geometry reconstruction from multiple images

Hi,

I am interested in geometry reconstruction using multiple high resolution camera images from different points in space, the application is for aerial mapping of landscape terrain.

The images are taken from a camera at a recorded GPS coordinate and camera XYZ angle, based on this data how feasible is it to create a relative height map of the landscape below assuming there are many high resolution images taken from various positions, all with the camera pointing relatively straight down?

I would like to generate a relative height map this way and then map that onto known height points to create a realistic absolute heightmap.

cheers, Jamie

Reply to
Jamie M
Loading thread data ...

It's possible, assuming your accuracy requirements aren't too high, but the programming task is far from trivial. For one thing, if the images aren't made at the same time, the shadows will be different.

Cheers

Phil Hobbs

--
Dr Philip C D Hobbs 
Principal Consultant 
 Click to see the full signature
Reply to
Phil Hobbs

I can imagine sliding images across one another and doing 2D cross-correlations to determine relative height in any small zone. But I don't know how to get absolute height, given that the camera angle could change between shots. The hybrid, relative adjusted for a known reference point, might work.

Or use Google Earth... if it still works on your PC.

--
John Larkin         Highland Technology, Inc 
picosecond timing   precision measurement  
 Click to see the full signature
Reply to
John Larkin

THe classic way of doing this is by stereo imaging. Has been used since at least before WW2, all contour mapping was originally done this way - by hand ! At least that way you know precisely the image separation, and the time interval ( if any ) between images.

--
Regards, 

Adrian Jansen           adrianjansen at internode dot on dot net 
 Click to see the full signature
Reply to
Adrian Jansen

If you have an actual airplane, as opposed to a webcam on a drone, you have a lot of advantages. Simultaneous sampling is important in so many situations. ;)

Cheers

Phil Hobbs

--
Dr Philip C D Hobbs 
Principal Consultant 
 Click to see the full signature
Reply to
Phil Hobbs

Hi Jamie,

Here's a reference I found useful on a similar project:

Multiple View Geometry in computer vision by Hartley and Zisserman

formatting link

It's spendy, but good.

If you Google photogrammetry, you'll find lots of useful references.

ChesterW

Reply to
ChesterW

from memory, the stereo images are generated with a simultaneous exposure taken with a split lense [is that the name?] camera system which has a separatedd view, something like 3 ft apart? or more. Then the two images are developed and if you're good you can simply look at them and see in stereo, cross your eyes and see inverted 3D, orput into a stereo viewer that keeps you from going woky while looking at them.

Reply to
RobertMacy

This article should convince you to forget it.

formatting link

Reply to
bloggs.fredbloggs.fred

This is one of those things where if it exists, it is an imagej plugin.

formatting link

Reply to
miso

Sometimes they also had side looking lenses, projecting the image of the horizon to the edges of the film. That way, you could later on determine the actual attitude of the plane, when the picture was taken.

These days various accelerometers and gyros are used to determine the instant attitude.

Reply to
upsidedown

beforem the invention od SAR that was how contour lines (isoheights) were drawn.

for details on the method try asking in a cartography forum.

--
umop apisdn
Reply to
Jasen Betts

If you have an aerial stereo camera, you know very accurately the relative position and aiming of the frames. Individual pictures taken from a moving car or drone require a lot more work, because the pointing and spacing aren't known very well, especially if it's just done with GPS. A few metres' error in the separation between camera positions will completely screw up the range information.

Cheers

Phil Hobbs

--
Dr Philip C D Hobbs 
Principal Consultant 
 Click to see the full signature
Reply to
Phil Hobbs

Hi,

Thanks for the link, so it's already been done! :D

Around page 14 they use a simple UAV glider, just a single camera and GPS, and it takes 3 images per second, and thousands over a 5 minute flight, and then all these images (with GPS coordinates) are fed into a supercomputer and the algorithm can output a detailed heightmap.

Ok maybe I should see if their software is opensource! :)

cheers, Jamie

Reply to
Jamie M

Hi,

The GPS spacing doesn't have to be that accurate I think when the quantity of images and high cross correlation of many images is effectively like oversampling to reduce the GPS error maybe.

The paper linked to by Fred shows a detailed heightmap was already created from a simple glider with a 3Hz camera and GPS:

formatting link

cheers, Jamie

Reply to
Jamie M

Hi,

I was reading a bit more of that paper, it looks like GPS coordinates aren't needed at all to generate a relative 3D model from unknown source input images, however it takes more processing as all images need to be compared with each other since their locations are initially unknown.

Also if some of the images have known parameters ie GPS etc, that can be transferred back into correlated images that previously didn't have the GPS coordinates too.

cheers, Jamie

Reply to
Jamie M

It has been done commercially for ages. I knew someone who worked on producing accurate maps for the first Gulf War (the justified one). The existing desert maps were simply not accurate or adequate at the time.

Automated technology for doing the analysis goes back even further as a spin off from high energy physics who digitised insane numbers of cloud chamber photographs and with modest hardware alterations could do the same for aerial photographs. The history of Laser-Scans development from a research contraption to modern digital map making is online at:

formatting link

They went bust in 2003.

Use of stereo pair aerial photographs for assessing topography of military targets dates back much further probably to WWI. The human brain is remarkably good at doing this trick in realtime.

--
Regards, 
Martin Brown
Reply to
Martin Brown

It should be open source, but the part about parallel processing on supercomputers is a major turn-off.

Reply to
bloggs.fredbloggs.fred

It shouldn't be, you can rent compute time on massive clusters of cloud servers by the second for not much money at all. it's a popular trick used by security researchers.

--
umop apisdn
Reply to
Jasen Betts

I thought security researchers just distribute the work to their botnets :)

Clifford Heath.

Reply to
Clifford Heath

ElectronDepot website is not affiliated with any of the manufacturers or service providers discussed here. All logos and trade names are the property of their respective owners.