, giving an absolute count error rate of 4% relative to the reference. However, undercounting entries and overcounting exits can offset each other in a daily total. Entries and exits should therefore be tested separately.

How should acceptance criteria be chosen?

Acceptance criteria should be agreed in advance, taking account of the device model, entrance type, intended data use, traffic distribution and contract. A manufacturer’s accuracy figure for laboratory or suitable field conditions cannot be treated as a guarantee for every installation. For example, the official Axis user guide recommends a manual reference crossing test for a particular AXIS People Counter application; that procedure is not a universal threshold for every brand and model.

The intended use of the measurement also matters. A report tracking general traffic trends may have a different error tolerance from data underpinning commercial KPIs such as store conversion rates.

Do not overlook reporting after the test

A successful manual counting test does not prove that all data reaches the platform. Network interruptions, incorrect time mappings, late records, empty transfers and duplicate transfers require separate checks. Acceptance documentation should therefore include both sensor-to-reference count comparisons and sensor-to-platform consistency.

The CS Otomasyon approach to people counting systems considers installation, accuracy checks and reporting needs together. Before replacing an existing device, a field test plan helps establish the source and extent of the problem.

Frequently asked questions

How often should calibration be performed?

There is no universal schedule. Checks are needed after initial installation, device repositioning, changes to entrance decorations or unusual deviations. Periodic intervals depend on how critical the use is and the risks at the site.

If the device is online, is the report accurate?

No. Being online may indicate connectivity; it does not prove counting accuracy, the arrival of every hour’s data or that a record has not been processed twice.

Conclusion

People counting accuracy is as much a field validation process as a device characteristic. CS Otomasyon can help assess installation, counting lines, manual references and reporting layers separately to identify correction priorities for your existing system.