Generic Fitted Primitives (GFP): Towards Full Object Volumetric Reconstruction for Service Robotics

Abstract

Service robotics applications, such as mobile manipulation in domestic environments, require 3D representations of the objects of interest to be grasped. Simple object recognition or segmentation cannot provide structural shape information mandatory for obtaining reliable grasp configurations. In this paper, the Generic Fitted Primitives (GFP) technique for volumetric reconstruction is introduced. The goal of the method is to build full 3D object shapes from a single camera perspective. In order to obtain the shape of the 3D primitive, we propose an energyminimization algorithm based on the concept of Active Contours applied directly on 3D visual data. Our modeling approach produces compact closed surfaces (volumes) describing the objects of interest which can be further used for service robotics tasks, such as grasping or manipulation. The performance of the proposed technique has been evaluated against two different methods, i.e. generalized active contours and superquadric approximations.