Unfortunately not quite so good. Maybe there exists a model that can do facial recognition. But the model I have loaded on mine just spits out “dog”, or “person”, or “car”. The false-positives I was referring to it not having, is what you’d typically get with a pixel-based motion detection camera. Where if it sees a leaf on a tree move, it alerts you.

Mine, at least that leaf needs to look convincingly like a person.

You can read more about the Coral at https://coral.ai/models/

I have a frigate setup. I run it through Docker and I even have the Coral AI processor chip hooked up. Which is pretty neat, runs local pattern recognition for people, annimals, etc. I use generic IP cams on their own network. I think pretty much anything that supports RTSP would work. Then hooked up to HA via MQTT, again all in docker. With the coral, I only get notifications if it actually detects a person. The false positives are extremely rare. And I use Tailscale for access from outside the LAN

You could be right. I am not a pro so I don’t really want to speak on the best practice approach. Really the only reason I containerize my services is the ease-of-deployment and the ease of potential re-deployment if my server did crash.

I personally am not too stressed about bad actors, being as this is a hobby server and the payout for a bad actor would be pretty low.

But your point does make sense to me.

I also do this. Just run Tailscale on bare metal and then I can access my all my services the same as if I was on my LAN, essentially.

+1 on duplicity. I run it directly on the host, outside of my docker containers. Grabs the data from the different volumes for my Nextcloud etc, puts it all into an AWS infrequent access bucket. Costs me ~3$USD/month. Pretty simple. Runs on cron