Investigations on the potential of binary and multi-class classification for object extraction from airborne laser scanning point clouds

Martin Weinmann, Rosmarie Blomley, Michael Weinmann, and Boris Jutzi
In: Tagungsband der 38. Wissenschaftlich-Technischen Jahrestagung der DGPF (2018), 27(408-421)
 

Abstract

In this paper, we address the strategies of a binary classification and a multi-class classification for the pointwise semantic labeling of airborne laser scanning data. For both strategies, we use a collection of spherical and cylindrical neighborhoods as the basis for extracting geometric multi-scale features for each point of a considered point cloud. The extracted features, in turn, are provided as input to a standard Random Forest classifier. The results achieved for multi-class classification indicate a better classification across different classes, which is important for a subsequent spatial regularization. The results achieved for binary classification addressing the detection of cars, buildings and trees, respectively, show the potential for a subsequent extraction of individual objects.

Bibtex

@ARTICLE{weinmann-2018-dgpf,
    author = {Weinmann, Martin and Blomley, Rosmarie and Weinmann, Michael and Jutzi, Boris},
     pages = {408--421},
     title = {Investigations on the potential of binary and multi-class classification for object extraction from
              airborne laser scanning point clouds},
   journal = {Tagungsband der 38. Wissenschaftlich-Technischen Jahrestagung der DGPF},
    volume = {27},
      year = {2018},
  abstract = {In this paper, we address the strategies of a binary classification and a multi-class classification
              for the pointwise semantic labeling of airborne laser scanning data. For both strategies, we use a
              collection of spherical and cylindrical neighborhoods as the basis for extracting  geometric 
              multi-scale features for each point of a considered point cloud. The extracted features, in turn,
              are provided as input to a standard Random Forest classifier. The results achieved for multi-class
              classification indicate a better classification across different classes, which is important for a
              subsequent spatial regularization. The results achieved for binary classification addressing the
              detection of cars, buildings and trees, respectively, show the potential for a subsequent extraction
              of individual objects.}
}