CloudCompare forum

Hello,


I have a question of general interest, and not that much cloud compare related.
But since I am using the ICP implementation on CC, I am asking for an advice here:

So, I have 2 datasets of 2 different Lidar flights over a forest.
I have exctracted the points that correspond to the tree tops, and I want to use them for the coregistration of the 2 different flight.
I have tried different registration methods and all seem to fail, even though I was expecting ICP to give reasonable results.

Do you think that maybe the reason is the fact that the 2 datasets do not have a uniform misaligment(see screenshot attached) and thus its impossible to extract the rotation and translation matrices for all the points?

Thanks in advance for any interesting opinion!

Kalliopi

ICP will rely on the nearest neighbors to match points. So considering your ultra low density, you'd better have as many points in both clouds, and that they are not too close to other points. You also need to be sure that the points represent the same physcical point in real world (ICP assumes the two clouds represent the same shape, and it will look for a rigid transformation between the two).

Hi Daniel,

Thanks a lot for your reply.

My main question is basically if the 2 datasets need to have a uniform misaligment through the whole point cloud,
for helping the ICP to find a global rotation and trnslation matrix.
Unofrtunately as you see, this is not the case in my tree tops points..

Well you don't "need" the two clouds to be exactly the same, but in this case the transformation matrix will simply minimize the error over all the points, and you'll end halfway.

I chose the points that are not further than 1m from each other, and I rerun the ICP.
It minimizes the error, but there is till a shift of 30 cm.

Yeah, that's definitely possible if the distances can't be minimized more than that (with a rigid registration matrix).

Yes.. My main question though is if the misalignment within the 2 PCs should be uniform.
Otherwise if the signal is not uniform then no matter how close the points are or how high my density is,
the ICP or any other registration algorithm will fail in extracting a uniform tranformation matrix, right?