Photogrammetry

I have just started playing with photogrammetry which is where you can build 3D models from multiple 2D images. I have known about this method for years, but I had not realised quite how advanced (easy) it had got. I thought it might be of interest to others here.

I don't have a CNC or 3d printer, but my neighbour has a large CNC for wood, and my objective is to copy a small bass relief in brass to a larger wooden version.

I am using the open-source project Meshroom which I am quite impressed with.

I bought a hand held 3D scanner which works well for large objects, scanned my girlfriends daughters head then printed out. I also have a desktop laser scanner for small objects, higher resolution scans again works fine.

I haven’t had much success with photogrammetry, probably need more practice and I’m sure software has improved since my last attempts

I shall be following your exploits with great interest, Adrian

MichaelG.

Hi there, all,

I took an interest in this subject some (many) years ago. I borrowed a book entitled just 'Photogrammetry' from my local library. I don't remember the author's name or the publisher, sorry.

I do remember one example from the book - a single photograph of a North Sea oil rig, taken from a helicopter on a sunny day. With just that photograph and knowing the time (and date? ) it was taken, photogrammetric techniques enabled the orientation of the rig to be determined with impressive accuracy.

My interest in the subject did not last - the mathematics made my head hurt!!!!

Best regards,

Swarf, Mostly!

Very much used by Surveyours - I developed an autonomous electric UAV carrying a 40Mpixel stills camera which is used by a number of instances here in Namibia for this purpose. Land surveyors save hundreds of hours of mark-out and walkabout time manually mapping terrain by these techniques. The local Uranium mines also use the method extensively, on a monthly basis, to determine stockpile volumes and pit removed volumes.

The UAV is pre-programmed to fly a ladder pattern up and down over the terrain taking photos with a 50% forward and a 70% side to side overlap - MANY photos...These are then processed by software - AgiSoft being a very popular package - and a complete 3D image is generated. This can be rotated in space and you can see the detail down gorges, crevices, down the side walls of crevices, etc. The software has many tools allowing determination of volume of entities ( stock piles of crushed rock..) and adaptable, very accurate estimations of mass, which are compared and updated with the actual mass measurements of truck loads of carted material - the trucks have weight sensors on there axels, with Blue-tooth connection to the Load site exits...

Also used here to lay a water supply pipeline, 40km long, first surveying the terrain with the UAV.

In all cases a number of photo-visible markers ( black and white targets) need to be laid in the area, with accurate GPS position in XYZ, so that the 'stitched' image can be tied down to real earth plane...

About 6000 40Mpixel photos in a 1 hour flight, for a 6x6km area @ 400meters AGL

Fun stuff..

( example of a photo-target against the bakkie's wheel in the bottom photo below)

Wow that is impressive stuff, I don't think mine will be half as impressive. Hopefully, today will be a good day for the lighting. So far it is a nice light overcast day, so far I have had shadows or varying light levels due to clouds.

Given some photos of (say) a Tinker tool & cutter grinder, it should be possible to reconstruct the plans?

It could be useful. There is no scale to the 3d model so you would either need to know one of the original dimensions or add something to the scene of known size before photographing. Also, the accuracy is only as good at the image resolution and number of photos. So I would not expect to be able to measure high accuracy. I could see it being great for castings. building the 3D model then printing it as a lost PLA form.

A lot depends on the software used to stitch it all together. There are many packages that do this very well. As I indicated, for land survey use with Agisoft we required 4 accurate GPS coordinates (XYZ) minimum per square km.

The long pipeline we did had 40 positions with targets - a pair of targets place 500meters from the centerline of the 'pipe' , left and right, every 2 km. If we turn off those positions during the stitch process, the 3D 40 long image 800meters wide, did 7 360deg twists about the pipe centerline! Looked like a DNA sequence...The closer to the flight centerline those targets are, the easier the model will twist.

To do physical models in a studio we used a turntable with many accurately marked spots on it, place the camera on a arced beam over the turntable, taking photos as the turntable rotated 360 deg in maybe 10 to 20 degree increments, then move the camera to a new angle, and repeat ( the camera moves in an arc over the model, so the lens is always the same distance from the model center. For a model around 1/2 cubic meter in volume , about 600 to 800 40Mpixel photos. Then in Agisoft we enter the camera line of sight distance ( focal length to model), the camera XYZ position per photo sequence ( actually in polar form), and the turntable angle - all these are referenced to each photo - and if you do all that, you can get 3D model resolution down to 0.1mm over a 1/2 cubic meter model..

Lens distortion begins to be the limit here..

Agisoft took 7 hours to process 800 photos..

So it is a very viable way of developing a 3D copy of an object - Plans will however not pop out of the model...

Joe