3D Object Modeling and Segmentation Using Image Edge Points in Cluttered Environments

摘要

Object models are indispensable for robots to recognize objects when conducting tasks. This paper proposes a method of creating object models from images captured in real environments using a monocular camera. In our framework, an object model consists of a 3D model composed of 3D points reconstructed from image edge points and 2D models composed of image edge points, each having a SIFT descriptor for object recognition. To address the difficulty in creating object models of separating objects from background clutter, we separate the object of interest by finding edge points which cooccur in images with different backgrounds. We employ supervised and unsupervised schemes to provide training images for segmentation. Experimental results demonstrated that detailed 3D object models are successfully separated and created.