A Voxel- and Graph-Based Strategy for
Segmenting Man-Made Infrastructures Using
Perceptual Grouping Laws:
Comparison and Evaluation
Yusheng Xu, Ludwig Hoegner, Sebastian Tuttas, and Uwe Stilla
Abstract
In this paper, we report a novel strategy for segmenting 3D
point clouds using a voxel structure and graph-based cluster-
ing with perceptual grouping laws. It provides a completely
automatic solution for partitioning point clouds of man-made
infrastructure. Two different segmentation methods using
voxel and supervoxel structures are presented and evaluated.
To increase the efficiency and the robustness of the segmenta-
tion process, the voxelization with octree-based structure is
introduced, which can suppress effects of noise, outliers, and
unevenly distributed point densities as well. The clustering
of over-segmented voxels and supervoxels is achieved using
graph theory on the basis of the local contextual information,
which is commonly conducted merely with pairwise infor-
mation in conventional clustering algorithms. The graphical
model is constructed according to perceptual grouping laws,
considering geometric information associated with points.
Experiments using both laser scanning and photogrammetric
point clouds have demonstrated that the proposed methods
can achieve good results, especially complex scenes and
nonplanar object surfaces, with F1-measures better than
0.67 for all the testing samples. Quantitative comparisons
between the proposed approaches and other representative
segmentation methods also confirm the effectiveness and the
efficiency of the former. Moreover, a series of experiments
is carried out, to investigate the methods’ sensitivity with
respect to various parameters on the segmentation results.
Introduction
In recent years, the reconstruction of 3D scenes using point
clouds obtained from laser scanning, stereo matching, and
range imaging cameras is attracting increasing attention
for many tasks such as constructing virtual reality, creating
digital surface models, or monitoring construction projects. In
such settings, point clouds are considered to be suitable data
sources for recognizing and reconstructing geometric objects
from 3D scenes. In general, individual objects should be iden-
tified and separated from the scene prior to the recognition
or modeling procedure. This is because the majority of both
indoor and outdoor scenes usually contain different types of
objects, complex structures, and surfaces of various shapes,
so that it is hard to directly recognize or model certain kinds
of objects from the scene (Yang
et al.
, 2015). To this end, for
unstructured raw point clouds, segmentation is a fundamen-
tal step and commonly applied to partition the 3D scene into
the largest possible meaningful segments, namely groups of
points having one or more characteristics in common (Grilli
et al.
, 2017; Vosselman
et al.
, 2017).
Theoretically, a well thought-out segmentation algorithm
can remove irrelevant objects in the scene and largely lessen
the burden of workloads of computing and storing. However,
the performance of conventional point cloud segmentation
algorithms commonly deteriorates in complex environments
of real outdoor scenes, especially for scenes containing build-
ings, where occlusions frequently occur. Such complex scenes
degrade the performance of commonly used methods, because
most of the segmentation criteria use merely pairwise infor-
mation between elements (e.g., differences of normal vectors
calculated from points), which is sensitive to missing points
and incomplete structures caused by occlusions. Moreover,
the data quality is also a leading cause for erroneous segmen-
tation results. For instance, outliers and unevenly distributed
points densities significantly affect the boundary quality of
segments owing to wrongly estimated point-based geometric
features (e.g., normal vectors or local density). Hence, apart
from the effectiveness, the reliability should be considered in
the development of segmentation algorithms as well. On the
other hand, as point cloud segmentation is computationally
intensive, efficiency is also relevant and should be taken into
consideration when having to cope with large datasets.
To tackle the aforementioned problems, we propose a point
cloud segmentation strategy to efficiently acquire geometric
segments from large-scale point clouds of man-made scene. In
Figure 1, we give an illustration of how the point cloud of the
scene is segmented. Here, the entire point cloud of the scene
is split into individual segments relating to semantic objects
Photogrammetry and Remote Sensing, Technische Universität
München, 80333 Munich, Germany (
).
Photogrammetric Engineering & Remote Sensing
Vol. 84, No.6, June 2018, pp. 377–391.
0099-1112/18/377–391
© 2018 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.84.6.377
Figure 1. Segmenting a 3D point cloud: (a) Original
unstructured point cloud (textured with
RGB
colors), and (b)
Segmented result rendered with different gray values.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2018
377
