Modreg: A Modular Framework for RGB-D Image Acquisition and 3D Object Model Registration

Abstract

Abstract. RGB-D sensors became a standard in robotic applications requiring object recognition, such as object grasping and manipulation. A typical object recognition system relies on matching of features extracted from RGB-D images retrieved from the robot sensors with the features of the object models. In this paper we present ModReg: a system for registration of 3D models of objects. The system consists of a modular software associated with a multi-camera setup supplemented with an additional pattern projector, used for the registration of high-resolution RGB-D images. The objects are placed on a fiducial board with two dot patterns enabling 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 able to generate consistent 3D models of objects.Keywords: ModReg, textured stereo, RGB-D image, point cloud, registration Motivation of the workEmergence of new types of sensors stimulates robotics. Up to the late nineties the most popular sensors used for obstacle avoidance were sonars, whereas the commercialization of laser rangefinders (such as the Hokuyo UTM and SICK LMS series) enabled rapid development of methods for 2D robot pose estimation [21] or techniques for simultaneous localization and map building (SLAM) [38]. Similarly, the appearance of inexpensive RGB-D sensors [33,15], such as Microsoft Kinect released

Bulletin of the Polish Academy of Sciences: Technical Sciences

“…Thus it is required that the system should posses such 3D models of objects. Those models can be generated using solutions such as MODREG [40]. Models used in our experiments are publicly available as part of the WUT Visual Perception Dataset [72], examples are presented in Fig.…”