Incremental surface extraction from sparse structure-from-motion point clouds

Abstract

Extracting surfaces from a sparse 3D point cloud in real-time can be beneficial for many applications that are based on Simultaneous Localization and Mapping (SLAM) like occlusion handling or path planning. However, this is a complex task since the sparse point cloud is noisy, irregularly sampled and growing over time. In this paper, we propose a new method based on an optimal labeling of an incrementally reconstructed tetrahedralized point cloud. We propose a new sub-modular energy function that extracts the surfaces with the same accuracy as state-of-the-art with reduced computation time. Furthermore, our energy function can be easily adapted to additional 3D points and incrementally minimized using the dynamic graph cut in an efficient manner. In such a way, we are able to integrate several hundreds of 3D points per second while being largely independent from the overall scene size and therefore our novel method is suited for real-time SLAM applications.