3D point cloud registration based on planar surfaces

This paper focuses on fast 3D point cloud registration in cluttered urban environments. There are three main steps in the pipeline: Firstly a fast region growing planar segmentation algorithm is employed to extract the planar surfaces. Then the area of each planar patch is calculated using the image-like structure of organized point cloud. In the last step, the registration is defined as a correlation problem, a novel search algorithm which combines heuristic search with pruning using geometry consistency is utilized to find the global optimal solution in a subset of SO(3) ∪ R³, and the transformation is refined using weighted least squares after finding the solution. Since all possible transformations are traversed, no prior pose estimation from other sensors such as odometry or IMU is needed, makeing it robust and can deal with big rotations.