FLoW-Vision: Object Recognition and Pose Estimation System based on Three-Dimensional (3D) Computer Vision

Abstract

This paper presents a three-dimensional computer vision-based object recognition on FLoW-Vision in RoISC (formerly ER2C) has entered its second phase. Previously, the robot had a basic vision that was used to replicate ‘human-like’ visual skills using 2D computer vision. As a result of the above discussion, we proposed the design and implementation of an object recognition and pose estimation system based on three-dimensional computer vision to handle object recognition and pose estimation tasks in real-world environments simultaneously. In the object recognition process, a point-cloud segmentation method is used to obtain possible object clusters before starting the calculation of feature description. Then, a keypoints-based two-stage matching process is performed to speed up the computation of finding correspondences between the object clusters of the current scene and a colored point cloud model of an object. Next, a Hough voting algorithm is employed to filter out matching errors in the correspondence set and estimate the initial 3D pose of the object. Last process process the pose estimation from clustered object using RANSAC to search the largest surface as Z surface. Experimental validate the object recognition can work correctly with percentage 100% and pose estimation accuracy of the proposed system can work correctly with percentage 60% in a complex real-world scene.