Also logistically, it is nearly impossible. The idea falls apart when attempting comprehensive surveillance. There is simply too much noise and too much uncertainty from the low quality, at distance, uneven illumination video. I write FR software, and large installations typically fall apart because the humans are simply too lazy, they delegate too much that should not be delegated, they are lax and frankly incompetent. Outside of developing tech orgs, FR is too complicated for under-educated persons like the police. At minimum, you need to understand probabilities better than a typical consumer.