Skip to content

Anyone Using Thermal Imaging Cameras?

edited July 6 in SecuritySpy

Is anyone using or have tested a worthy thermal imaging security camera with SecuritySpy? The last discussion on this topic is a few years old.

I installed one of the first Axis thermal imaging cameras for a friend in 2014 and was unimpressed. It was way, way out of my price range and with a resolution of just 160x128, it just didn't seem worth the trouble unless I was running a witness protection safe house. However, I have recently seen the prices come way down on Ebay and with the resolution up to 640x480 that is AI pixel refined. Thermal imaging would be a great addition to my army of cameras living in a remote rural area, but I am still on the fence about it all. UGH, but it would indeed help me with a coyote verses livestock issue just for the fact that the predators would be very easily identified in brush verses any of the night vision enhancement technologies available (such as Axis Lightfinder 2.0, which is excellent otherwise).

So, I would be interested in any experiences or other knowledge that anyone can offer. Like all other geeky bells and whistles, it is only a matter of time before you see the CIA (MI6 for you, Ben) level stuff on sale at your local discount store.


Peace,

Dr. Z.

Comments

  • edited July 15

    Update:

    I was able to purchase an Axis Q1941-E thermal imaging bullet-style camera at a surprisingly affordable price (considering the ridiculous retail price tag). Yup, it is a discontinued 2019 model, but brand new out of the box. While I have owned a basic handheld FLIR thermal imaging unit for general maintenance diagnosis over the last six years and wouldn’t be without it now, the Axis unit is my first personal security camera application of this type. The 28 degree field of view and 13mm lens limits what can be monitored on a 12 acre ranch. Axis does make a pan/tilt motor accessory for some of their units and software drivers for web browser control, but not this particular model. I could probably hack together something if I really need that functionality. As usual though, if you want something in technology to work seamlessly and actually be useful, it takes money that I can’t justify spending just yet.

    Anyhow, the first nights test found a few wildlife guests on my fenced in property. Detection is as far away as 800 feet, yet recognition starts around 400 feet. My primary concern is in predator tracking against my livestock; coyote mostly, but mountain lions also a few times a year. The trick will be to tweak the AI animal motion sensing of SecuritySpy. Axis analytics only track humans and vehicles inside the camera. Unfortunately the 768 x 576 upscaled pixel resolution of the camera and amount of area to cover limits the efficiency of this currently. It will just have to take some trial and error to finally find a central camera mounting spot for the investment to be useful 24/7.

    I have tested and/or installed a security camera from just about every manufacturer over the last 25 years personally and professionally. My first exposure to a thermal camera was as a firefighter in the 2000’s and a test program to use such technology right alongside my axe. It made the job much safer and was directly responsible for finding victims so quickly in otherwise blind smoke conditions. Now thermal imaging units are found on every fire engine in the USA. While thermal imaging may only appeal to a sliver of the average consumers now, whether it be on a sunny day or no moon foggy night, I expect that this technology will also become a standard in security camera features as the price comes down. Ben has already shown his willingness to adapt SecuritySpy features as the need arises.


    Peace,

    Dr. Z.



  • TSITSI
    edited July 24

    I am using thermal cameras. They will produce two feeds, one thermal, and one visible. Animals and people produce very defined images on the thermal as shown in the post above and the images below. I wonder if the thermal channel could be used to trigger the SS motion detection, being able to leverage the AI to discern differences between animals, cars and people. This channel could then be set to trigger the visible camera stream recording as well as other actions. Think of it as as Intelligent PIR. Regular PIR does not work well outdoors because it does not know the difference between people and pets. I plan on putting this idea to the test in the near future.


    The attached images show Foxes playing in my front yard. One is a juvenile fox. In the video they roll around on the grass and play.




  • Very interesting to read about both your experiences with these cameras!

    Thermal cameras are still rare, and SecuritySpy doesn't have any specific features for them. In particular, I doubt whether the AI object detection will be accurate for thermal cameras. This is because there are no thermal camera images in our training dataset that we use to train our AI algorithms, due to their rarity. Also, thermal images are low resolution and somewhat fuzzy and lacking in detail. For high AI detection accuracy you instead need high-resolution, clear images. So my suggestion would be that standard pixel based motion detection with the AI turned off would be the only reliable configuration, even if this does result in some false-positive detections.

    We'll definitely keep an eye out for the future of this technology, but until the image quality improves, and the products dramatically reduce in price, I don't see them being widely adopted.

  • Ben, what framework are you using for your AI models? Is the Apple CoreML stuff useful in your Mac dev environment, or are you using something else?

  • BenBen
    edited August 12

    Hi @photonclock we use TensorFlow to train our model, on hundreds of thousands of real CCTV images, and then we convert the resulting model into a form that is compatible with CoreML. SecuritySpy then uses CoreML to execute the model locally on the Mac that it is running on, allowing it to take full advantage of any hardware acceleration available (e.g. Neural Engine) for the fastest possible execution times (~2ms per classification on Apple Silicon Macs).

  • At the moment I'm running a a FLIR II "Marine" that I got lucky enough a few years ago to get at a bargain (relatively...) from a boat builder on the east coast. It's a 320x240 res. the PTZ board stopped working way out of warranty, but the IR cam inside the housing works well, so it's now a "stationary" cam. Also own a Flir E8 for general "techie" stuff of all sorts. FLIR imagers are still very expensive, and the lenses are usually made of Germanium, which can pass the long wave IR. Here's a pic of a "Kitah"....


  • Ben, have you considered a system that would allow local AI learning, where the end user can "teach" the AI the correct classifications? For example if SecuritySpy mistakes a big dog for a person, the user can teach the AI not to make that mistake again. Even individual human faces could be taught in such a fashion I think, so that, for example, the user could get a notification only if a stranger is detected in the home. (Back to my hope for Security Spy completely replacing old fashioned home alarm systems.) I could imagine that such a system, if fed back into Security Spy's built in AI system, could refine that as well over time with thousands of users collectively fine tuning the system.

Sign In or Register to comment.