Thanks for starting this thread @Surfingpsychedelicchicken! 

I've been working with folks mostly in the Gathering for Open Science (GOSH) community on giving wildlife camera traps stereovision capabilities. See this thread for our extensive discussion on the subject. Stereovision allows camera traps to capture spatial data about the scene in addition to images of wildlife, and this piece of information can greatly help estimate population size.

Recently, one of my colleagues developed a 3D-printed addon that could provide that function. It's looking promising and we're at 95% for a crowdfunding campaign on Experiment.com with just under a week left (also posted here). Sorry to sound self-promoting, but I genuinely believe we've got something going here, and would definitely appreciate some support big or small!

Happy to discuss more here or in the other threads linked to.

P.S. I love your point about anti-disciplinarity!

Super duper getting that mesh network working. Valuable resources. I have been working with IP camera systems since 2008 and have tried a lot of them. Ubiquiti itself is a great networking brand, the only comment I have about their cameras though is that they do not support a motion jpeg (mjeg) stream of any sort. What this means is that in order to get a still image for use in an AI computer vision system you have to waste GPU power somewhat getting the still. Generally speaking I would suggest to new things that have in their mind future AI integration is to choose a camera that supports an mjpeg stream with a custom frame rate. For example, I have 15x cameras around my home. They are all being analyzed in real time with the highest scoring object detector (Currently)  (That is published using the coco dataset from microsoft) yolov6. I also double check matches sometimes with yolov7 to provide extra resilience to false positives.

For this I tend to consume an mjpeg camera stream at 640x480 resolution at a frame rate of just 6 frames per second. That means that any system that is performing local object detection on something like the jetson series SBCs do not have to use any of their capacity first decoding mpeg streams to images.

Hi! @Surfingpsychedelicchicken ! yeah we run a conference kind of about this (well actually i've been helping run two conferences about this!) Dinacon and the GOSH gathering! Thanks @StephODonnell !

www.dinacon.org is about an anti-disciplinary gathering of people to explore field biology and design and it's a long form conference (minimum stay 7 days, whole conference is 1 month+)

GOSH is another group kind of like wildlabs here, a bit less focused on field tech, but still has lots of international folks working on open science hardware of many sorts

GOSH Community Forum

Let's talk about Open Science Hardware

GOSH Community Forum

And they also have an "unconference" that happens every couple of years (last year it was in panama) that is more traditional (4-5 days) but chock full of wonderful folks!

If you have a server and can host a solution, install on the server your own instance of  Next Cloud, Pydio  or similar - it's exactly what you need without reinventing the wheel. 

Strongly supporting the Nextcloud/owncloud suggestion -
apart from bringing everything you list in your requirements, it would give you a lot of features that might be useful as the project grows (versioning, failover for lost/failed data, sharing, ..)

Available for more advice, if needed.

Hi Jim, could you please update us on your final choice and experience so far?