This is really cool and really scary at the same time. Sometimes being in this industry is absolutely amazing and sometimes you have to step back and say, Did I really just do that? Referring to making big brother even more powerful!!
I'm not too concerned yet. The article is very light on technical detail, and knowing Brian's inclination to nail that sort of thing, it's probably because the vendor is light on sharing it. A cop's body camera that beeps away with false alarms is not going to be accepted. And high accuracy is at odds with low cost and small form factor edge devices. It's a great idea, a good vision and deep learning holds the promise of a lightweight footprint, but there is no detail yet. What complexity can it learn? What is the impact of environmental conditions (shadow, low light, direct light, motion, etc.)? What is the accuracy in terms of false reject and false accept? It ain't real until this is all quantified, hopefully by a credible third party.