Global Localization of Vehicles Using Local Pole Patterns

Accurate and reliable localization is an important requirement for autonomous driving. This paper investigates an asymmetric model for global mapping and localization in large outdoor scenes. In the first stage, a mobile mapping van scans the street environment in full 3D, using high accuracy and high resolution sensors. From this raw data, local descriptors are extracted in an offline process and stored in a global map. In the second stage, vehicles, equipped with simple, inaccurate sensors are assumed to be able to recover part of these descriptors which allows them to determine their global position. The focus of this paper is on the investigation of local pole patterns. A descriptor is proposed which is tolerant with regard to missing data, and performance and scalability are considered. For the experiments, a large, dense outdoor LiDAR scan with a total length of 21.7 km is used.