The thermal camera can catch micro expressions much better than what you would catch with a flash if the animal is close enough and the resolution is sufficient. If that is a important enough. That been my experience with the videos we have captured.  We also have 1280x1024 thermal. There is typical no motion blur with that.

As to how an animal would act under continuous surveillance with one of our cameras I cannot say. That experiment has never been done with our gear before. There are variety of fixed lens available. The lens I’ve shown are general purpose reasonably wide angle lens. But if you wanted to study behavior from much further away there are lens to observe from very far. They are just expensive of course.

The modules we use have impressive onboard image processing that really bring out the details. Unlike ones I’ve seen before.

Have a look at the videos on our channel and see what you think. The earliest video on the channel is in 1280x1024.

https://youtube.com/@wildlifesecurityinnovations

The thermal modules that we use with our system are quite new and I've used a lot of thermal gear over the years and none had images as good as these ones. As to suitability for behavioral research I can see that there is much more detail available with this gear than what I see with flashing traditional gear. But I'm not a behavioral researcher in the field. So the suitability would ultimately be up to the researcher themselves. But never before was there a 1280x1024 thermal camera available for use in behavioral studies, so the suitability thereof has never been evaluated before.

What resolution, age was the thermal gear you used before ? And what distance were you filming at ?

I'll definitely check out your channel! To be honest, I'm much more on the animal science and behaviour side of things than the hardware engineering side, so I can't speak to the exact specs of the older gear I've seen, but having that high of a resolution without motion blur definitely sounds like a massive leap forward.

The way I see it, the trickiest part about any remote tech, no matter how high-res it gets, is that you can never truly measure the 'avoidance factor' from behind the lens. If a cautious animal senses a foreign object in its home ground and decides to completely steer clear of that zone, the camera will never catch it. You only ever get data on the animals that don't mind the camera, which can unintentionally skew the behavioural picture.

It’s the classic observer dilemma. It’s why some of the most famous animal behaviourists in history only truly understood nuanced behaviour by actually embedding themselves in the environment and becoming part of the pack, rather than relying solely on a fixed lens. But as a tool to bridge the gap where humans can't go, it's definitely exciting to see how much clearer the visibility is getting!

I would love to have feedback from a behavioral researcher. When I made the comment, I was mostly thinking about macro behavior. I haven't dived into the requirements for micro research, but it sure is interesting.

The area where I hope to make the most impact is in human-wildlife conflict mitigation. I would be thrilled if it turned out to be useful to behavior analysis as well.

Occasionally we get a close close up of an animal. Such as this hare. I would love to know whether in you consider it contains sufficient detail for behavioral purposes.

Hare closeup in Thermal

And here, even better

Ring side view of a hare close up

Over time, the animals do get comfortable with our gear. I'm sure that a bird built a nest under the panel recently. I just haven't been out there in a while to check, but it keeps flying up from below in the area. We have a visible view of that.

Baby bird living under a solar panel

Most of our wolf videos are on our other channel. Here the wolves indeed were very wary of the gear at first. Mostly they would glance up, however at first they would have been looking at the camera, but over time I think that most of the time they were looking across the field to the road on the other side.

We also have a 4K ultra low light camera that we were lighting with invisible (940nm) lighting. This we have also recording continuously.

Wolf with 4K ultra low light camera
 
We custom design all weather enclosures for out thermal modules. They are design such if you wanted you could remove them and use them in a stealth custom made enclosure of your own. They are USB based modules, so the main recording unit can be hidden away from the camera. Here is a photo of a 640x512 unit

There's a camera mounting fitting underneath so you can can ball joint camera mounts to mount them on and only a little bit sticks up into view.
 
The 1280x1024 resolution module is a bit bigger

That would be an incredible addition to Sparrow Studio! You are free to fork the repo and add to it, and through a pull request we can even fold it into the one we're developing. 

These are very helpful information, Kim. Also, we are moving away from Jetson Nano to Raspberry Pi 5 and Pi zeros. We have made a dedicated inference engine for these devices. Now, we can run megadetectorv6 (Yolov11 models) on the pi zero, and two Resnet18 models concurrently on the pi zero. I am specifically mentioning the pi zero here because it is the lowest level hardware we are working with, and it only cost ~20 usd. With two resnet18 and 4 threads, the power consumption is around 4-5W. If we are only using 2 threads, it is 2.8W. These data make the application of these lower end chips very feasible in the field. We have recently put a pi zero 2W sparrow on a bouy with gsm. the average power consumption with all sensors and modem together is around 7-9W. In terms of Pi 5, except for transformer based models, we can run most of the CNN models with a efficient power load.

@carterjandrew These are really good questions. For now, we do not have an independent website for Sparrow. The Linkedin page in my previous post is the best place for now. That is our main communication channel for updates and progress of projects like Sparrow, MegaDetector, and PytorchWildlife. Please follow us there. 

In terms of deforestation, we do have plans to add this functionalities in the future because we have a line of work specifically on deforestation assessment under the Project Guacamaya initiative. It will be very similar to the overhead animal detection professing but on forest segmentation. 

For people sharing thoughts, we will ask some of our partners who are currently using Sparrow to share some public feedback to us. It will also likely be on Linkedin. Please keep an eye on it. 

In terms of tooling, we are designing Sparrow to be as tool agnostic as possible, so it is reduced down to a simple data collection, management, and processing platform. How data are collected, managed, and processed is totally up to different users and projects. That's our end goal. 

Same as our software. We are it to be as open and extensible as possible. We want people to be able to create their own "Sparrow Studio" on top of our base Sparrow Studio. Once it is open sourced, it should be possible. 

Please let us know if there are any questions, and feel free to send us emails. Thanks! 

Hi there! Thanks @Adrien_Pajot for the call-out as I didn't see this. Sandro, I completely feel for you with the challenge! We had attempted to use the motion trigger feature of the cameras, but ended up just recording video and using an object detection approach within the video. I should mention too that our results were not perfect, but our aim was to mostly work out the process, not refine it so there's certainly room for improvement and increase in sample size and training development. But I did think that there was a bit better luck with the nighttime moth detection as compared to daytime (if you take a look at about minute 4:25 in the video at the bottom of this link https://www.oceanscienceanalytics.com/terrestrial-pollinators). I think if we were to continue development, nighttime monitoring would be an interesting route to go. I am not sure about the 850 nm for the Voopeak IR but it didn't seem to attract the insects in the data I reviewed (but I could certainly be wrong!). 

I'll be interested in learning more about how your project goes!  

Hello! The "MIB Paradox" is absolutely real, and PIR sensors are a dead end for insect detection due to their lack of thermal mass. However, falling back on a regular-interval time-lapse is a trap that will likely kill your battery and cost you the most interesting data.

Writing to an SD card is highly energy-intensive (causing current spikes often over 100mA). Doing this blindly every few minutes to save static pictures wastes a massive amount of power on empty frames. Worse, if a bug does show up, you only capture a single frozen moment and miss the entire temporal dynamic of its behavior.

Instead of a blind time-lapse, you can use your camera and a lightweight microcontroller (like an ESP32) as an Intelligent Optical Trigger. Here is how you can bypass the paradox:

• The Low-FPS Radar: Configure your camera to capture images at a very low framerate (1 or 2 frames per second) and in low resolution.
• The RAM Buffer: Do not write these interval images to your SD card. Keep them strictly ephemeral, stored only in the RAM.
• The Binary Judge: Run a tiny, highly optimized machine learning model (TinyML) on that low-res frame to answer a single question: "Is there an insect?"
• The Duty Cycle: If the frame is empty, the system discards the image and immediately drops into a fractional deep sleep for the remaining milliseconds of that second. The SD card is never powered on.
• The Action Trigger: If the AI detects a target, the system wakes up entirely. It triggers your MOSFET to turn on the IR lights, switches the camera to a higher framerate, and begins writing the full video stream directly to the SD card.

This architecture gives you the best of both worlds. You get the energy efficiency of a motion trigger by treating disposable low-res frames as a "virtual PIR," and you only pay the heavy energy cost of lighting and SD card writing when there is guaranteed action to record.

Hi Kim,

I come from automotive CV where false positives around vulnerable road users are a constant challenge, especially with edge cases at night and in low-visibility conditions (in Greenland or Canada winter conditions might skew the video clarity).

I’m curious about how this is handled in conservation/anti-poaching setups, particularly in IR-based detection systems that can pick up humans at range in darkness.

In automotive we rarely try to classify object intent, rather just direction of movement and proximity, so I’m wondering how systems in your context avoid over-interpreting a detection (e.g. differentiating a hiker or worker from a genuine threat scenario), and what role something like restricted location, known poacher trails, activity, or time of day might play into interpreting the detection.

Is the system usually designed to be triggered based with a manual triage backend or if there might be some degree of automated triage? Or if the methods you use are mostly for animal detection a la camera traps and human detections are an added benefit?

Would be great to hear how you structure that pipeline in practice.

Thanks,

Ron

Elionai - your point about lessons from past deployments and "what tends to fail first" really resonates. I think that gap between ideal-condition performance and what actually holds up in the field is one of the most underrated questions in this whole space.

I'm building something that integrates environmental monitoring, so I'd love to pick your brain on the edge/deployment side. Messaging you to connect!

Thank you Tessa. I have filled up the google form as well (Bioacoustics software database submissions).