Towards Robotic Semantic Segmentation of Supporting Surfaces

Abstract

Perceiving the geometry of environmental structures surrounding is a crucial prerequisite for robotic understand the indoor environments autonomously. A new framework for parsing RGB-D images aimed at supporting surfaces segmentation is proposed. First, the surface normal is extracted from depth information using PCA and normal clusters with 3D mean shift clustering. Then the main planes such as floor, wall will be detected with gravity vector estimation. Finally supporting surface and its corresponding objects are segmented using graph optimization with energy functions. The approach can offer a robotic semantic segmentation for better understanding the indoor environment. The experiment results based on Berkeley 3D Object Dataset demonstrate that our framework works well on indoor RGB-D cluttered scenes.