I'd been trying to get my hands on some of those sensors but was having a heck of a time with it, how much do they run?

(I originally sent this to you through KS, but then realized that the input of the HaD crew could be *very* helpful to you.)

(unless on a tripod), you can use superresolution techniques to increase
the (resulting) resolution of the thermal image using the constantly
changing position of the regular camera images. (If your unit contains
accelerometers/etc., then the calculations will take even less time, as
the software will already know what direction to start looking for image
matches. As well, motion blur can be taken into account.) Even
single-image SR techniques can quadruple the resolution of an image (http://www.wisdom.weizmann.ac.il/~vision/SingleImageSR.html), so combining the results of multiple image SR with single-image SR, you could potentially approach 1MP thermal images!

Here are some good links to get them started:

http://photoacute.com/tech/superresolution_faq.html

https://www.google.ca/search?q=github+super+resolution

http://www.deconvolve.net/bialith/Research/BAR_SuperRes.htm (software used: 

http://www.bialith.com )

Thanks Keith, 

superresolution can be a great method to increase resolution and these are some excellent resources.  The practical limit is about the square root of 2 in each dimension, so 1.4 times the pixel count. We could expect to get a real resolution of 115x87, but our frame rate would drop by at least 4. 

Note I haven't done any serious work with SR, just read a few MRI SR papers and follow it a little. I wasn't familiary with the single image SR, which increases the apparent resolution and it produces very nice images in many cases of typical high-contrast images, but it seems it doesn't actually increase resolution. It makes choices based on local image texture and the probability of finding a given texture. If you look at some of the examples, you can see "cartoonish-ness" in some of the fainter parts of the image, such as clouds and palm trees (like a median+bilateral filter). Honestly, its not bad, it gives really good results, despite these smallish image recon artifacts (any image processing algorithm produces _some_ artifact). Thermal images have lower contrast than these which could reduce the benefit, but given how useful it could be, it is definitely something to try out. Thank you again Keith. Have you gotten the single-SR or the accelerometer/multiple-SR methods working?

Erik

Most SR solutions can only depend on the relatively low XY counts of their own pixels for position change accuracy.  You, on the other hand, have a 5MP secondary camera to base it on, which gives you ~30 visual pixels per thermal pixel.  If, for example, you determine that a second *visual* frame is shifted 15 pixels to the right, then you know that all *thermal* pixel results are 50/50 split over the last ones.  The only calculations needed are to figure
out the offset for each succeeding frame, so you know what thermal value to assign to each visual pixel.  (I'm not saying that you could get a 100% 5MP matchup, but it would be close enough for any application that I could think of.  It would also have a great deal of difficulty with non-static and high depth-of-field situations, but that would be true of *any* solution.  A second visual camera on the other side of the thermal imager would allow depth information to be gathered, which would allow more accurate placement of thermal data within the recorded 3D space.)

I personally haven't used any of the programs that I linked to in my previous post, but I have used the programs that I linked to here: https://deadlydad.wordpress.com/2012/11/15/image-rescaling-utilities/

Keith, I actually remember reading your blog post about a year and a half ago when I looked into pixel scalers. These are potentially very useful and I'll look into them!

Very interesting stuff. I really appreciate seeing some of the development process.

Thanks suicidal.banana!

Cool, nobody can deny the need for more affordable thermal cams :) your KS link is broken btw, remove the dot at the end of it ;)