Efficient interpretation of 3D point clouds by assessing feature relevance

The semantic interpretation of 3D point cloud data acquired with mobile laser scanning (MLS) systems has become a topic of major interest for photogrammetry, remote sensing and computer vision. In this paper, we propose a methodology for the semantic interpretation of point cloud data in terms of assigning each 3D point a semantic label. Our methodology involves (1) individual neighborhoods of optimal size in order to provide distinctive geometric features for each 3D point and (2) feature relevance assessment in order to reduce the computational burden with respect to processing time and memory consumption. More specifically, our approach for feature relevance assessment relies on a general relevance metric composed of seven different, classifier-independent feature selection strategies and thus addresses different intrinsic properties of the given training data. The results derived for a labeled benchmark dataset with about 1.3 million 3D points reveal that, instead of including as many features as possible in order to compensate a lack of knowledge about scene and data, a crucial task such as the semantic scene interpretation can be carried out with only few relevant features without a significant loss in classification accuracy.