Online 3D SLAM by Registration of Large Planar Surface Segments and Closed Form Pose-Graph Relaxation

Abstract

A very fast pose-graph relaxation technique is introduced in this article for enhancing the consistency of 3D maps created by registering large planar surface patches. The surface patches are extracted from point-clouds sampled from a 3D range-sensor. The registration method offers an alternative to state of the art algorithms and provides advantages in terms of robustness, speed and storage. It especially results in accurate rotation determination, although a lack of predominant surfaces in certain directions may result in translational uncertainty in those directions. A loop-closing and relaxation problem is hence formulated which gains significant speed by relaxing only the translational errors considering the full translation covariance determined during pairwise registration. This leads to very fast 3D Simultaneous Localization and Mapping (SLAM) suited for online operations. The approach is tested in two disaster scenarios that were mapped at the NIST 2008 Response Robot Evaluation Exercise (RREE) in Disaster City, Texas, USA. The two datasets from a collapsed car park and a flooding disaster consist of 26, respectively 70 3D scans. It is shown in these experiments that the approach can generate 3D maps without motion estimates by odometry or similar and that it outperforms Iterative Closest Point (ICP) based mapping with respect to speed and robustness. published in: