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
EDGE-BASED REGISTRATION FOR AIRBORNE IMAGERY AND LIDAR DATA
L.C. Chen**, C.Y. Lo®
* Center for Space and Remote Sensing, National Central University, Taiwan - lcchen(g)csrsr.ncu.edu.tw
? Dept. of Civil Engineering, National Central University, Taiwan - freezer@csrsr.ncu.edu.tw
Commission III, WG III/4
KEY WORDS: Aerial Imagery, Building, Detection, LIDAR, Model, Registration
ABSTRACT:
Aerial imagery and LIDAR points are two important data sources for building reconstruction in a geospatial area. Aerial imagery
implies building contours with planimetric features; LIDAR data explicitly represent building geometries using three-dimensional
discrete point clouds. Data integration may take advantage of merits from two data sources in building reconstruction and change
detection. However, heterogeneous data may contain a relative displacement because of different sensors and the capture time. To
reduce this displacement, data registration should be an essential step. Therefore, this investigation proposes an edge-based approach
to register these two data sets in three parts: (1) data preprocessing; (2) feature detection; and (3) data registration. The first step
rasterizes laser point clouds into a pseudo-grid digital surface model (PDSM), which describes the relief with the original elevation
information. The second step implements topological analyses to detect image edges and three-dimensional structure lines from the
aerial image and PDSM. These detected features provide the initial positions of building shapes for registration. The third part
registers these two data sets in Hough space to compensate for the displacement. Because each building may have prominent
geometric structures, the proposed scheme transforms these two groups of edges, and estimates the correspondence by the Hough
distribution. The following procedure then iteratively compares two groups of Hough patterns, which are from an aerial image and
LIDAR data. This iterative procedure stops when the displacement is within a threshold. The test area is located in Taipei City,
Taiwan. DMC system captured the aerial image with 18-cm spatial resolution. The LIDAR data were scanned with a 10-point
density per square meter using the Leica ALS50 system. This study proposed a 50 cm spatial resolution of PDSM, which is slightly
larger than the point spacing. The experiment selected two buildings to evaluate the performance of the proposed scheme. The
manually edited building boundaries from the stereo aerial images are the reference data for validation. Comparisons indicated that
the registration procedure could adjust the displacement within 50 cm, which relates to PDSM resolution. These preliminary results
also demonstrated the possibility of providing locations for building reconstruction.
1. INTRODUCTION
Buildings are essential objectives in a geospatial area. To
achieve building modeling, the detection and reconstruction of
building models comprise two major steps. Within the literature,
three common data sets are used frequently. These include
topomaps, aerial images, and LIDAR data. These three data sets
represent building geometries in different phases. Topomaps
describe existing building shapes with polylines or polygons, so
that the locations are well known in a single time point. These
vector data can generate models directly with accurate contour
lines. However, topomaps may lack structure lines on rooftops.
In certain places, even topomaps do not contain the elevation
information. For the reconstruction of new buildings, changed
structures, and detailed buildings, topomaps are insufficient for
building reconstruction. Aerial images and LIDAR data then
become two important data sources.
Aerial images implicitly exhibit building geometries by edges
and corners. Building extraction thus becomes a necessary step.
The stereo pairs provide a basis to manually measure corners
(Gruen and Wang, 2001) or structure lines (Rau and Chen, 2003)
for building reconstruction with semiautomatic processes. By
improving modeling automation, matching and positioning
techniques generate the digital surface model (DSM) to
compare with the designed primitives and reconstructed
* Corresponding author
building models by using the model-based concept (Hammoudi
and Dornaika, 2011) or extract building areas by image
classification (Zebedin et al., 2006). Although these approaches
can enhance the modeling ability, the quality of generated DSM
restricts the details of focused buildings directly.
Conversely, the laser scanning system can provide dense point
clouds to illustrate relief. This device blindly collects three-
dimensional information of all objects along flight strips with
discrete points (Ackermann, 1999). Building extraction is thus a
high priority. In comparison with aerial imagery, the
distribution of point clouds is helpful to identify building
locations by using the elevation constraint. However, using this
unorganized format may be difficult to delineate building
boundaries directly by using image processing. Additional
processes are necessary. Accordingly, numerous studies have
proposed varied strategies to estimate three-dimensional lines
from LIDAR data (Sohn and Dowman, 2007; Pu and
Vosselman, 2009; Wang and Tseng, 2010). The major concept
is the segmentation to derive point clouds of building facades
and intersect lines for model reconstruction. The segmentation
results depend on the selected mathematical models for point
collection.
Based on the combination of merits from heterogeneous data
sets, data integration may be helpful to improve the degree of
autc
pert
eta
Bec
con
buil
a né
revi
mar
exis
dete
stuc
for
201
they
Zar
Hat
Cori
pro
Ho
con
disc
mo:
To
ana
ima
det:
the:
ind
mo
loc:
Thi
aer
edg
fea!
inl
2.1
Be
dis
eac
prc
dat
Tec
