Wow, what a great project.

This is a great project! Some comments:
RaspberryPI though accessible is not the best fit for video pipelines and AI workloads or off grid deployments:
- it lacks onboard ISP which means either software implemented ISP, distorted data or on camera non energy optimized ISP.
- it lacks any power management techniques, low power modes, etc.
- it runs from SDCard using the same one for OS, swap and data, any corruption can lead to full loss.
- it runs any AI/ML workload on CPU which is extremely non efficient and any addon accelerators such as Hailo8 add a lot to power consumption and heat dissipation representing more challenges.

The advantages are of course plenty of documentation, community and all kind of makers addons, hats, etc.

For something more realistic, real life suitable I would suggest using something based on SoC with integrated NPU such as Hailo 15, Renesas RZ/V, Synaptics SL1680, MediaTek Genio or even the I.MX8M Plus for very light AI/ML workload. All of these have variety of SBCs, kits or even standalone smart camera oriented designs available from different vendors.

Hi, are you looking to import these? Do you have any import tax considerations? This could impact which models you buy. I have been using Acorn models, very reliable and provide photo and 4K video with sound options.

Best wishes

Susan

Great guide — this is exactly the kind of resource the community needs. A few additions from a hardware embedded perspective that might be worth including:

On PIR sensors — the standard Fresnel lens + PIR combination has a fundamental limitation in hot environments: when ambient temperature approaches body temperature (~35°C in African savannah), the thermal contrast between the animal and the background drops dramatically and trigger reliability degrades. This is worth calling out explicitly for tropical and arid deployments, where the standard PIR may miss animals during the hottest part of the day. Some teams have moved to passive radar (Doppler microwave) as an alternative trigger for hot environments — less species-selective but more temperature-independent.

On power architecture — one thing I'd add to the component deep-dive is the power switching circuit. Most commercial cameras use a simple battery holder with no protection. For DIY builds, a proper battery management IC with overcurrent protection, low-voltage cutoff, and reverse polarity protection adds almost no cost but prevents a lot of field failures, especially when using lithium primaries in extreme temperatures.

On IR illumination — the choice between 850nm (faint red glow, better image quality) and 940nm (truly invisible, lower image quality, shorter range) is well covered in most guides, but what's often missed is thermal management of the IR LEDs themselves. High-power IR LEDs run hot and can significantly raise the enclosure temperature in a sealed housing — worth mentioning as a factor in enclosure thermal design, particularly for cameras that run night-long video.

On the shift away from hardware — curious what drove that decision. Was it the enclosure/thermal challenges, the PIR reliability issue, or something else entirely?

Thank you for sharing.

HI

This what I am trying currently todo

https://wildlabs.net/en/discussion/looking-collaborators-ai-powered-backyard-bee-monitor-citizen-science?language_content_entity=en

I have already some algorithms and some monitoring set up in development

feel free to reach me maybe we could join forces

Hi Antoine,

I had not seen these before, but I'll echo Rob in wondering if the radio links in these are truly what most would consider 'LoRa'.  That tech/protocol generally has very low data transfer rates and would be quite challenged in sending pictures.  That said, what they call it may not be relevant if it works for you. I would just be cautious of thinking it could integrate with other 'LoRa' devices or networks.  Some other web sites that mention this system describe the radio link as 'proprietary'.

Kyler

Antoineede they are a mesh style of camera, one links to the other and then send pictures back to the home unit where you either send them via cellular or you check the sd card. The cover Lora and cuddielink cameras do this but they play hell on battieries.

I had a cuddelink system and got rid of it , the home unit was to hook up to a pc and then from there you could easily wept a scrip to send to txt message or email etc but they scrapped that idea 

I can share some practical perspective on the LoRa camera trap architecture for remote high-relief terrain with poor connectivity.

The core concept — cameras not connected to network, base station at a connectivity point relaying via LoRa — is sound and well-proven. A few things to consider for your specific conditions:

On LoRa range in strong relief — this is where the technology shines and where it disappoints unpredictably. In open terrain, 5-15km gateway range is achievable. In steep valleys or dense canopy, a node in a gully might only reach 200-300m. The solution is careful gateway placement on ridgelines or elevated points, and in complex terrain, one or two dedicated relay nodes at intermediate heights. Test before committing to a layout.

On reliability in heavy rain — LoRa itself is very robust in rain (the signal is largely unaffected by precipitation). The vulnerability is the hardware: connectors, antenna connections, and enclosures. For the gateway, use N-type or SMA connectors with proper weatherproofing, and position the enclosure under a simple rain shield. Cheap LoRa modules with U.FL connectors are more vulnerable — consider a fiberglass enclosed gateway with a proper outdoor antenna.

On the commercial options you mentioned — the Covert LoRa uses a proprietary LoRa implementation that requires their own base station, not standard LoRaWAN. This limits flexibility significantly. If you want to integrate with open platforms like The Things Network or ChirpStack and use standard sensors alongside the cameras, a system built on standard LoRaWAN is more future-proof even if it requires more initial setup.

Happy to discuss specific gateway options or architecture for your terrain.

Congratulations. The thermal images look great!

Woah!! Amazing videos. Super cool project!

That's a great idea @MandyEyrich ! Similar to your idea, I wrote up an article with geospatial highlights from ICTC 2026: https://wildlabs.net/en/article/wildlabs-geospatial-group-ictc-2026. 

Is the Human-Wildlife Coexistence article available yet? Would love to read it and share it with colleagues at Fauna & Flora.

What about this year? Who will be there? 

https://www.naturetechweek.com/

I am planning to be there for the Unconference and some satellite events.

I've looked into adding external triggers to camera traps.  I've documented that effort here.  Basically, it involves board level work to hijack the trigger signal.  But as this signal is open drain, it's straightforward to wire-OR several of these signals from multiple cameras.  In your case, you can perform this OR operation using simple wireless units.

I'm afraid I don't see a way to abstract and extract the trigger functionality cleanly into a drop-in product.  Perhaps the best that can be done is to convert all participating cameras into slave units by replacing the IR sensor with a connector to which a master triggering source is attached.  This still requires individual board work, but is at least straightforward.

This looks amazing and I look forward to trying it out when I get the chance! 

Just wondering, when it comes to the AI recognitions, is there a way to "rename" the recognitions that were incorrect?