Geometric features for robust registration of point clouds

Abstract

Several feature detectors for 3D point data have been proposed in the literature. They have been applied to various problems in computer vision and robotics. We use them to solve two fundamental problems in real-time robotics, namely the registration of laser scans as well as the detection of loops and places. We extend and modify existing feature detectors, combine them in a smart way and create a system, that solves these problems efficiently and better that existing other solutions. We evaluate our system with data sets provided by other groups as well as our own data and we compare our results to those obtained with other algorithms.