Haven't kept up on H/W acceleration - WOW!!! "Jetson" does indeed looks interesting (actually, pretty incredible..) !  



I was first thinking about establishing a standard way to just getting the raw reverse-proxy NAT stuff working for auditable, admin-level access... Beyond that, a "standard" UI resource that would run on the installed box for crunching the SD cards full of data folks are already routinely collecting from Audiomoths, drones, and camera traps... perhaps within the "EcoScope" workflow framework. 

Read "GitHub - wildlife-dynamics/ecoscope: Conservation data analytics · GitHub" on GitHub

GitHub - wildlife-dynamics/ecoscope: Conservation data analytics · GitHub

Conservation data analytics. Contribute to wildlife-dynamics/ecoscope development by creating an account on GitHub.

GitHub

Realtime alerting from advanced systems like yours ... that's a whooooole 'nother level...indeed, the tradeoff of how much "extra" inferencing capacity a Jetson has after its primary task to handle ... what ... 1, 2, 3 ..? additional A/V streams? Listen for gunshots and chainsaws or human conversation? Run other much more complex models?   



Just wondering what a baseline SBSA host configuration with a few terabytes of SSD would look like / cost...that can work in the tropics... And if there's interest among readers for the monitoring scenario ... "if" it were user-friendly and could be remotely administered ... (i.e. easy to install and start effectively using the sytem..)

I worry about overheating in the tropics ... perhaps with lots of heat sinks and some fan spares one might get away with an RPi5 with something like: 

Read " Vilros 6 Piece Heatsink Set for Raspberry Pi 5 – Vilros.com " on Vilros.com

Vilros 6 Piece Heatsink Set for Raspberry Pi 5 – Vilros.com

Set of 6 heatsinks to be applied as shown in the image. Heatsink has a Preapplied Thermal Tape on the back for easy application.Each heasink has an engraved number to easily be able to follow the graphic and install it in place. 1. Wi-Fi Chip Heatsink Size2. Power Management IC Heatsink Size 6mm x 6mm 3. 4. BCM2717 Pro

Vilros.com

Read " Vilros "WAVE" Raspberry Pi 5 Compatible Aluminum Alloy Active + Passiv – Vilros.com " on Vilros.com

Vilros "WAVE" Raspberry Pi 5 Compatible Aluminum Alloy Active + Passiv – Vilros.com

The Vilros "WAVE" Raspberry Pi 5 Compatible Aluminum Alloy Active + Passive Cooling Case is a Case designed to passively provide massive cooling to a Raspberry Pi 5 board and also provide active cooling for when the passive cooling is not enough.Not only does the case contact the onboard chip to pull heat away from the

Vilros.com

and ~2TB of SSD storage ... optional expansion via USB-C

make the Jetson board an optional awesome add-on ... 



What would this thing have?

• remote admin/upgrades via reverse proxy (to where...?)
• Samba server with known user and directory layout so folks could easily upload files and get results on shared drive
• basic user (Ubuntu?) desktop and apps
• Birdnet-Analyze batch; local 'Chirpity'
• OpenDroneMap
• camera trap processing (Zamba cloud, Pytorch Wildlife, ..?)
• Ecoscope workflow engine for batch process management
• perhaps an optional SMART connect server ? 

Read " SMART Connect " on

SMART Connect

- 

If you are dealing with memory constraints and slow processing times, you might want to migrate Ecoscope to Rust. Its more lightweight, memory efficient, safe to parallelize, as fast as C in some use cases and doesn't come with all the python environment setup headaches.
Its also much easier to read than C++. You can migrate one module at a time and port them into your python projects using Py03. There are pandas, pytorch and numpy analogs in Rust and multiple ways to expose UIs for the underlying components, including web based UIs and native Rust GUI's. For example Tauri allows you to have a Rust backend with a Web UI and runs on linux, mac, windows, ios and android. There are also other options if you need a UI on hardware that doesnt have a dedicated GPU.

yeah; definitely intend to do things as modules in EcoScope if possible...

Seems like a full ecoscope on the edge, processing audio and image and drone video, would be like carrying an elephant on a bicycle. 
If your first interest is your AudioMoths, then to start simple you could use snakemake (https://snakemake.readthedocs.io/en/stable/) as your workflow management, and have someone like @Leoman99 write some audio processing in a fast, memory efficient language like C or Rust. Get that working on the edge first, and then add modules incrementally. 
Do you need a UI if these devices are going to be out in the bush? Would it not be enough to just take your laptop into the field, do a data dump whilst you are out there, and visualize it when you get back to base? 

Interesting stuff! 

I can certainly understand/appreciate the need for complex topology VPNs, but this use case is way simpler: massive amounts of data production at one location, with a skinny cloud connection. This is just to co-lo a compute and storage appliance that can be remotely managed; (Kim, please correct me if I'm wrong) spoke-hub topology to one or two jump-host admin networks using straight-up open source Wireguard should work just fine...

 

It works for me me in any case. I have used tailscale in the past. But it wants to try and lock you into subscriptions. So indeed, I use pure wireguard now.

Our solar powered wolf monitor has been in the field for 6 months now. It think it's proven that it can stay in the field indefinately, whilst running local AI now :-) My systems don't use SD cards, they use an nvme SSD so data reliability is very high. When I get a wolf detection, I copy over snapshot stream for further training data then copy a full frame (25 fps) sections of h264 video to a .mov file. I never have to bring the unit back to the office or visit the site to copy data off it.

This is what our pi-based portable versions look like:

could such a thing be generalized to handle batch oriented bioacoustics, camera trap data, and/or multiple streams via directed HaLow wifi, etc.? 

(with consulting, support offered by companies in the Netherlands, etc)

I don't quite follow what you suggestion is actually?

In terms of halow/directed wifi. Yes, that's just networking and it is a fully fledged linux, so indeed it handles multiple streams now. Camera streams and also inference streams.

Making a general purpose platform out of it is difficult not being a general purpose computer with open access.

Our focus now is goals towards sustainability. So that's over the air updates, then scalability. Because sustainability means volume sales in the areas that people want. We see that in human-wildlife conflict migitation at this stage.

I'm also not sure what you mean by "support offered by companies in the Netherlands, etc". We are a company in the Netherlands, but just one company in the Netherlands.

Nice you are thinking along.

It could be that training at a remote station is likely only viable for very small models unless you have significant amount of power at the remote site. For our polar bear detector with @Lars_Holst_Hansen  which uses yolov9. We trained for a week on an rtx4090 to get a reasonable level of interations. That's got 16384 CUDA codes. On a midrange jetson with 1024 CUDA cores that rougly translates to 16 weeks. Likely more due to batch sizes.

Inference on the edge with better quality models is absolutely viable. But it requires accepting that you need more power than microcontrollers and dealing with it.

yeah - no question training complex models, such as real-time video, requires substantial compute, and hence considerable power.

I was mainly thinking of inferencing against pre-trained models, and basic motion or embeddings/hyper-parameter similarity detection to aid with human labeling.  And of course just grunt work like running OpenDroneMap orthophoto and mesh generation, etc. And as a platform for other compute-intense, on-site stuff. (eDNA processing..? etc)

Oh yeah. On a Jetson that sort of thing would be viable.

Our Pi 5 is really pushing the boundaries with what we have it doing.

To put it into perspective. Our polar bear detector is round robin checking 22 cameras with a more than 20,000,000 parameter model on a middle of the range Jetson.

Yes, I was considering inference only in this case. I see it as a filter applied to raw images and audio before uploading to cloud/server to save bandwidth (if it is necessary at all?), which I think was the OP's request. 

It is then a balance of power requirements and power stability (and physical security?) to keep a small server running and connected. But a starlink system is relatively power hungry for a remote station too, so they may have this covered.

Software wise, the "TRAPPER" system might do what you need, with a local/LAN installation and allow for admin access externally. https://os-conservation.org/trapper/

nothing like a concrete, interesting problem to find focus
:-]

my partner is into chickens ... trust me, it's a whole subculture 
🤣

We were talking with friends about pecking orders , and thought... hmmmm

so, after riffing with Claude a bit, assembling RPi4+ USB Coral acclerator + Tapo 120 to tinker with RTSP streams Frigate Plus , COCO, Pytorch, etc. to automatically determine a pecking order, from an RTSP webcam.
Claude's got the strategy all figured out! 🤣

And we've got an Audiomoth and a StealthCam we use on occasion.

SO, managing/evolving that RPi4-based from behind the firewall at her house, versus mine, will be fun. 

Just installed the Wireguard client; off we go...

(Viv, top of the pecking order, sole survivor of a previous weasel attack; all the others are newcomers...)