Partial view modeling and validation in 3D laser scans for grasping

Abstract

Humanoid robots performing every day tasks in human environments need a strong perception system in order to operate successfully. As 3D data acquisition devices like laser scanners and time of flight cameras get better and cheaper, we expect three-dimensional perception to become more important. We describe a new method for detecting surfaces of revolution in point clouds within our Sample Consensus Framework. Cylinders, cones and arbitrary rotational surfaces can be reliably and efficiently detected. Symmetry assumptions can be hypothesized and verified in order to complete the model from a single view, i.e. to generate data on the occluded parts of the object. These complete models can be used for grasp analysis. Additionally, we propose a new method for scoring models within the Sample Consensus Framework in order to get better shapes.