International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
STRUCTURE LINE DETECTION FROM LIDAR POINT CLOUDS USING
TOPOLOGICAL ELEVATION ANALYSIS
C. Y Lo'*,L CChen?
“ Department of Civil Engineering, National Central University, Taiwan - freezer@csrsr.ncu.edu.tw
? Center for Space and Remote Sensing, National Central University, Taiwan - lcchen@csrsr.ncu.edu.tw
Commission III, WG III/2
KEYWORDS: Detection, Automation, Analysis, Point cloud, LIDAR, Building, Model
ABSTRACT:
Airborne LIDAR point clouds, which have considerable points on object surfaces, are essential to building modeling. In the last two
decades, studies have developed different approaches to identify structure lines using two main approaches, data-driven and model-
driven. These studies have shown that automatic modeling processes depend on certain considerations, such as used thresholds,
initial value, designed formulas, and predefined cues. Following the development of laser scanning systems, scanning rates have
increased and can provide point clouds with higher point density. Therefore, this study proposes using topological elevation analysis
(TEA) to detect structure lines instead of threshold-dependent concepts and predefined constraints. This analysis contains two parts:
data pre-processing and structure line detection. To preserve the original elevation information, a pseudo-grid for generating digital
surface models is produced during the first part. The highest point in each grid is set as the elevation value, and its original three-
dimensional position is preserved. In the second part, using TEA, the structure lines are identified based on the topology of local
elevation changes in two directions. Because structure lines can contain certain geometric properties, their locations have small
relieves in the radial direction and steep elevation changes in the circular direction. Following the proposed approach, TEA can be
used to determine 3D line information without selecting thresholds. For validation, the TEA results are compared with those of the
region growing approach. The results indicate that the proposed method can produce structure lines using dense point clouds.
1. INTRODUCTION building boundaries are well known, the processes partition the
collected points according to their fit with the designed
Three-dimensional building modeling is essential for various primitives. A dormer on the roof can also be successfully
applications, such as 3D visualization, urban planning, geo- segmented using this scheme. However, these designed
database updating, and decision support. Key components primitives restrict the capability under considered building
including corners, edges, and planes can be used to delineate geometries.
the building geometry. Therefore, feature detection is the
highest priority for building procedures using aerial imagery, To overcome the limitation of the model-driven approach, the
LIDAR point clouds, and integration strategies. Among these data-driven approach is used to directly analyze the point cloud
data sources, aerial imagery provides the spatial information ^ distribution and estimate building geometries. Many studies
with spectral features. Conversely, LIDAR data shows the have employed various strategies including the Hough
building geometries with point cloud distribution. Following the transform, RANdom sample consensus (RANSAC) (Fischler
development of lidargrammetry, an airborne LIDAR system can and Bolles, 1981), binary space partitioning (BSP) (Sohn and
provide more point clouds to illustrate the relief of ground Dowman, 2007), knowledge-based criteria (Pu and Vosselman,
surfaces. From a building modeling perspective, laser scanning 2009), the octree-based split-and-merge segmentation algorithm
systems can provide more points with substantial accuracy and (Wang and Tseng, 2010), TIN-merging and reshaping (TMR)
high point density, facilitating the improvement of automatic (Rau and Lin, 2011), and the level set method (Kim and Shan,
modeling processes. However, the LIDAR system is a blind 2011). According to literature, segmentation approaches can be
device that cannot directly identify specific objects or features employed to manage unorganized point clouds for modeling.
(Ackermann, 1999). Because irregular point distribution is the The differences of local patches were analyzed under several
main aspect, analyzing point cloud distribution according to criteria including normal vectors, elevation differences, shapes,
building geometry, with either explicit constraints (model- ^ or mathematical weighting functions. Then, coplanar points
driven approach) or implicit constraints (data-driven approach), ^ were clustered together to fit the surface and calculate the
is necessary. structure lines through intersecting surfaces. Nevertheless, these
threshold-dependent approaches developed building surfaces by
To estimate building geometry using point clouds, the model- ^ adjusting the thresholds for different cases. The problematic
driven approach employs predefined model primitives. Maas issues may contain the starting position of segmentation
and Vosselman (1999) fitted point clouds to identify the basic processes, iterative procedures, and threshold determination.
primitives and their model parameters from invariant moments. Therefore, threshold modification and post-processing have
Point clouds can also be integrated with vector maps for become necessary steps.
building modeling (Vosselman and Dijkman, 2001). If the
* Corresponding author
140
Basec
LIDA
elevai
expli
phenc
analy
thresl
origir
and C
point
Becai
need
conve
detec
raster
proce
propc
to ea
spacii
LIDA
imple
struct
grid «
each
are tl
detec
map,
dime
result
merg
asses:
propc
To id
the t
estab
study
struc!
the c
radia
basic
elem
three
that i
and (
thres]
spaci
spaci
eleva
regar
