I am currently trying to implement a GraphSLAM/SAM algorithm for LIDAR. From papers I've read, you generate a directed graph from expected LIDAR measurements to landmarks similar to the image below (taken from the Square Root SLAM Paper by Dellaert).
However in practice the point clouds you obtain from LIDAR are similar to this (taken from the KITTI car collected dataset):
It seems algorithms such as SIFT for 3D point clouds aren't as accurate yet. Is there a commonly used technique to efficiently find features in consecutive point clouds to find landmarks for SLAM algorithms without using >30,000 points in a point cloud?