Robust Volumetric 3-D Reconstruction in a Dynamic Environment

Abstract

As a part of a research project at SINTEF Ocean AS named Neodroid aiming at developing a general purpose humanoid robot to execute a variety of complex tasks, the need for perceptive ability arises. The robot consists of two robotic manipulators and is intended to solve tasks demonstrated by a human in a virtual environment. This study aims at developing a vision system that is capable of keeping an accurate reconstruction of the robot workspace. To achieve this, a volumetric representation method consisting of a three-dimensional grid of voxels is used. To keep this reconstruction up to date, the changing regions of the volume need to be scanned. This is achieved by performing novelty detection with RDE1 on depth images from a camera viewing the entire volume in a top-down configuration. The novel regions are then reconstructed in three dimensions and clustering is performed to draw bounding boxes around these. Camera viewpoints are generated on a circular path around each novel region to obtain an accurate measurement.