Fisheye solving

[Steinar Midtskogen]

I've been trying to use astrometry.net to solve images taken with full frame fisheye lenses.  This does not work well with astrometry.net, but I've managed to work around it reasonably well, so I thought I should share my workarounds.

Astrometry.net will solve a narrow field, say about 40 degrees, around the centre of such an image.  This can be pushed higher be reprojecting the image to orthographic projection, but it will still only work for a small part of the original image and the risk of incorrect matches increases.  So what I resorted to in the end was to reproject keeping the fisheye but changing the centre of the projection and do that in 10 degree intervals over the entire field of view, solve for every reprojection, remap the results in the .corr file back to the original file and eliminate duplicates.  This is of course very slow, but it works.  Hugin/panotools offer the tools needed to reproject and map coordinates back.

Here's a video showing some results using this technique: https://www.youtube.com/watch?v=IDymepBQdoo

[Dustin Lang]

Nice!

[Jean-Paul Godard]

Did you try to use "Iris" (freeware from Christian BUIL) : see http://www.astrosurf.com/buil/iris/new530/new530_us.htm

 I did not had a try at this time.

JP

[Benjamin Hill]

Very nice stuff!  I'm working on something similar (full frame bad-at-lowlight stills from a gopro) and attempting to align and stack them.  I think you've had better luck than me, judging from our output in your youtube video, but maybe my naive star-finder algo can be of use?

[Dominic Ford]

Quick reply to add that I'm using a similar technique myself, and it works really well.

I'm working on a schools astronomy project where kids will leave cameras pointing at the sky overnight, and we use astrometry.net to figure out orientation. The kids can use the images to illustrate all sorts of bits of their astronomy and compsci curriculum, as well as hunting for meteors, planes and satellites.

But because the lenses are fish-eye, and can only solve a small rectangle in the centre of each image.

That loses me some accuracy, but I get it back by taking the average of many fits. I know each camera was static overnight, so the central declinations of all the frames should be the same, meanwhile the RAs should differ by sidereal time.

My impression is that part of the problem is that most wide-angle lenses have significant barrel distortion, and astrometry.net is very sensitive to that.

When I first start using a new lens, I solve for the barrel distortion parameters by attempting to stack the solved frames whilst measuring how well they overlay each other. If I then go back and feed the corrected wide-angle images into astrometry.net, I can often solve much wider fields quite comfortably.

If any astrometry.net team members are reading this, I'll be in touch with you shortly to ask about publicity, etc. We're happy to put your logo on stuff!

[Dustin Lang]

Cool!  Would be happy to hear more details about this project.  I've been thinking it would be fun to try to do something like this with my own kids.

--dstn

[Steinar Midtskogen]

I can add that it also helps to pick a suitable star database with no more stars than those over the limiting magnitude of your camera.  Also, if the goal is to model the lens as a Hugin .pto file, feedback from Hugin will tell you whether astrometry's solution was correct or not.

My project has progressed since last I wrote.  We've set up a rig of three cameras on one site and four cameras at another which combined cover the entire sky.  All cameras have been calibrated using astrometry and Hugin.  The purpose is to monitor the sky for meteors, and in order to calculate the trajectory, it's important to find the path across the sky with great precision.

One recent fireball was this one (the view from one single camera with a grid drawn from the Hugin .pto file): http://norskmeteornettverk.no/bilder/2015/harestua-20150428003216-grid.jpg
(this meteor missed zenith by just a few arc minutes!)

The markers for the stars serve as a verification that the camera has not shifted since it was calibrated.  The markers are drawn using the timestamp and .pto file, and marks where the stars ought to be.

Since I have a .pto file for the lens, I can reproject the image to rectilinear to straighten the path: http://norskmeteornettverk.no/bilder/2015/harestua-20150428003216-reproj.jpg

And to find the path I can get closeups of the start and end:
 http://norskmeteornettverk.no/bilder/2015/harestua-20150428003216-start.jpg
 http://norskmeteornettverk.no/bilder/2015/harestua-20150428003216-end.jpg

Using the .pto files from all four cameras at this site, I can also stitch together a full view:
 https://www.youtube.com/watch?v=8UdXy7u-Q0I

Or like this:
 https://www.youtube.com/watch?v=gN40Za4HOUc

Unlike ordinary stitched panoramas these views were stitched solving each lens individually using the stars, not by stitching overlapping sections.

And here's a live view using equirectangular and fisheye projections: http://voksenlia.net/met/solobs/

The cameras record a 2560x1920 image each at 5 frames per second.  They give me a view of the entire sky at ~3 arcminute resolution.  And since there are multiple lenses, the distortion near the edges of the fisheye view is low unlike true fisheye lenses.

-Steinar

[高宝全]

I have a question if u can help me? if u use some small rectangle in the centre of each image to stack  , and the distortion paraments how to measured?
在 2015年5月21日星期四 UTC+8下午10:02:38，Dominic Ford写道：