A descriptor and voting scheme for fast 3D self-localization in man-made environments

In contrast to the increasing availability of affordable and lightweight 3D sensors, navigation sensors are still big, expensive (IMU), and prone to GPS errors. In view of a lightweight, affordable and robust 3D mapping solution, it is preferable to aim at a low-cost IMU and GPS-less system. Therefore, some capabilities provided by navigation hardware should be replaced by methodical solutions. We present an approach for data-based self-localization in a large-scale 3D model of a man-made environment (e.g., an urban area, an indoor environment), which solves substantial parts of this problem. Our approach uses a rotation invariant descriptor and a 3D voting scheme to determine the own position and orientation within available 3D data of the environment. While our methods can support loop closing during mapping, the main result is the ability for fast and GPS-less initial self-localization.