Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environments

Abstract

Abstract-Accurate and reliable localization and mapping is a fundamental building block for most autonomous robots. For this purpose, we propose a novel, dense approach to laserbased mapping that operates on three-dimensional point clouds obtained from rotating laser sensors. We construct a surfel-based map and estimate the changes in the robot's pose by exploiting the projective data association between the current scan and a rendered model view from that surfel map. For detection and verification of a loop closure, we leverage the map representation to compose a virtual view of the map before a potential loop closure, which enables a more robust detection even with low overlap between the scan and the already mapped areas. Our approach is efficient and enables real-time capable registration. At the same time, it is able to detect loop closures and to perform map updates in an online fashion. Our experiments show that we are able to estimate globally consistent maps in large scale environments solely based on point cloud data.I. INTRODUCTION Most autonomous robots, including self-driving cars, must be able to reliably localize themselves, ideally by using only their own sensors without relying on external information such as GPS or other additional infrastructure placed in the environment. There has been significant advances in visionbased [6,7] and RGB-D-based [18,33,3] SLAM systems over the past few years. Most of these approaches use (semi-)dense reconstructions of the environment and exploit them for frameto-model tracking, either by jointly optimizing the map and pose estimates or by alternating pose estimation and map building [21]. Dense approaches have a prospective advantage over feature-based and sparse approaches as they use all available information and thus do not depend on reliable feature extraction or availability of such features.In contrast to these developments, current 3D laser-based mapping systems mainly accomplish the estimation relying on feature-based solutions [34,35], reduced map representations [14,13], voxel grid-based methods [16], or point sub-sampling [30], which all effectively reduce the data used for alignment. Compared to most indoor applications using RGB-D sensors, we have to tackle additional challenges in outdoor applications using 3D laser sensors, i.e., (1) fast sensor movement resulting in large displacements between scans, (2) comparably sparse point clouds, and (3) large-scale environments.In this paper, we present a dense mapping approach called Surfel-based Mapping (SuMa), which builds globally consistent maps by tracking the pose, so-called odometry, and closing loops using a pose graph as illustrated in Figure 1. To this end, we employ a surfel map to efficiently generate synthetic views for projective data association and additional

Bayesian Spatial Kernel Smoothing for Scalable Dense Semantic Mapping

“…SemanticKITTI dataset is a large-scale dataset based on KITTI odometry dataset. It provides dense annotations for each scan of sequences 00-10 including camera poses estimated from a surfel-based SLAM approach (SuMa) [43]. The input data of this dataset is collected by a Velodyne HDL-64E laser scanner.…”

3D Surface Reconstruction from Voxel-based Lidar Data

“…Other methods create local descriptions of the surface by a set of unoriented discs (-surfels-) calculated from the points distribution inside defined neighborhoods. Surfels have been used for traversability assessment [8], and more recently to perform simultaneous localization and mapping (SLAM) in outdoor urban environments [9]. Implicit methods: For other applications such as physical modeling or detailed terrain traversability, a more accurate and continuous representation might be preferred.…”

Tigris

“…Point Cloud Registration Point cloud registration pipelines generally fall under three categories depending on the density of points that are used for registration. On one extreme is algorithms that use all the points for registration [7,9,48], which tolerate outliers but are computationally prohibitive in real-time. On the other extreme are algorithms that use only (sampled) feature points [15,33], which are efficient in compute, but could suffer from local minima.…”