And I'm positive that people in Greenland would be very familiar with what actually works and what doesn't in people/wildlife conflicts. You would actually expect that sensible people would look to Greenland and places where these conflicts occur before forming their policies.

Nice to see this, I think it's the right approach.



What I want to bring to this is the use of the larger and more effective models. Models that won't run on micro-controllers or even raspberry pi's. Specifically the Jetson series SBCs and the more recent one's that have more processing power and memory.



Over the years I've seen just how much more accurate empirically the improved models are and eliminating false positives is always very important. W.R.T the comment below about not letting them get used to the sounds, here again I think that making the sounds less often due to less false positives would also be a benefit.



The highest scoring object detector model (Based on the coco dataset) that I know about right now is the yolov6 model. This is a pytorch model. These are often harder to get working on the Jetson series boards because they have a lot of build dependencies and are not installed by default. With my project:



 

I have solved the dependencies and made pytorch and torchvision builds for the Orin series SBCs. Although... I haven't updated the installation process just yet, I've been side tracked by my gunshot detection/ARU project, but if anyone particularly wants to try this out I'm be happy to help more directly.



I am ambititions to convert some of the annotated wolf images from imagenet to the coco format and then re-train yolov6 to include specific wolf categories (Or "a" wolf category) and then make that available for running on the Jetson platforms.

 

For a lot of use cases one can simply use the dog category. It can be enhanced also to potentially detect dogs and alert only if there is also no person detected for now.



I think a specific wolf training would likely only be needed for wider monitoring uses, such as monitoring public areas or places where a live stock threat is not present, then in order to reduce the noise alerting specifically when wolves are detected would be helpful.



I run yolov6 

on some night image captures of a friend of mines husky to see how well and far it can detect a dog with the pre-trained model and let you all know.



Note that my software system also has a lot of hooks for custom responses as well, including out of the box support for USB connected Phidgets so you can easily integrate relay responses as well as connecting buttons to control with it. This is handy for the Jetson series computers because they are expensive (Around 1000 euros for a 16GB recomputer J4012 Orin, with 1024 CUDA cores.) and you wouldn't want to blow up the board with static electricity. Plus Phidgets 

are very easy to use here, just plug it in and it's working with gui configured input and output.

Very nice to get a chance to learn more details about the Romanian bear project! I have been toying with the idea of doing something like this with polar bears and have also been in contact with the Hack the Planet guys regarding this. First I am going to try and get people with polar bear camera trap images together and see if we can assemble a proper training data set. Please contact me if you have polar bear trap images and if you are interested in participating in such an effort!

Thanks for the plug, Rob :)
Boombox has the ability to play MP3s based on a trigger input or just on a timer. If there was an existing AI detection system that would send out a pulse on detection of a wolf, then we could trigger off that. It's currently being used with coyotes and also a study with wolves in the Italian Alps, but being targeted by specialists whose study species are wolves. 

Ohhhh, the Italian alps, that sounds very attractive, let me know if you want to try out an Orin system there and need some help :)