Even if the false positive rate is really low, it's still a problem. All of these dragnet-style things mess with your priors. In a traditional investigation, the police have a handful of suspects and then gather evidence. Circumstantial evidence showing that the chances are 1 in 1000 that this one of the five suspects was not the culprit is pretty convincing, and probably enough to convict.
Cast a dragnet over a city of 1 million, and suddenly that same 1-in-1000 error rate will be giving you 1000 innocent people who, under the old system, are 'guilty beyond reasonable doubt'. They've been pre-selected to look pretty frikkin' bad regardless of whether they're guilty, which means any one of them will have a hell of a time proving their innocence.
I expect that if you knew that the cameras were watching, you could trick their AI's pretty easily (e.g., by faking a limp), but I'd prefer to confirm that before relying on it :-).
Kind of reminds me of Dunes "Walk without rhythm so you don't attract the worm"
What constantly amazed me was how you could see someone walking a quarter mile away and instantly know exactly who it was.
The South Pole is a smaller population than China so I'd be worried about false positives (probably mostly within families as kids learn to stand and walk like their parents and have similar dimensions when fully grown) but I suspect it's actually a pretty robust identification technique.