So I’ve had this (perhaps naive and not novel?) idea for a while and am curious what people think. What if it were required by law that all surveillance video is “anonymized” in real-time via computer vision software? In other words, use an image segmentation model like Mask R-CNN to cut out or blur each person appearing in frame. Violent/criminal behavior can still be detected in real-time either via software or by human operators (you don’t need to see someone’s face or skin color to determine if they’re e.g. committing a violent act).

If a crime is committed and the video is needed as evidence or for identifying the person(s), then the archived video can be “unmasked” via court order or other process. There is some precedent for this, relating to when US citizens incidentally appear in foreign phone calls collected by intelligence agencies (see: https://en.m.wikipedia.org/wiki/Unmasking_by_U.S._intelligen...).

This way, there is a reasonable level of privacy preservation (=anxiety reduction) for people innocently going about their lives. It also limits opportunities for abuse, like stalking and discrimination. The effectiveness of surveillance systems as a crime deterrent and source of evidence is still maintained under this policy. That is, it’s basically a Pareto improvement—no one is made worse off, while many are better off.