A Hybrid Semantic Point Cloud Classification-Segmentation Framework Based on Geometric Features and Semantic Rules

Martin Weinmann, Stefan Hinz und Michael Weinmann
In: Journal of Photogrammetry, Remote Sensing and Geoinformation Science (2017), 85(183-194)
 

Abstract

In this paper, we focus on semantic point cloud classification taking into account standard failure cases reported in a variety of investigations. We present a hybrid two-step framework integrating classification, segmentation and semantic rules in a common end-to-end processing pipeline from irregularly distributed points to semantically labelled point clouds. The first step of our framework consists of a point-wise semantic point cloud classification based on rather intuitive, handcrafted, low-level geometric features extracted from local neighbourhoods of locally adaptive size. The second step of our framework consists of refining the point-wise classification results by considering semantic rules applied to geometric features extracted on the basis of an over-segmentation of the derived class-wise point clouds. We demonstrate the performance of our framework on a standard benchmark dataset for which we obtain a semantic labelling of high accuracy and high plausibility.

Bibtex

@ARTICLE{weinmann-2017-pfg,
    author = {Weinmann, Martin and Hinz, Stefan and Weinmann, Michael},
     pages = {183--194},
     title = {A Hybrid Semantic Point Cloud Classification-Segmentation Framework Based on Geometric Features and
              Semantic Rules},
   journal = {Journal of Photogrammetry, Remote Sensing and Geoinformation Science},
    volume = {85},
      year = {2017},
  abstract = {In this paper, we focus on semantic point cloud classification taking into account standard failure
              cases reported in a variety of investigations. We present a hybrid two-step framework integrating
              classification, segmentation and semantic rules in a common end-to-end processing pipeline from
              irregularly distributed points to semantically labelled point clouds. The first step of our
              framework consists of a point-wise semantic point cloud classification based on rather intuitive,
              handcrafted, low-level geometric features extracted from local neighbourhoods of locally adaptive
              size. The second step of our framework consists of refining the point-wise classification results by
              considering semantic rules applied to geometric features extracted on the basis of an
              over-segmentation of the derived class-wise point clouds. We demonstrate the performance of our
              framework on a standard benchmark dataset for which we obtain a semantic labelling of high accuracy
              and high plausibility.},
       doi = {10.1007/s41064-017-0020-5}
}