Utilization of textured stereovision for registration of 3D models of objects

Abstract

RGB-D sensors triggered a rapid progress in the field of robot visual perception. A typical visual perception subsystem relies on finding the correspondences between features extracted from RGB-D images retrieved from robot sensors and models of objects. In this paper we introduce a multi-camera setup supplemented with an additional pattern projector used for registration of high-resolution images of objects. The objects are placed on a fiducial board with two dot patterns enabling robust extraction of masks of the placed objects and estimation of their initial poses. The acquired dense point clouds constituting subsequent object views undergo pairwise registration and at the end are optimized with a graph-based technique derived from SLAM. The combination of all those elements resulted in a system for generation of consistent 3D models of objects. We present details of the developed system and conclude the paper with discussion of the achieved results.