One of the simple tricks in getting high performance for a global 3D point cloud registration is combining "3D feature matching + ICP". For example FPFH matching + RANSAC for initial pose estimation and ICP for pose refinement.
Does anyone know a good reference paper for this method? I am referring to this method in my paper but don't know who first proposed it.
Have you reviewed this article?
An Explicit Loop Closing Technique for 6D SLAM
It consists of building the 3D model of environment with heuristic loop closure using ICP and reliable feature extraction.
I'm not sure if this is the paper where the method was first proposed, but the 1992 paper A Method for Registration of 3-D Shapes by Best and McKay (published in IEEE Transactions on Pattern Analysis and Machine Intelligence 14(2):239-256 · March 1992) does describe a method of the kind you are looking for.
More importantly from your perspective, it includes a fairly detailed review of the literature at that date.
