Authors:
Natálie Vítová
;
Jakub Frank
and
Libor Váša
Affiliation:
Department of Computer Science And Engineering, Faculty of Applied Sciences, University of West Bohemia, Univerzitní 8, 301 00 Plzeň, Czech Republic
Keyword(s):
Registration, Point Cloud, Dynamic, Series, Kinect, Sensor, Alignment.
Abstract:
Surface registration is a well-studied problem in computer graphics and triangle mesh processing. A plethora of approaches exists that align a partial 3D view of a surface to another, which is a central task in 3D scanning, where usually each scan only provides partial information about the shape of the scanned object due to occlusion. In this paper, we address a slightly different problem: a pair of depth cameras is observing a dynamic scene, each providing a sequence of partial scans. The scanning devices are assumed to remain in a constant relative position throughout the process, and therefore there exists a single rigid transformation that aligns the two sequences of partial meshes. Our objective is to find this transformation based on the data alone, i.e. without using any specialized calibration tools. This problem can be approached as a set of static mesh registration problems; however, such an interpretation leads to problems when enforcing a single global solution. We show
that an appropriate modification of a previously proposed consensus-based registration algorithm is a more viable solution that exploits information from all the frames simultaneously and naturally leads to a single global solution.
(More)