Dear colleagues,
I am working with Point Cloud alignment and I would like to know, based on your experience, which was the best alignment algorithm suitable in your application:
* ICP (Iterative Closest Point);
* CPD (Coherent Point Drift);
* NDT (Normal-Distributions Transform).
Also, good practices as filtering and other methods related would be very appreciated.
Thank you very much.
Kind regards
In my experience, ICP is the first go-to method for any registration related problem as it works "decently" in most situations. Often filtering outliers, having multiple initial start conditions and random sub-sampling of points in conjunction with ICP would give good results. Go-ICP is great if you want globally optimal solutions, however, in my experience Mixed integer programming methods offer better solution than Go-ICP, albeit at greater computation cost.
If you know that the problem you are dealing with is a deformable registration, then ICP's counterpart first go-to method is CPD. It is easy ti implement and gives good results out of the box provided you filter out obvious outliers and try multiple initial starts. If you have noise in the data, which can be characterized well enough then iterative most likely method (IMLP) is quite good. There are variations of IMLP that do deformable registration as well. Of course all of this assumes you want to use a conventional method, otherwise deep learning techniques such as PointNet-LK, PCRNet, DCP, Deep global registration, PR-Net, RPM-Net are good choices.
You may try ICP. https://www.youtube.com/watch?v=djnd502836w
https://www.codeproject.com/Articles/865830/Point-cloud-alignment-ICP-methods-compared
Arun Srivatsan Rangaprasad , thanks for you answer, actually my data is almost free of outliers. I believe my main problem is that not all of my point cloud has a correspondent point in both point cloud. I am wondering if "Point-to-Plane" approach wold solve it. For now I would like to find a deterministic approach rather than deep learning.
Vicky Barua, this work was very useful until now, thanks a lot. Chapter 2 was particularly interesting for understanding concepts.
Mahmudul Hasan, I did see the full video yet, but this professor have a very interesting didactic. Also, he has a lot of short-videos of complex or not so well explored concepts. Thank you!
Thiago Neves Cunha : Aah ok, if that is the case, then here is one recipe to solve the problem:
1) Generate a mesh from the point cloud ( You could use something simple like Delaunay triangulation)
2) Use point-to-plane correspondences from the source points to the template mesh. This operation will essentially project the point onto the faces of the mesh and find the correspondence.
3) Use the projected point correspondence to find the registration.
Here are three papers that may be useful. They use simple conventional techniques, no deep learning:
1) Low, Kok-Lim. "Linear least-squares optimization for point-to-plane icp surface registration." Chapel Hill, University of North Carolina 4.10 (2004): 1-3.
2) Billings, Seth D., Emad M. Boctor, and Russell H. Taylor. "Iterative most-likely point registration (IMLP): A robust algorithm for computing optimal shape alignment." PloS one10.3 (2015): e0117688.
3) Segal, Aleksandr, Dirk Haehnel, and Sebastian Thrun. "Generalized-icp." Robotics: science and systems. Vol. 2. No. 4. 2009.
- Badges
- Science topic