Detection of Feature Areas for Map-based Localization Using LiDAR Descriptors

Abstract

Map-based localization is an essential challenge for the development of autonomous vehicles. Popular localization solutions depend on static, semantic objects, like road signs. In this paper, we introduce a novel approach to extract feature areas (FAs) within LiDAR point clouds enabling the detection of non-semantic map (MFAs) as well as on-board (KFAs) areas. KFAs compose a set of connected points with similar geometry-based descriptors which are extracted based on their benefit for the localization task. As opposed to other extraction methods based on LiDAR descriptors, our approach selects areas rather than detecting single key points. This input is used by our extraction approach in a two-stepped clustering and discarding process resulting in non-semantic segments. Our simple localization algorithm following the feature-based approach is more accurate than point-based localization on a real-world data set. We show that the feature extraction works persistently over data sets spanning one and a half year.