Authors:
Simon Thompson
;
Masahi Yokozuka
and
Naohisa Hashimoto
Affiliation:
Smart Mobility Research Group, Robot Innovation Research Center, National Institute of Advanced Industrial Science and Technology, Tsukuba and Japan
Keyword(s):
Point-Cloud Registration, Inconsistent Maps.
Related
Ontology
Subjects/Areas/Topics:
Informatics in Control, Automation and Robotics
;
Intelligent Transportation Technologies and Systems
;
Mobile Robots and Autonomous Systems
;
Robotics and Automation
Abstract:
Accurate point cloud registration techniques such as Iterative Closest Point matching have been developed to produce large scale 3D maps of the environment. Typically they iteratively register point clouds captured from adjacent sensor scans resulting in point clouds which are largely consistent. However, merging two seperate point cloud maps constructed at different times can lead to significant inconsistencies between the point clouds. Existing point based registration techniques can be sensitive to local minima caused by such inconsistencies. Feature based approaches can overcome local minimum but are typically less accurate, and can still suffer from correspondence errors. We introduce Large Scale Persistent Features (LSPFs), sub regions of point clouds that have orthogonal planar regions that are consistent and persist over a large spatial area. Each LSPF is used to calculate an individual transformation estimate using traditional registration techniques. Sampling Consensus is t
hen used to select the best transform which is used for registration, avoiding local minima. LSPF registration is applied to simulated point cloud maps with known inconsistencies and shown to perform with more accuracy and lower computation time than other popular approaches. In addition, real world registration results are presented which demonstrate LSPF registration between MMS maps and low cost sensor maps captured 6 months apart.
(More)