3D measuring - Iterative closest points algorithm

[Spin69]

Hello,

Is there someone knowledgeable in the filed of 3D measuring and so on...? I have trouble programming the ICP algorithm and would really appreciate some help.

I think I understand the theory, but my implementation isn´t working at the moment..

Thank you

[altenbach]

Do you have a typical set of 2 data clouds for testing.

[Spin69]

I have created some basic 2D and 3D shapes for testing. Didn´t use the Stanford bunny or other point cloud.

[altenbach]

Well, the actual meaning of my question was if you could share a typical dataset. i.e. attach it here.

Lets start with the 2D problem (fewer degrees of freedom)> I assume you allow translation and rotation.

[Spin69]

Sure, I´ll attach them. Fix is the stationary one and Mov is to which I apply the calculated translation and rotation. Sadly the translation vector seem to be always wrong. Don´t know about the rotation matrix. I have tried to figure out the mistake, but no luck so far. The Mov file is just rotated and translated version of the first one (checked on that).

[altenbach]

Can we assume that the points of the two sets are ordered the same way?

[Spin69]

Yes, here they are in order. Though the real data will not be, so my application needs to work for both ordered and unordered clouds. I have coded a function to assign each point from Fix cloud a coresponding point in Mov cloud. So that I can minimize the error function based on euclidian distances between the corespondencees.

[altenbach]

A very primitive attempt using Levenberg-Marquardt (assuming points are ordered).

[Spin69]

Many thanks for your idea and your time spent on this. I will look into it and hopefully get something out of it for my work.

[altenbach]

It would be trivial to rewrite the model to output N distances. The Y input would then be an array of N zeroes, i.e. you want to translate/rotate until all pairwise distances are as small as possible.