3D Semantic labeling of ALS point clouds by exploiting multi-scale, multi-type neighborhoods for feature extraction

The semantic labeling of 3D point clouds acquired via airborne laser scanning typically relies on the use of geometric features. In this paper, we present a framework considering complementary types of geometric features extracted from multi-scale, multi-type neighborhoods to describe (i) the local 3D structure for neighborhoods of different scale and type and (ii) how the local 3D structure behaves across different scales and across different neighborhood types. The derived features are provided as input for several classifiers with different learning principles in order to show the potential and limitations of the proposed geometric features with respect to the classification task. To allow a comparison of the performance of our framework to the performance of existing and future approaches, we evaluate our framework on the publicly available dataset provided for the ISPRS benchmark on 3D semantic labeling.