Object Recognition and Pose Estimation from RGB-D Data Using Active Sensing

Abstract

Factory automation has been growing rapidly in production sites. On the other hand, it is mentioned that inspection work is still often done manually in factories where many dangers lurk. Although, safety can be ensured by substituting robot automation for many tasks that require dexterity such as opening and closing valves. Therefore, more accurate object recognition and pose estimation of the target object is required. In this paper, the task of automatically opening and closing a globe valve using a robot arm from different angles is considered. A system was developed and introduced to make a precise object recognition and pose estimation using an RGB-D camera attached to the robot arm. The system consists of three main sub-systems for object recognition, pose estimation, and arm-control. By using an RGB-D camera, it was possible to make color-based segmentation of the globe valve to the detect valve area using the Color Signature Of Histograms of Orientations (CSHOT) algorithm by detecting features. Then apply singular value decomposition to the precise position of the valve. By considering the angle of the robot arm to the globe valve, from 0° to 30° was taken as the front mode, and from 30° to 90° was taken as the side mode. Detecting the objects from the side is novel in this research area. RANdom SAmple Consensus (RANSAC) algorithm was used to verify the pose estimation of the front mode of the valve. Centroid connection and Pose Integration Hough-Voting methods were used to verify the pose estimation of the side mode. From the results, it can be concluded that the system is accurately detecting object recognition and pose estimation of a globe valve from the side mode to the front mode.