Building multi-resolution volumetric 3D maps in real-time using CPU

Abstract

We propose an approach to fuse sequences of point cloud data from RGB-D cameras or other Laser sensors into one multi-resolution volumetric map. Our approach can work with one or multiple generic sensors and bases on an octree-like data structure to compress data into a memory efficient and multi-resolution map. Such map correctly models free, occupied as well as unknown space and can be suitable for robot navigation. We make use of a fast neighbour search on octrees to reach real-time registration on CPU. Our system does not require availability of a state-of-art GPU and hence can be run on systems with limited on-board computational resources like flying robots. We present our first experimental results using such approach.