A workshops at the computer vision conference I'll attend soon attempts to address this doozy of a problem in surveillance AI inaccuracy, described euphemistically:

"Computer vision methods trained on public databases demonstrate performance drift when deployed for real-world surveillance."