Assuming ATMs take and store a photo every time they're used, then it'll be a matter of matching the photos to photos of people already in the corpus of training data several times. That makes it quite easy.