Interesting discussion for sure, just joining in
I agree that installing openCV is something you would rather avoid. Since I needed a couple of RPi3's running the same stuff, I decided to make an image with openCV and other "personal defaults" configured. Then just burn a new sd, boot it up and re-configure the ip for the network. But once an update is needed, o boy, well that will be the day
When I looked into using openCV in general for video streaming, I noticed that the cpu load was much, much higher than when doing the same simple task with Motion. When it comes to object analyze & detection, openCV with the various dnn models are great, they are pretty accurate. I think the key is not to send openCV too many pictures for analyze. I stop analyzing once the object "person" is detected and then notifies me with a picture of that detection (but I continue to record images pre-and post event allowing me to make an investigation at any time later)
I believe (and hope) that a RPi3 would be able to handle more IP-cameras than USB connected ones. I would like to have devices or similar like the ones below. But it will require more cabling, at least in my house. So maybe next house.
