How accurate is camera-based people counting?
Published figures for camera-based and sensor-based counting sit in the high nineties, with documented conditions that pull real-world performance below the headline number. The Frictionless retail report cites systems at 99% accuracy while noting that "low confidence events" still occur, caused by factors ranging from a blocked camera lens to the wrong product location 1. Beyond Retail in the Age of AI describes a system that forgoes biometric recognition and instead combines behavioural and sensor data, reporting an accuracy rate of 99.7% 2.
What moves the accuracy figure
The DOOH and In-Store Retail Media Playbook 2024 states that the accuracy of sensor-based systems, including cameras, can vary with changes in lighting or store layout, and that regular recalibration is required 3. Traffic measurement in that playbook is performed with sensors such as cameras or passive WiFi detection to give insight into customer flow and behaviour 3. A bottom-up methodology uses always-on detection technology at or near the in-store screen, with the exposed audience calculated from sensors capturing the presence of individuals in front of the screen 3. That bottom-up approach avoids the assumptions and estimations an Impression Multiplier applies to raw counts in a top-down approach 3.
What the count represents
Accuracy also depends on the definition being counted. Norwich monitors footfall as the number of people going into a shop or business, since not all footfall results in a purchase 4. Poitiers buys external data from MyTraffic to monitor monthly pedestrian flows and map footfall per street 4.