Thanks Jared, yes we've considered this for distance measurement in the past. And from that could infer a size, at least for images where the animal is clear enough to detect the animal edges. But my primary goal here is a bit more than just size, I would like to reconstruct a 3d model of the animal for fine-grained classification tasks. For example juvenile Kiwi vs adult, or small rats vs large mice.

I'll assume for a microcontroller then. In which case I can't help you with thermal. For Raspberry Pi type of implementation I have thermal modules that go down to 256x192 pixels with very good image quality.

I haven't discounted thermal imaging!  Long term the dream is for a multi-sensor device.  But I have no capacity right now for building new devices, and it isn't super clear to me which of these various sensors should be prioritised. My focus is on building datasets and figuring out which of these options could suit our needs should the hardware become practical in future.

Or use my ones. My thermal cameras come in a variety of sizes and in have the AI image detection as well 

Read "Wildlife Security Innovations thermal cameras on The Inventory | WILDLABS" on The Inventory | WILDLABS

Wildlife Security Innovations thermal cameras on The Inventory | WILDLABS

A range of tiny, low-power state of the art thermal cameras with flexible manufacturer support. 1280x1024, 640x512, 384x288x 256x192. Integrates with our AI camera traps.

The Inventory | WILDLABS

As usual Kim, you are -way- ahead of me
:-)

I was shown some tech today with cooled thermal imaging cameras that can detect the presence of a human at 20km distance. Mind blowing stuff. Using tiny sterling engines that cool the sensor to -150 degrees Celsius. I took a picture of one. I can sell you one if you like 😀

Not looking to make devices, I was just curious to know if anybody has already tried LiDAR already.  Even if it's impractical now, it's a field that is improving rapidly.  I try could build useful datasets in anticipation of them being needed in future.

I'm only interested in post-processing for the forseeable future.  I was imagining a combination of PIR sensor to turn the thing on when movement is detected, and a camera trap (also with PIR activation) so we know what it is we're looking at. The goal is to produce a 3d image with sufficient resolution to distinguish the various classes of interest that we struggle with using only un-sized 2D imagery.

Ah okay. Thanks for the clarification.

But are not the  difficulties that New Zealand research has  had to the past with thermal due to the use of very low resolutions ?

I understand that the equipment used there are flir leptons of just 160x120 or 80x60 pixels. It’s quite understandable that that is not going to give very good results.

The modules that I’m providing can be 640x512 pixels, come with a variety of lenses and draw just 0.7W. And they are tiny.  That’s a huge difference.

Plus with different lenses to give you a zoom magnification, the equipment doesn’t have to be so close. You can have large and high resolutions thermal images to do object detection and measurement calculations on.

I really don’t think you will ever be pushed to having to use lidar.

Here’s a recent dog detection in thermal with the 640x512 camera.

 

There are certainly a lot of difficulties you are facing. Waking up on PIR sensors can work but it's only fast if waking up a microcontroller, then you have the problem that they don't have much compute for anything. You can set a Jetson into a very deep sleep and wake it up. But I suspect the wakeup time will be somewhere between 20s-1m. So if you want to wake it up because there's a kiwi in the area for example, then that's okay, but if you are expecting to also trigger on the images etc then you might be disappointed.

Add to that the PIR sensors go off falsely a huge amount and maybe the very worst citicism of them is that they do not detect to very far away. Kiwi's are pretty small, I would expect that you could only detect them with a PIR sensor if they are very close to the sensor. Personally, I think that waking up on PIR sensors have had their day, the limitations are too great. We can only speculate as to how much data we have lost due to the limitations of PIR triggering. How many kiwi's are you currently missing? Good question. I've worked with PIR triggering of video capture of my own systems since 2008-2019, at which point I switched to AI computer vision and never used PIR triggering again, but I have observed their behavior for a very long time. Sure, heavy power usage is not fun but the gains that could be had are pretty large. If you want to go from A to B someday then you have to start somewhere. Experimentation and research is needed. 

There are a great many research opportunities available to see how much monitoring improves when you take the limitations of PIR sensors out of the equation.

Assuming that you could get lidar modules to work okay. They likely generate a very large amount of point data so what kind of compute are you going to use to process that? I somehow doubt that microcontrollers are going to cut it any time soon, then you are back to Raspberry Pi type of compute in which case you can get great results with thermal.

The photo of the dog above was triggered with continuous AI object detection with about a 1s inference time and the module in the photo below  on a Raspberry Pi based engine, together with a 25 million parameter high quality object detector model. Something you could not even get close to on a microcontroller.

Reach out if you want to have a meeting to see the modules I'm using if you like.

Thanks Kim, yes we're aware that PIR has many pitfalls, especially when it comes to small things far away, which is often the case for us.  But we can accept missing some detections for indexing purposes if we win in other areas like battery life, size or affordability.  We're currently able to detect small birds from a useful distance with Browning trail cameras. 

If I do dust off my soldering iron & actually make something it will be on a scale of 1, for dataset building proof of concept stage, not for large scale use.  It would likely include a variety of sensors so I can make meaningful comparisons & propose a sensible combination.  The detection part could be as simple as pixel subtraction from video footage I imagine.   

That’s pretty cool. It took me a while to realise that you were referring to a software package. So for others like me, here’s a link

 

Thanks Jack, that's an interesting repo.