An AI age check for the self-checkout at Albert Heijn to go. A camera estimates whether a customer is clearly over 25, and an employee always makes the final call.
Problem
At a self-checkout, buying alcohol or tobacco means waiting for an employee to check your age. In our survey, 80% of respondents run into this check, and waiting for staff is the biggest frustration (58%). A manual check takes about a minute.
The system
With Smart Scan, customers can choose to let a camera estimate their age. Two ResNet-50 models work together: one sorts faces into nine age groups, and a second looks closer at ages 15 to 35, where the line at 25 matters most.
The estimate goes to the employee’s handheld terminal with one of seven confidence levels, and the employee decides. The whole check takes 8 to 15 seconds.
Design
The interface went through five versions, from eleven screens that tried to do too much to a short flow with large touch targets, plus separate screens for the employee terminal. We also worked out how to stop people from fooling the camera, for example with a photo, through liveness detection.
Research
Alongside the project I researched how glasses, sunglasses and caps affect a ResNet-50 age model trained on UTKFace. Without anything on the face, the model was off by 4.78 years on average. Glasses raised that to 7.69 years, and glasses with a cap to 8.27. People over 60 were affected most.
For a self-checkout this matters: the system has to notice when a face is covered and leave the decision to the employee.
Storyboard, flow & screens
From storyboard to the final screens, plus my research poster (in Dutch). Tap a page to see it large.