PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
December 2016
955
Planar-Based Adaptive
Down-Sampling of Point Clouds
Yun-Jou Lin, Ronald R Benziger, and Ayman Habib
Abstract
Derived point clouds from laser scanners and image-based
dense-matching techniques usually include tremendous
number of points. Processing (e.g., segmenting) such
huge dataset is time-consuming and might not be neces-
sary. For example, a planar surface just needs few points
to be defined. In contrast, linear/cylindrical and rough
features require more points for reliable modeling since
during the data acquisition process, only a portion of lin-
ear/cylindrical features is present in the point cloud.
This paper introduces an adaptive down-sampling strate-
gy for removing redundant points from high density planar
regions while retaining points in planar areas with sparse
points and all the points within linear/cylindrical and rough
neighborhoods. To demonstrate the feasibility and perfor-
mance of the proposed procedure, a comparison of segmenta-
tion results using original laser and image-based point clouds
as well as the adaptively, uniformly, and point-spacing-based
down-sampled point clouds are presented while commenting
on the computational efficiency and the segmentation quality.
Introduction
3D point cloud processing has been a critical task due to the
increasing demand of a variety of applications such as urban
planning and management, virtual reality, as-built mapping of
industrial sites, cultural heritage documentation, and change
detection (Gonzalez-Aguilera
et al
., 2013). Laser scanning and
optical imaging systems are the two main sources for point
cloud acquisition and derivation. Laser scanners are capable
of directly acquiring high precision point cloud along object
surfaces in an efficient manner. On the other hand, captured
imagery by optical sensors can provide spectral information,
high spatial resolution, and point clouds through a photogram-
metric space intersection. More specifically, point clouds can
be derived from overlapping images after the identification of
conjugate points in such imagery, which could be established
through modern dense image matching strategies. The derived
point clouds from laser scanners and image-based dense-match-
ing techniques usually include numerous points. Data process-
ing (e.g., segmentation and 3D modeling) of such huge datasets
is time-consuming and might not be necessary. In order to attain
high computational efficiency while maintaining the charac-
teristics of the local surface (i.e., planar, linear/cylindrical, and
rough features), an appropriate down-sampling is necessary.
As already mentioned, optical imagery and laser scanners
are the two major sources for indirectly or directly deriving
point clouds. Traditionally, point cloud generation from opti-
cal imagery relied on area-based and feature-based matching
techniques. Area-based matching evaluates the similarity of
grey values within a small template in one image and a search
window within overlapping images to determine conjugate
points. Pratt (1978) proposed the normalized cross-correla-
tion that can compensate for local brightness and contrast
variations between the grey values within the template and
search window. Feature-based methods use extracted features;
for example, through Feature from Accelerated Segment Test
(
FAST
) (Rosten and Drummond, 2006) to determine conju-
gate features. 3D coordinates of conjugate point and linear
features, which can be used for a wide range of applications
such as building model reconstruction (Suveg and Vosselman,
2004), can be derived through a photogrammetric space inter-
section. Area-based image matching is not rotation invariant,
and both area-based and feature-based image matching tech-
niques are incapable of providing a detailed object descrip-
tion. Compared to area-based and feature-based matching
techniques, recently-developed dense-matching algorithms
can provide precise point clouds with high density through
a global matching constraint (Furukawa and Ponce, 2010).
Hirschmuller (2005) proposed Semi-Global Matching (
SGM
)
that performs pixel-wise matching using mutual information.
Haala (2013) showed that pixel-wise dense matching and cur-
rent software tools are capable of generating high definition
landscape digital surface models (
DSM
) from airborne imagery.
Depending on the utilized platform, laser scanners can be
categorized into Airborne Laser Scanners (
ALS
), Stationary
Terrestrial Laser Scanners (
STLS
), and Mobile Terrestrial Laser
Scanners (
MTLS
).
ALS
was developed in the early 1990s, and is
mainly used for collecting surface data over large areas. The
level of detail in
ALS
data depends on the flying height, pulse
rate, scanning rate, aircraft speed, field of view (
FOV
), among
other parameters. The point density for
ALS
data usually ranges
from 1 to 40 pts/m
2
(Hyyppä
et al
., 2009) which is suitable for
surface description, digital terrain model (
DTM
) generation (Liu,
2008), and rough building model generation (Habib
et al
., 2010;
Kwak and Habib, 2014). Due to the nature of the data acqui-
sition mechanism,
ALS
systems cannot provide the necessary
details for extracting objects that do not belong to building roof-
tops and terrain such as building facades, light poles, trees, and
fences. As a result of their proximity to the objects of interest,
STLS
and
MTLS
systems can acquire dense point clouds for the
extraction and modeling of building facades, fences, trees, and
light-poles.
STLS
systems were introduced almost 15 years ago
and with the improvement in modern terrestrial direct geo-ref-
erencing systems,
MTLS
systems are now capable of collecting
high density point clouds relative to a global reference frame.
Pu
et al
. (2011) identified road corridors, lanes, curbs, poles,
barriers, building facades, and vegetation from
MTLS
data to sat-
isfy the needs of transportation management applications.
STLS
and
MTLS
point clouds have been also used for tree reconstruc-
tion (Côté
et al
., 2009), road extraction, curb detection (Yang
et
al
., 2013), and facades reconstruction (Becker and Haala, 2009).
Derived point clouds from
ALS
,
STLS
, and
MTLS
systems
as well as image-based dense-matching techniques usually
include excessive number of points. Processing (e.g.,
Lyles School of Civil Engineering, Purdue University, West
Lafayette, IN 47906 (
).
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 12, December 2016, pp. 955–966.
0099-1112/16/955–966
© 2016 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.82.12.955
