Hi George! Ah that's great, this is going on my reading list - thanks a lot!

In case anyone else was trying to read this, here's a link that is outside the csu login :) 

Also this early wolf system also using tags: 

and the bear one I mentioned originally (both this and the wolf one led by the inimitable Stewart Breck): 

The study testing automatic hazing of wolves seems very relevant for the ideas you have regarding wolves in the Netherlands, @kimhendrikse !

No, you are spot on Rob. I refrained from commenting because I didn't want to come over as domineering in the topics, I've commented on a lot of topics lately. I think that if you have power and networking or some other way to send an alert this approach is the ideal early warning detection system and my system lends itself to this purpose out of the box :) But perhaps it's time to show something.

Rob, maybe we can setup a live demo. Do you have a site whereby you are in a position to expose a rest endpoint/url to the outside world with a password protected mjpeg stream of around 640x480 sizeish at around 6 fps ? If so, then I could hook into it from here and demo this live during a variety hour sometime. I could provide a URL for the viewers to connect to and explore the user interface, and I could demo the remote popup system that works over the websocket interface? What does that lookup, attendees could connect to a URL that I provide and when the animal triggers it they would get a popup on their screen. Nice demo huh?

If Rob can't do it, then anyone else want to do this? The network details of your stream would be white washed through my system.

It's also good to demo sometime long distance detection. The following image of roe deer was taken through a night vision device and a swarovski telescope and an iphone. At night, with no additional lighting at a distance of around 340m. Detected with deepfaune (Processed today).


As you can see it matched just fine.

This is an optical telescope. I suspect that with a large aperture reflection telescope of the kind used for astronomy maybe this is possible to a kilometer away? I also suspect that a night vision device (expensive and limited tube lifetime) is also not needed. I suspect that an ultra sensitive HIKvision colorvu camera would likely also work.

How would you like an early warning system that can detect bears or poachers at night from 1km away? :)

I saw an amazing demo once at an arms fair that I was able to attend whereby a special telescope with able to read the licence numbers of ships 16km away. It used a pulsed laser and an electronic gate in sync with the laser such that only the photons that would have been reflected from that distance were gated through. Totally amazing tech.

By being able to detect from really long distances away from lookout towers you can take away the problem that the requirement for a power source provides. This should be coupled with really high end full sized models. Like in my system.... :)

PS. No that's not a wolf in the background, I think it's another deer but it's head is not showing.

Hi @kimhendrikse I don't have a site for wolves (or dingoes) where I have permission/ethics clearances at the moment, although I could possibly arrange one with another researcher, but it might take some time to arrange...can your system also process pre-recorded footage in a similar way, just as a demonstration?  

Cheers,

Rob

I could probably hack up a Python script to do that.

That’s indeed one approach.

Maybe some practioners can send a few random videos with wolves (or in my case, dingoes in them) and you could run them through your system? Or what's the easiest way for you? I don't want create even more work for you @kimhendrikse !!

Kind thoughts Rob 😀

I think that would likely involve more effort than the value.

I’ll give some thought to the approach.

Ok thanks Kim. Cheers,

Rob

I am looping back into this discuss as I'm finally going deep into this topic - awesome, thank you for sharing these rob, this is exactly what I was hoping for!

Hi Henrik,

I've had a look around your site, but I can't find the details. I was wondering, what model are you using in The Sentinel and in what size ? So in the case of mobilenetv2 for example, the scale factor, for other models they often come in sizes, large, medium, nano etc.

I've look up the specs on the TPU SOM. Which size are you using? I see they come in 1GB, 2GB and 4GB versions.

It's really hard to compare models. Because it depends not only on the model but also the data that you use to train it. The closest thing we have to a model comparision is the coco dataset. I assume you have tested your model against the coco dataset, I was wondering what mAP (mean average precision) values do you get? These provide a strong indication as to what quality of matching one can expect.

I think it's important to publish mAP values for models used for wildlife purposes it's the only way that people can develop an intuition about that quality they can expect from a model.

To provide an idea as to it's important, if I lookup the TPU I find sales references referring to mobilenetv2. The lartest model for mobilenetv2 provides a mAP score on coco as 22%. Back in 2019, yolov3 came out in many models. That's when I first integrated a computer vision model into my system. yolov3 tiny model has a mAP value of 23%, higher than the highest scoring mobilenetv2 model. I found it to be completely unusable for detecting people (Think poachers), it was thinking that the drain pipe along the house was a person.
 

Also speed of processing is important if you have to use larger models. The TPU SOM 1GB version advertises a speed of 1 TFOP of INT8 calculations. The NVidia Jetson orin NX would have about 85 TFOPs of INT8 calculations. But it would draw more power. The whole unit is drawing around 20W in my usage. I think the 1GB TPU SOM would draw around 1/2W, but you have to add the power consumption of whatever processor you are using, that would not be much for a microcontroller, but significant for even a Pi.

I'm going to open up another thread on this topic because I think it's important to know these values in order to weigh up the importances of accuracy versus things like power consumption and the cost of sending out an anti-poacher team to a false positive and indeed also the cost to replace stolen units when they make their choices. Let's get a discussion going! I'll create a new thread this evening :)

Awesome, thanks a lot henrik! Interesting to hear - let's have a chat, I'm curious to hear more about the applications you're seeing! 

Hi @Henrikcox this is very exciting news about the Sentinel. The LoRa and cell options sound really great for a lot of urban/semi-urban areas, so a big potential winner there. Does the Sentinel offer any sort of GPIO pins that could possibly be connected to something external (e.g. like a sound unit that can playback scary sounds) if a detection occurs? 

Also @StephODonnell I'd be super keen for something like a 'WILDLABS working group' or similar on this, as it such an important topic. 

Rob

INTERESTING. Okay hold tight on that idea. We've got some events coming on this topic (hence my questions), so a working group could follow on from that? 

Hi @Henrikcox 

I hope you are well. 

I am not sure if you may remember myself and my colleague from CLS. We had a meeting and met in-person at Earthranger last year. 

I would be very keen to discuss your current satellite limitations and suggest options based on your required outcome. Please let me know if I can assist?