this PDF file - Journal of Geodesy and Geoinformation

Transcription

this PDF file - Journal of Geodesy and Geoinformation
UCTEA
Chamber of Surveying and Cadastre Engineers
Journal of Geodesy and Geoinformation
TMMOB
Harita ve Kadastro Mühendisleri Odası
Jeodezi ve Jeoinformasyon Dergisi
Vol.2  Iss.1  pp.1-8  May 2013  Journal No.107  Doi: 10.9733/jgg.250913.2
www.hkmodergi.org
Photogrammetric features for the registration of
terrestrial laser scans with minimum overlap
Sibel Canaz1,2,* Ayman Habib1
1
Department of Geomatics Engineering, University of Calgary, Schulich School of Engineering, Calgary, AB, Canada
Deparment of Geomatics Engineering, Karadeniz Technical University, Trabzon, Turkey
2
Abstract
Volume: 2
Issue: 1
Page: 1 - 8
May 2013
The interest and demand for 3 Dimensional (3D) documentation of man-made structures have increased with
the continuous improvement in data acquisition systems and the expanding range of potential applications.
Currently, 3D data can be obtained through two technologies: photogrammetry and laser scanning. In spite
of the proven quality of static laser scanning, the collection and processing of laser scans is a time consuming process and the derivation of complete 3D models requires multiple scans with each scan having its own
coordinate system. Therefore, the different scans should be aligned in a common coordinate system. This
alignment process is known as “registration”. Current registration techniques require large overlap area
between the collected scans in order to obtain reliable estimation of the transformation parameters relating
these scans. The main objective of this research is to avoid the requirement of large overlap areas among
the laser scans using photogrammetric data for the registration process. Registration methods generally use
points as primitives to relate the different laser scans. In this research, photogrammetric linear and planar
features are used for the proposed registration method. Finally, qualitative and quantitative quality control
measures are proposed for analyzing the result of the proposed registration method.
Keywords
Terrestrial laser scanning, Close range photogrammetry, Registration, Feature extraction
Özet
Minimum çakışan yersel lazer taramalarının birleştirmesi için fotogrametrik özellikler
Cilt: 2
Sayı: 1
Sayfa: 1 - 8
Mayıs 2013
Günümüzde, insan yapımı binaların 3 Boyutlu (3B) modellemesine olan ilgi ve talep veri toplama sistemlerindeki gelişmeler ve büyüyen potansiyel uygulama aralığı ile birlikte arttı. Mevcut durumda, 3B veriler iki
teknoloji ile sağlanmaktadır: fotogrametri ve lazer taraması. Lazer taramalarının toplanması ve islenmesinde zaman kaybı söz konusudur. Bütün bir 3B model elde etmek için birçok lazer taraması gerekmektedir. Her
bir lazer taraması ayrı koordinat sistemlerinde olduğu için; 3B modeller ancak bu farklı koordinatlardaki
lazer tamamlarının bir araya dizilmesi ile mümkündür. Bu dizilim “birleştirme” olarak bilinir. Birçok birleştirme teknikleri doğru 3B model için toplanan lazer taramaları arasında büyük oranda çakışma ister. Bu
araştırmanın temel amacı fotogrametrik veri kullanarak lazer taramaları arasındaki büyük oranda istenilen
çakışmadan kaçınmaktır. Birleştirme tekniklerinde genellikle noktalar kullanarak lazer taramaları arasındaki ilişki bulunur. Bu önerilen araştırmada düzlemsel ve doğrusal özellikler çıkarılmış ve kullanılmıştır.
Son olarak, nitel ve nicel kalite kontrol yöntemleri önerilen birleştirme metodu sonucunu analiz etmek için
yapılmıştır.
Anahtar Sözcükler
Yersel lazer tarayıcı, Yersel fotogrametri, Birleştirme, Özellik çıkarma
*Corresponding Author: Tel: +90 462 377 27 76 Fax: +90 462 328 0918
E-mail: sibelcanaz@gmail.com (Canaz S.), ahabib@ucalgary.ca (Habib A.)
© 2013 HKMO
2 Photogrammetric features for the registration of terrestrial laser scans with minimum overlap
1. Introduction
Representation of an object in a digital environment is denoted as 3D modeling. 3D models of man-made structures are
necessary for their digital documentation, which is useful for
a wide range of applications. For instance, 3D models of historical buildings can help in their reconstruction if they are
damaged by natural or human events. Another 3D modeling
application can be given in archeology where sites/objects
can be studied through an investigation of their 3D models.
Some other areas that frequently use 3D models include architecture, forestry, geology, urban planning, and civil engineering. Due to the wide spectrum of applications that are
based on 3D models, several researchers have been focusing
on creating 3D models of our environment in general and
man-made structures in particular.
Laser scanning technology has become an established
technique for creating 3D models due to their ability of
direct acquisition of 3D data. A more traditional approach
for deriving 3D models of objects of interest is photogrammetric reconstruction. In contrast to laser scanning, 3D data
can be only derived from overlapping imagery whenever
conjugate features have been identified and the intersection
of the respective spatial rays is mathematically described.
Considering the characteristics of laser scanning and photogrammetric systems, the user community has been interested in utilizing the former for 3D model generation due
to the direct derivation of the coordinates of scanned surfaces. However, in many applications, it is almost impossible to cover a complete structure with a single Terrestrial
Laser Scanner (TLS) scan. Therefore, several TLS scans
from different stations are necessary for complete coverage
of the surveyed structure (i.e., different scans from different
stations produce sub-models of the object in question). The
captured scans/sub-models are related to different coordinate
systems associated with the laser scanning units at the different stations. To obtain meaningful 3D models, alignment of
the collected scans in a common coordinate system is a prerequisite. This alignment process is known as registration.
An effective registration paradigm has four components
that should be carefully established: appropriate primitives,
transformation parameters, similarity measure, and matching strategy (Al-Ruzouq 2004). First of all, registration
primitives, which can be reliably identified/extracted from
the overlapping sub-models, should be established. The used
primitives for the registration of TLS scans can be either
point, linear, or planar (areal) features. Identification of conjugate points in different TLS scans is almost impossible due
to the irregular distribution of the TLS point cloud. On the
other hand, it is possible to identify planar and linear features
in TLS scans (Habib et al. 2008). For this reason, planar and
linear features are chosen as the registration primitives in this
research. The second element of a registration paradigm is
establishing the transformation parameters, which describe
the relationships between the reference frames of the involved datasets (i.e., sub-models covering different portions
of the object in question). Usually, three translations, three
rotations, and a scale factor are used to define the relationship between the reference frames of the sub-models. The
third element of the registration paradigm, which is the similarity measure, incorporates the matched primitives in over-
lapping scans together with the transformation parameters
to mathematically describe the coincidence of the involved
primitives following the registration process (Renaduin et al.
2011). Finally, the matching strategy, which can be considered as an overall scheme for solving the registration problem, needs to be designed to manipulate the primitives for
the estimation of the necessary transformation parameters
using the established similarity measure (Al-Ruzouq 2004).
There are many different registration techniques available in the literature. The Iterative Closest Point (ICP),
which is based on minimizing the point-to-point distance
in the overlapping area between different TLS scans, is the
most popular and generally used algorithm for registering
3D point clouds (Besl and McKay 1992). Basically, the sum
of squares of the Euclidean distances between the nearest
points of two data sets is minimized in the ICP registration
technique. Similar to the ICP, Chen and Medioni (1992) developed a method, where the point-to-plane normal distances
in the overlapping area among the TLS scans are minimized.
A variant of these methods, the Iterative Closest Projected
Point (ICPP) can be considered as both a point-to-point and
point-to-plane registration technique (Al-Durgham et al.
2011). The ICPP minimizes the distance between a point
in one scan and its projection on the closest three points in
the other scan. Some other researchers used a camera that
was mounted on the top of the TLS system (a hybrid system) to integrate photogrammetric data and TLS scans for
registration purposes. The collected images were used for
registering the TLS scans using the estimated mounting
parameters between the two systems. These mounting parameters were determined from a prior system calibration
process. Dold and Brenner (2006) conducted a registration
study using a hybrid system. Al-Manasir and Fraser (2006)
established another example of registration using a hybrid
system. Coded targets on the object were used to register
TLS scans automatically. The major disadvantage of using
a hybrid system for registration purposes is that it is not always possible to have a good observation station for collecting both a TLS scan and photogrammetric data at the
same time. Global Positioning Systems (GPS) device can
also be used to estimate relations between TLS scans and
reference coordinate system while collecting the scans. For
instance Balzani et al. (2002) conducted a study for geo-referencing scans using GPS device. The drawback behind this
methodology for registering scans is that the effort in using
GPS to collect control points for the geo-referencing of the
scans that have minimum overlap is more than the required
effort in the image data acquisition. The ICP and many other registration methods generally require large overlap areas
among the TLS scans for deriving reliable estimates of the
transformation parameters relating the different sub-models.
In this research, the large overlap area requirement among
the scans is eliminated/reduced using photogrammetric data,
which can be acquired in a relatively short time. Since linear
and planar features can be easily derived from the photogrammetric data, identified linear and planar features in both
the photogrammetric model and the TLS scans are used for
the co-alignment of different scans to a common reference
frame.
Overall, there are many advantages of the proposed registration method. The main advantage of having minimum
Sibel Canaz, Ayman Habib / Vol.2  Iss.1  2013 3
overlap area among the TLS scans and using photogrammetric data to acquire 3D model is that the proposed procedure
will lead to reduction in the field data collection time, which
will lead to reduction in the acquisition cost. Moreover, in
the presence of some restrictions during the data acquisition
(e.g., busy traffic, cluttered environment, and/or unstable
soil), fewer scans would be needed to cover the object of
interest. On the other hand, the image data acquisition can
be done quickly with minimum time and site stability requirements. Furthermore, the proposed procedure will lead
to reduction in the amount of collected data, which will
have a positive impact on the data processing and manipulation activities. In this regard, it should be noted that the
massive amount of collected data by terrestrial laser scans
would increase the complexity and storage requirements
for the subsequent data processing activities. In addition,
without image data, significant overlap among the scans as
well as the presence of variation in the slopes/aspects within
the overlap regions are necessary for the registration these
scans. In some situations, this would not be possible. For
example, if we have a structure with a large extent that does
not exhibit good geometry in the overlap regions (i.e., the
overlap region regardless of its extent would not have good
geometry for reliable registration – e.g., building with a large
facade). In those situations, the proposed procedure would
facilitate the possibility of registering these scans as long
as the scan has a minimum of two non-parallel linear features that could be identified in the photogrammetric model.
Another advantage of the proposed procedure is that it will
ensure the co-registration of multi-sensory data to a common
platform. One of the main benefits of such a registration is
the straightforward ability of linking the spectral information
from imagery with the laser scanning point cloud. Moreover,
the same procedure can be used to for the alignment of terrestrial laser scans with minimal overlap with airborne laser
scanning data or airborne imagery that have common features that could be identified in the TLS scans. Finally, with
the increasing interest in using overlapping imagery together
with dense matching techniques, the possibility of generating a dense point cloud from imagery is becoming a reality.
The proposed procedure would allow for the registration of
the laser and image-based point clouds. Such a registration
would have a positive impact on the calibration of laser scanning systems as well as the refinement of the dense matching
algorithms and its results.
Following the registration process, a quality control procedure is then adopted to ensure/verify the quality of the
estimated transformation parameters. A qualitative quality
control is conducted by plotting and visually inspecting the
registered scans. A quantitative quality control, on the other
hand, is established by calculating the point-to-plane normal
distances between the registered surfaces. As an additional
quantitative quality control, results from the proposed registration method are compared with those derived from the
ICPP method. The next section deals with the extraction of
the registration primitives from the photogrammetric data
and the TLS scans. Then, the registration and the quality
control process are explained. The experimental results section evaluates the performance of the proposed procedure.
Finally, conclusions and recommendations for future work
are presented.
2. Feature extraction from TLS scans and
photogrammetric data
2.1 Feature extraction from photogrammetric data
In this research, we start by collecting a set of overlapping
images that cover the object of interest. Following the camera calibration and the acquisition of the images that cover
the object in question, a set of tie points are identified in the
captured imagery to establish the position and orientation
of these images in an arbitrarily-defined coordinate system
using a bundle adjustment procedure. For the alignment of
the TLS scans, additional linear and planar features are incorporated in the bundle adjustment. Those features will be
later identified in the TLS scans. A planar surface is defined
by some points along its boundary. More specifically, three
or four points are observed in multiple images, and their
model space coordinates are estimated through the bundle
adjustment procedure. On the other hand, linear features are
represented by their end points. The coordinates of the end
points are derived from a coplanarity constraint, which can
be added to the bundle adjustment procedure (Habib et al.
2004; Habib et al. 2007; and Renaudin et al. 2011). The end
points are only identified in one of the images while a set
of intermediate points, which are mono-scopically identified
in the overlapping images, are used in the coplanarity constraint. Interested readers can refer to Habib et al. (2004),
Habib et al. (2007), and Renaudin et al. (2011) for more details regarding the derivation of the linear features from the
photogrammetric data. In summary, a set of points representing planar and linear features will be derived from the photogrammetric bundle adjustment procedure. These points are
denoted as Photogrammetrically Reconstructed Data (PRD)
in this research. The PRD are used together with the extracted features from the TLS scans for their registration.
2.2 Feature extraction from terrestrial laser scans
For the feature extraction from the TLS scans, a segmentation procedure is utilized to identify the points that belong to
the same plane. In this research, a novel parameter-domain
segmentation approach, which is introduced by Lari et al.
(2011), is used. In this approach, the neighborhood of each
point is established using an adaptive cylinder, whose axis is
changing its orientation to be aligned along the normal to the
local planar surface through the point in question. The adaptive cylinder is used to derive the attributes of the point in
question. These attributes are stored in an accumulator array.
Rather than using a tessellation for the manipulation the accumulator array, a kd-data structure is implemented to store
the attributes of the point cloud. Finally, an efficient peak
detection procedure is utilized to identify clusters of similar
attributes (Lari and Habib 2012). Following the identification of the planar regions, three or four points from each of
the planes that have been identified in the photogrammetric
data are manually chosen. It is worth mentioning that the
chosen points in the TLS scans are not necessarily conjugate
to the selected points in photogrammetric data to define the
same planar patch. For the linear feature extraction, neighboring planes are intersected. The intersection of such planes
provides infinite linear features. The end points of the TLS
444Photogrammetric
Photogrammetric
features
for
the
registration
terrestrial
laser
scans
with
minimum
overlap
Photogrammetricfeatures
featuresfor
forthe
theregistration
registrationofofofterrestrial
terrestriallaser
laserscans
scanswith
withminimum
minimumoverlap
overlap
⃗
⃗
⃗⃗⃗ ⃗
⃗⃗⃗ ⃗
⃗⃗⃗ ⃗
⃗
⃗
⃗
linear
linear
features
are
established
through
the
projection
of
linearfeatures
featuresare
areestablished
establishedthrough
throughthe
theprojection
projectionof
of ⃗
⃗
⃗
(1)
(1)
(1)
(1) (1)
the
the
point
cloud
in
the
neighboring
planes
within
givthepoint
pointcloud
cloudin
inthe
theneighboring
neighboringplanes
planeswithin
withinaaagivgiv(1)
where:
where:
en
where:
en
buffer
around
their
intersection
onto
the
infinite
linear
enbuffer
bufferaround
aroundtheir
theirintersection
intersectiononto
ontothe
theinfinite
infinitelinear
linear
where:
where:
(1)
feature
and
2012)
where:
feature
(Lari
andHabib
Habib2007)
2012)Similar
Similarto
tothe
theplanar
planar
features,
feature(Lari
(Al-Durgham
to
planarfeatures,
features,
- Scani indicates the ithth
scan;
indicates the
the iiththTLS
TLS
scan;
Scanii indicates
where:
the
the
points
defining
the
lines
the
photogrammetric
and
TLS
Scani indica
th scan;
thepoints
pointsdefining
definingthe
thelines
linesinininthe
thephotogrammetric
photogrammetricand
andTLS
TLS
--the i TLS
TLS
scan;scan;
ScanScan
i indicates
TLS
-Scan
i indicatesththe i
data
data
need
not
be
conjugate.
The
sets
of
non-conjugate
points
dataneed
neednot
notbe
beconjugate.
conjugate.The
Thesets
setsof
ofnon-conjugate
non-conjugatepoints
points
Scani indicates the i TLS scan;
is the observation vector of the point
along
along
corresponding
linear
and
planar
surfaces
are
used
for
alongcorresponding
correspondinglinear
linearand
andplanar
planarsurfaces
surfacesare
areused
usedfor
for
the ith] isTLS
-⃗⃗ Scani indicates
coordinates
either in
the ith
TLS
[
the scan;
observation
vector
of the
the
point⃗ coordinat
coordinat
-- point
]]isisthe
vector
of
the
observation
vector
of the
coordina
- -- ⃗Xscani⃗/PRD) = [[ [
the
] or
isobservation
the
observation
vector
ofpoint
the point
coor
the
registration
process.
One
should
note
that
this
stage,
scan
the photogrammetrically
retheregistration
registrationprocess.
process.One
Oneshould
shouldnote
notethat
thatatatatthis
thisstage,
stage,
⃗
[
]
is
the
observation
vector
of
the
point
coordinat
constructed
data;
the
photogrammetric
model
generation,
including
the
derithe
the photogrammetric
photogrammetric model
model generation,
generation, including
including the
the derideri⃗
[ reconstructed
] is the
observation vector of the point coor
-photogrammetrically
data;
vation
vation
of
linear
and
planar
features,
based
on
manual
photogramm
vationof
oflinear
linearand
andplanar
planarfeatures,
features,isisisbased
basedon
onaaamanual
manual photogrammetrically
reconstructed
data;
photogrammetrically
reconstructed
data;data;
photogrammetrically
reconstructed
processing
processing
of
the
imagery.
The
automated
triangulation
and
processingof
ofthe
theimagery.
imagery.The
Theautomated
automatedtriangulation
triangulationand
and photogrammetrically reconstructed data;
is the translation vector between
derivation
derivation
of
linear
features
are
the
focus
of
current
and
fuderivationof
oflinear
linearfeatures
featuresare
arethe
thefocus
focusof
ofcurrent
currentand
andfufu- ⃗ photogrammetrically reconstructed
data;scan or the phoeither
TLS
is the
the ithtranslation
translation
vector- between
between
eith
⃗
--- - ⃗X-⃗T scan⃗i /PRD) = [[[ [
]]] is
the
vector
eith
ture
ture
research.
However,
the
linear
and
planar
features
are
is
the
translation
vector
between
eith
tureresearch.
research.However,
However,the
thelinear
linearand
andplanar
planarfeatures
featuresare
are
] is the reconstructed
translation
vector
between
togrammetrically
data
⃗
[
] andisthethe
translation
vector
between
eith
automatically
automatically
extracted
from
the
TLS
data.
The
matching
of
automaticallyextracted
extractedfrom
fromthe
theTLS
TLSdata.
data.The
Thematching
matchingof
of
ground coordinate systems;
⃗
[
] is the translation vector between
image
image
and
TLS
linear
features
are
based
on
manual
proceimageand
andTLS
TLSlinear
linearfeatures
featuresare
arebased
basedon
onmanual
manualproceproce- photogrammetrically
reconstructed
data and
and the
the ground
ground coordinate
coordinate
system
photogramm
photogrammetrically
reconstructed
data
system
reconstructed
data data
and the
coordinate
system
dures.
photogrammetrically
reconstructed
andground
the ground
coordinate
sy
dures.
Examples
of
automatically
derived
linear
and
planar
dures.Examples
Examplesof
ofautomatically
automaticallyderived
derivedlinear
linearand
andplanar
planar photogrammetrically
th
photogrammetrically
reconstructed
data
and
ground
system
thth TLS coordinate
ththe
is
the
scale
factor
between
either
i
scan
or
the
photo
i
features
from
a
TLS
scan
and
manually
extracted
features
th
features
from
a
TLS
scan
and
manually
extracted
features
/PRD)
S
is
the
scale
factor
between
either
the
i
TLS
scan
or
the
isisthe
factor
between
either
the
TLS
-features from a TLS scan and manually extracted features
theisscale
scale
factorfactor
between
eithereither
theii the
TLS
scanor
orthe
theorphoto
photo
the scale
between
i scan
TLS
scan
thesy
p
- scaniphotogrammetrically
reconstructed
data
andththe
ground
coordinate
from
the
photogrammetric
data
are
shown
in
Figure
1.
photogrammetrically
reconstructed
data
and
the
ground
coordinate
from
fromthe
thephotogrammetric
photogrammetricdata
dataare
areshown
shownininFigure
Figure1.1.
is thecoordinate
scale factor
between either the i TLS scan or
thethe
photo
and the
the ground
ground
systems;
and
grou
and
coordinate
systems;
and
coordinate
andground
the ground
systems;
systems;
is thecoordinate
scalesystems;
factor
between either the ith TLS scan or the p
- the
and
the
ground
coordinate
systems;
3.
3.
Similarity
measure
and
quality
control
3.Similarity
Similaritymeasure
measureand
andquality
qualitycontrol
control
is the rotation matrix relating eiis the
theith TLS
rotation
matrix
relating eith
eith
and the ground
systems;
- relating
⌊⌊⌊ ⌊ coordinate
⌋⌋⌋ther
is
--scan
ormatrix
the phois⌋ the
the
rotation
matrix
relating
eith
is rotation
the
rotation
matrix
relating
To
To
estimate
the
transformation
parameters
between
the
phoToestimate
estimatethe
thetransformation
transformationparameters
parametersbetween
betweenthe
thephophotogrammetrically
reconstructed
⌋ is the rotation matrix relating eith
- - R scani /PRD) = ⌊
data and the ground coordinate
togrammetrically-reconstructed
togrammetrically-reconstructed
data
(PRD),
the
TLS
scans,
togrammetrically-reconstructeddata
data(PRD),
(PRD),the
theTLS
TLSscans,
scans, photogrammetrically
reconstructed
data
and
the
ground
coordinate
syste
⌊ reconstructed
⌋ data
is defined
the
rotation
matrix
relating
photogramm
photogrammetrically
and
coordinate
syste
systems,
by ground
the
reconstructed
data
and the
the
ground
coordinate
syst
photogrammetrically
reconstructed
data
and
theangles
ground
coordinate
and
and
the
ground
coordinate
system,
similarity
measure
andthe
theground
groundcoordinate
coordinatesystem,
system,aaasimilarity
similaritymeasure
measureisisis photogrammetrically
Ω,
Ф,
and
Κ;
and
photogrammetrically
reconstructed
data
and
the
ground
coordinate
syste
and
Κ;
and
and Κ; and
used.
and
used.
Since
planar
and
linear
features
are
represented
by
used.Since
Sinceplanar
planarand
andlinear
linearfeatures
featuresare
arerepresented
representedby
by and
andΚ;
Κ;
and
and
Κ; and
photogrammetrically
reconstructed data and the ground coordinate
points,
a
point-based
3D
similarity
transformation
is
chopoints,
a
point-based
3D
similarity
transformation
is
choand
Κ;
and
points, a point-based 3D similarity transformation is chois the ground coordinates of the corresponding point in
⃗
sen
and
sen
for
relating
the
observed
conjugate
features
the
PRD,
[ Κ;
] is
the
ground
coordinates of
of the
the corresponding-point
point⃗ in
in the
the
obj
senfor
forrelating
relatingthe
theobserved
observedconjugate
conjugatefeatures
featuresinininthe
thePRD,
PRD,
[ obj
]
the
object
space.
coordinates
--- - ⃗-⃗XG =and
isthe
the
ground
coordinates
of theofcorresponding
corresponding
pointpoint
in the
⃗[[ ]][is
] isground
the
ground
coordinates
the corresponding
inob
th
the
TLS
scans,
and
the
ground
coordinate
system
(Equation
the
theTLS
TLSscans,
scans,and
andthe
theground
groundcoordinate
coordinatesystem
system(Equation
(Equation
⃗
[ ] is the ground coordinates of the corresponding point in the obj
1).
1).
One
should
note
that
3D
similarity
transformation
1).One
Oneshould
shouldnote
notethat
thataaa3D
3Dsimilarity
similaritytransformation
transformationisisis ⃗
[ ] is the ground coordinates of the corresponding point in th
commensurate
commensurate
with
the
transformation
parameters
(i.e.,
commensurate with
with the
the transformation
transformation parameters
parameters (i.e.,
(i.e., As
As
mentioned
earlier,
the
points
representing
corresponding
Asmentioned
mentionedearlier,
earlier,the
thepoints
pointsrepresenting
representingcorresponding
corresponding
three
three
shifts,
three
rotations,
and
scale
factor)
relating
the
threeshifts,
shifts,three
threerotations,
rotations,and
andaaascale
scalefactor)
factor)relating
relatingthe
the features
features
from
the
photogrammetric
model
and
the
TLS
scans
featuresfrom
fromthe
thephotogrammetric
photogrammetricmodel
modeland
andthe
theTLS
TLSscans
scans
reference
reference
frames
of
the
involved
data.
Although
that
there
referenceframes
framesof
ofthe
theinvolved
involveddata.
data.Although
Althoughthat
thatthere
thereisisis are
are
not
necessarily
conjugate.
Therefore,
the
traditional
LSA
arenot
notnecessarily
necessarilyconjugate.
conjugate.Therefore,
Therefore,the
thetraditional
traditionalLSA
LSA
no
no
scale
difference
among
the
TLS
data
and
that
its
scale
procedure
modified
by
changing
the
weights
of
the
error
noscale
scaledifference
differenceamong
amongthe
theTLS
TLSdata
dataand
andthat
thatits
itsscale
scaleisisis procedure
procedureisisismodified
modifiedby
bychanging
changingthe
theweights
weightsof
ofthe
theerror
error
correctly
vector
associated
with
the
observation
equations
(Equation
correctly
set
by
the
measured
ranges,
the
scale
used
this
vectorassociated
associatedwith
withthe
theobservation
observationequations
equations(Equation
(Equation
correctlyset
setby
bythe
themeasured
measuredranges,
ranges,the
thescale
scaleisisisused
usedinininthis
this vector
1).
The
conceptual
basis
of
the
weight
modification
is
1).
The
conceptual
basis
of
the
weight
modification
is
min1).
The
conceptual
basis
of
the
weight
modification
istototominminmodel
to
allow
for
the
scaling
of
the
photogrammetric
model
model
to
allow
for
the
scaling
of
the
photogrammetric
model
model to allow for the scaling of the photogrammetric model
imize
the
normal
distances
between
conjugate
features
rather
imize
the
normal
distances
between
conjugate
features
rather
imize
the
normal
distances
between
conjugate
features
rather
tototofit
the
TLS
data
since
the
scale
is
arbitrarily
established
in
fit
the
TLS
data
since
the
scale
is
arbitrarily
established
in
fit the TLS data since the scale is arbitrarily established in
than
than
minimizing
the
distances
between
non-conjugate
points
thanminimizing
minimizingthe
thedistances
distancesbetween
betweennon-conjugate
non-conjugatepoints
points
the
the
3D
model
derivation
process
from
imagery.
the3D
3Dmodel
modelderivation
derivationprocess
processfrom
fromimagery.
imagery.
Figure
Figure
Examples
automatically
derived
linear
and
planar
features
from
TLS
scan
(a)
and
the
manually
extracted
corresponding
features
Figure1:1:1:Examples
Examplesofofofautomatically
automaticallyderived
derivedlinear
linearand
andplanar
planarfeatures
featuresfrom
fromaaaTLS
TLSscan
scan(a)
(a)and
andthe
themanually
manuallyextracted
extractedcorresponding
correspondingfeatures
featuresinininaaa
photogrammetric
photogrammetric
data
(b)
photogrammetricdata
data(b)
(b)
Sibel Canaz, Ayman Habib / Vol.2  Iss.1  2013 5
along corresponding linear or planar features. The modification process is established by assigning zero weights to
points along the directions of the enclosing linear or planar
feature. Thus, only the weights in the normal direction to the
linear or planar feature are considered in the least squares
minimization process. Therefore, the transformation parameters relating the different scans and the photogrammetric
model are estimated to minimize the normal distances between conjugate features in the sub-models. For more details
regarding the weight modification process, interested readers can refer to Renaduin et al. (2011), Habib et al. (2011),
and Kersting (2012). Since the photogrammetric features are
referenced to a unique reference frame (i.e., the arbitrarily-chosen reference frame for the bundle adjustment), the
registration process will implicitly align the different TLS
scans to the photogrammetric model to form a single model
in the photogrammetric reference frame while simultaneously scaling the photogrammetric model to fit the TLS data.
Then, the augmented model is related to the ground coordinate system through the provided control. In this research,
the reference frame for one of the TLS scans, which is denoted as the reference scan, is chosen as the ground coordinate
system. Therefore, the least squares adjustment process will
align the different scans and the photogrammetric model until they fit as well as possible with the corresponding features
in the reference scan.
To test the feasibility of the proposed procedure, the registered scans are qualitatively and quantitatively evaluated.
Qualitative quality control is conducted by transforming all
of the TLS scans to the reference coordinate system using
the estimated transformation parameters and plotting them
together. Then, the quality of registration is evaluated by visually checking the closeness of the point clouds from the
different scans in the overlap area. In addition, quantitative
quality control is performed using either one of the following approaches. First, the point-to-plane normal distances
among conjugate planar patches are evaluated and analyzed.
The point-to-plane normal distances are evaluated for overlapping regions between the PRD and TLS scans and/or
overlapping regions between the TLS scans, if available. In
the second quantitative quality control approach, the estimated registration results are compared with those derived from
the ICPP registration, which is a variant of the ICP approach.
It is worth mentioning that additional TLS scans are used
for the ICPP registration since it requires large overlap areas
among the involved scans. The next section deals with the
conducted experimental results to highlight the capabilities
of the proposed procedure as well as the comparative performance of planar and linear features as registration primitives.
the results of the registration methods are presented. Four
minimally overlapping TLS scans were collected around the
Rozsa Center at the University of Calgary using the Trimble
GS200 scanner, whose maximum range and angular resolution are 200 m and 32 milli-radians (or 3 mm at 100 m), respectively. As mentioned before, the ICPP-based registration
requires large overlap area among the TLS scans. Therefore,
two additional TLS scans and an airborne scan, which have
large overlap areas with the other minimum overlapping four
scans, were added for the ICPP registration approach. For
the photogrammetric reconstruction, sixteen images of the
Rozsa Center are collected using a Canon EOS Rebel XS
camera with a 0.00571 mm pixel size. The camera has an
array dimension of 3888×2592 pixels and an 18 mm nominal
focal length.
4.1 Planar feature-based registration
For this experiment, 30 planar features within the Rozsa
Center are identified and extracted/derived from the TLS
scans and PRD. These features are used for the proposed
registration method, whose results are listed in Table 1. The
values between the parentheses are the standard deviation
of the estimated parameters. One should note that the coordinate system of TLS scan 1 has been used as the ground
(reference) coordinate system.
4.2 Linear feature-based registration
For this experiment, 20 different linear features within the
Rozsa Center are derived from the TLS scans and PRD, and
they are used for the proposed registration method. The estimated transformation parameters together with their standard deviations from this experiment are shown in Table
2. Comparing the estimated parameters and their standard
deviations to those in Table 1, one can hypothesize that the
linear features did not provide a registration quality of the
same level as the planar features. As can be seen in Table 2,
the standard deviation of the vertical displacement for TLS
3 and TLS 4 as well as the Ω and Ф rotation angles for TLS
2, TLS 3, and TLS 4 are significantly higher than those in
Table 1 (refer to the highlighted cells in Table 2). The cause
of the inferiority of the linear feature-based registration will
be discussed in the quality control section of the experimental results.
4.3 Quality control and comparison of the
proposed planar and linear registration
methods
In order to analyze the ability of the proposed registration
methods, real experimental data have been collected and
The quality control of the proposed linear and planar feature-based registration procedures are conducted through
both qualitative and quantitative inspection of the co-aligned
scans. For the qualitative quality control, the registered TLS
Table 1: Results of the planar feature-based registration procedure
Table 2: Results of the linear feature-based registration procedure
4. Experimental results
XT (m)
YT (m)
ZT (m)
TLS
scan1
0
0
0
TLS
scan2
-23.186
(±0.0287)
-14.801
(±0.0219)
-0.687
(±0.0805)
TLS
scan3
69.677
(±0.0221)
92.511
(±0.0293)
1.335
(±0.0431)
TLS
scan4
-41.693
(±0.0298)
91.370
(±0.0234)
-0.251
(±0.0916)
PRD
5.372
(±0.0184)
1.610
(±0.0163)
37.383
(±0.0143)
Scale
Ω (o)
Ф (o)
XT (m)
o
Κ( )
YT (m)
Ω (o)
Ф (o)
Κ (o)
ZT (m)
Scale
1
0
0
0
1
0.869
(±0.2571)
1.073
(±0.5158)
8.421
(±0.1569)
0.681
(±0.4161)
0.824
(±0.5422)
121.411
(±0.1071)
1
0
0
0
TLS
scan1
0
0
0
1
0.234
(±0.1133)
-0.429
(±0.2092)
8.373
(±0.0484)
TLS
scan2
-23.217
(±0.0685)
-14.792
(±0.0341)
-0.667
(±0.1226)
0.243
(±0.0624)
0.313
(±0.0781)
121.373
(±0.0441)
TLS
scan3
69.639
(±0.0661)
92.565
(±0.0646)
1.284
(±0.7231)
1
-0.291
(±0.1273)
0.165
(±0.0769)
-145.531
(±0.0447)
TLS
scan4
-41.787
(±0.0983)
91.305
(±0.0724)
-0.751
(±0.3747)
1
-0.803
(±0.4175)
0.009
(±0.2541)
-145.441
(±0.1263)
0.998
(±0.000)
30.584
(±0.2271)
-74.546
(±0.0419)
91.168
(±0.2261)
PRD
5.461
(±0.0722)
1.635
(±0.0393)
37.384
(±0.0336)
0.998
(±0.0006)
31.338
(±0.8272)
-74.372
(±0.1408)
91.914
(±0.7643)
1
1
6 Photogrammetric features for the registration of terrestrial laser scans with minimum overlap
Figure 2: Registered scans using the planar feature-based (a) and linear feature-based registration approaches (b)
scans are plotted and a visual inspection is conducted in the
overlap area. Figure 2 illustrates a portion of the registered
scans from both procedures. From that figure, it is quite
evident that the planar feature-based registration is showing better quality (Figure 2a) while the linear feature-based
registration exhibits a problem in the vertical alignment of
the different scans (Figure 2b); please refer to the zoomedin portion corresponding to the lightening rod on top of the
building rooftop, where the involved three scans are properly
aligned in Figure 2a while a misalignment in the Z-direction
is visible in Figure 2b. Such misalignment can be either attributed to improperly estimated displacement in the Z direction or the Ω and Ф rotation angles relating these scans
from the linear feature-based registration. This observation
is in agreement with the large standard deviations in Table 2
(please, refer to the highlighted cells in this table).
Quantitative quality control is performed by calculating
the point-to-plane normal distances for the planes between
the PRD and the TLS scans (i.e., quantitatively showing the
compatibility of the photogrammetric and TLS data following the registration process). Using the estimated transformation parameters, the PRD and the TLS scans are transformed into the reference frame coordinate system. For the
point-to-plane normal distance estimation, which is the first
quantitative quality control procedure, the plane parameters
are determined from the PRD and the normal distances from
the individual points in the segmented TLS scans to the PRD
plane are then estimated. Finally, a visual illustration of the
calculated point-to-plane normal distances is presented as
can be seen in Figures 3 and 4. For the planar feature-based
registration, the average normal distance turned out to be at
the 10 cm level. On the other hand, the linear feature-based
registration exhibited an average normal distance of up to 20
cm for horizontal and sloping planar features (i.e., building
rooftops in Figure 3). It is worth noting that these planes are
the ones that verify the quality of the estimated ZT shift as
well as the Ω and Ф rotation angles between the different
scans. Once again, these parameters are the ones that exhibit
large standard deviations as can be seen in Table 2. The reason behind this problem will be explained in the next paragraph. For the second quantitative quality control procedure,
a comparison of the results from the proposed registration
with those derived from the ICPP is presented in Table 3,
where large differences between the results from the different approaches are highlighted in yellow. Once again, the
parameters that are showing the largest deviations are ZT
shift as well as the Ω and Ф rotation angles for the linear
feature-based registration.
Figure 3: Point-to-plane normal distances using the linear feature-based
registration method (color is based on normal distances as seen in the
scale bar on the right)
Figure 4: Point-to-plane normal distances using the planar feature-based
registration method (color is based on normal distances as seen in the
scale bar on the right)
Sibel Canaz, Ayman Habib / Vol.2  Iss.1  2013 7
Table 3: Comparisons of the results from the ICPP and the proposed registration methods
Registration
Parameters
TLS scan2
TLS scan3
TLS scan4
PRD
XT (m)
YT (m)
ZT (m)
Scale
Ω (o)
ф (o)
Κ (o)
XT (m)
YT (m)
ZT (m)
Scale
Ω(o)
ф(o)
Κ(o)
XT (m)
YT (m)
ZT (m)
Scale
Ω(o)
ф(o)
Κ(o)
XT (m)
YT (m)
ZT (m)
Scale
Ω(o)
ф(o)
Κ(o)
ICPP
Method
-23.258
-14.733
-0.605
1
0.177
-0.189
8.168
69.613
92.579
1.375
1
0.095
0.194
121.388
-41.751
91.313
-0.259
1
-0.204
0.002
-145.53
-
Planar featurebased method
Linear feature-based
method
-23.186
-14.801
-0.687
1
0.234
-0.429
8.373
69.677
92.511
1.335
1
0.243
0.313
121.373
-41.693
91.37
-0.251
1
-0.291
0.165
-145.531
5.372
1.611
37.383
0.998
30.584
-74.546
91.168
-23.217
-14.792
-0.667
1
0.869
1.073
8.421
69.639
92.565
1.284
1
0.681
0.824
121.411
-41.787
91.305
-0.751
1
-0.803
0.009
-145.441
5.461
1.635
37.384
0.998
31.338
-74.372
91.914
Further investigation of the problematic linear feature-based
registration revealed the cause of observed anomalies in
Figures 2 and 3 as well as Tables 2 and 3. As mentioned earlier, the registration approach assumes that the derived features
from both the photogrammetric data and the TLS scans are
conjugate. While this assumption is true for the planar features, it is not necessarily valid for the linear features. More
specifically, the extracted linear features from the TLS scans
might not correspond to physical linear features that could be
ICPP vs. Planar
feature-based
methods
-0.072
0.067
0.082
0
-0.056
0.239
-0.204
-0.063
0.068
0.040
0
-0.147
-0.118
0.015
-0.057
-0.056
-0.008
0
0.086
-0.162
-0.007
-
ICPP vs. linear
feature-based
methods
-0.041
0.058
0.062
0
-0.691
-1.262
-0.252
-0.025
0.014
0.091
0
-0.585
-0.629
-0.022
0.036
0.008
0.491
0
0.598
-0.006
-0.097
-
Linear vs.
planar
0.031
-0.009
-0.020
0
-0.635
-1.502
-0.048
0.038
-0.054
0.051
0
-0.438
-0.511
-0.038
0.094
0.065
0.500
0
0.512
0.156
-0.09
-0.089
-0.025
-0.001
0
-0.754
-0.174
-0.746
identified in the photogrammetric data. Figure 5 illustrates an
instance of such a problem. This type of problem generally
occurs in the roof portion of the scan, which commonly has
sloping or horizontal linear features. The non-corresponding
sloping or horizontal linear features from the PRD and the
TLS scans are the main cause for the incorrect estimation of
ZT, Ω, and Ф since these features are the ones contributing
to the estimation of these parameters in the least squares adjustment procedure.
Figure 5: Problem with the linear feature-based registration approach (i.e., TLS linear features that do not correspond to those identifiable in the
photogrammetric data)
8 Photogrammetric features for the registration of terrestrial laser scans with minimum overlap
5. Conclusions and recommendations for
future work
In this research, a new registration method for the co-alignment of TLS scans is proposed. While commonly used
registration methods, such as the ICP and its variants, fail
to register TLS scans with minimum overlap area, the proposed registration method has succeeded in registering the
scans using photogrammetric data, which can be quickly and
economically collected, covering the object of interest. The
proposed registration method uses linear or planar features,
which have been extracted from photogrammetric and TLS
data for the estimation of the transformation parameters relating the different scans. In spite of the fact that non-conjugate
points have been used to represent corresponding linear and
planar features, a point-based similarity measure has been
used after modifying the weights of the random errors in the
observation equations to eliminate the systematic component
resulting from the use of non-conjugate points and to minimize the normal distances between corresponding features in
the TLS scans and photogrammetric data.
Qualitative visualization of the registered scans has been
used to verify the performance of the proposed approaches.
In addition, quantitative quality control measures have been
introduced to evaluate the performance of the proposed registration. First, point-to-plane normal distances between the
registered surfaces are calculated. Second, we carried out a
quantitative comparison between the results from the proposed registration approaches and the ICPP method after adding more scans to increase the overlap between the involved
sub-models. According to the quality control analysis, the
planar feature-based registration method gives more reliable
results than those derived from the linear feature-based registration. The reason for the inferior performance of the linear
features is that extracted linear features from the TLS data
do not necessarily correspond to physical linear features that
could be identified in the photogrammetric data. Current and
future research will be focusing on the automated extraction
of the registration primitives from the photogrammetric data.
In addition, the automatic identification of conjugate features
in both photogrammetric and TLS data will be addressed.
Acknowledgement
The Authors would like to acknowledge the financial support
of the Natural Science and Engineering Research Council of
Canada and TECTERRA. In addition, the help of the members of the Digital Photogrammetry Research Group (DPRG)
at the University of Calgary in the photogrammetric and laser
scanning data acquisition is highly appreciated.
References
Al-Durgham M. (2007), Alternative Methodologies for the Quality
Control of LiDAR Systems. M.Sc. Thesis, Department of
Geomatics Engineering, University of Calgary, Canada, pp. 66-68.
Al-Durgham M., Detchev I., Habib A., (2011), Analysis of Two
Triangle-Based Multi-Surface Registration Algorithms of
Irregular Point Clouds, ISPRS Workshop Laser Scanning 2011,
XXXVIII-5/W12, Calgary, Canada, 29-31 August, 2011.
Al-Manasir K., Fraser C. S., (2006), Automatic Registration of
Terrestrial Laser Scanner Data via Imagery. The International
Archives of the Photogrammetry Remote Sensing and Spatial
Information Sciences(IAPRS), Dresden 25-27 September,
35(5), pp. 26-31.
Al-Ruzouq R., (2004), Semi-Automatic Registration of MultiSource Satellite Imagery with Varying Geometric Resolutions,
PhD Thesis, Department of Geomatics Engineering, University
of Calgary, Canada, pp.10-24.
Balzani M., Pellegrinelli A., Perfetti N., Russo P., Uccelli F., Tralli
S., (2002), Cyrax™2500 laser scanner and G.P.S. operational
flexibility: from detailed close range surveying, to urbanscale
surveying. In: Proceedings of International Workshop on
Scanning for Cultural Heritage Recording – Complementing or
Replacing Photogrammetry, Corfu, Greece, 1 – 2 September.
Besl P. J., McKay N. D., (1992), A method for registration of 3-D
shapes, IEEE Transactions on Pattern Analysis and Machine
Intelligence, 14(2), pp. 239–256, doi: 10.1109/34.121791.
Chen Y., Medioni G., (1992), Object modeling by registration of
multiple range images, Image and Vision Computing 10(3), pp.
145-155, doi: 10.1016/0262-8856(92)90066-C.
Dold C., Brenner C., (2006), Registration of Terrestrial Laser
Scanning Data Using Planar Patches and Image Data, IAPRS
Volume XXXVI, Part 5, Dresden 25-27 September 2006.
Habib A., Ghanma M. S., Tait M., (2004), Integration of LIDAR and
Photogrammetry for Close Range Applications, XXth ISPRS
Congress: Proceedings of Commission V, 7 July 2004, Istanbul,
Turkey.
Habib A., Bang K. I., Aldelgawy M., Shin S. W., Kim K. O., (2007),
Integration of Photogrammetric and LIDAR Data in a MultiPrimitive Triangulation Procedure, Conf. American Society
Photogrammetry Remote Sensing Conference, Tampa, Florida,
May 7-11 2007.
Habib A., Kersting A. P., Ruifanga Z., Al-Durgham M., Kim C.,
and Lee D. C., (2008), Lidar Strip Adjustment Using Conjugate
Linear Features in Overlapping Stripts, The International
Archives of the Photogrammetry, Remote Sensing and Spatial
Information Sciences, XXXVII. Part B1, Beijing 2008, pp. 386390.
Habib A., Kwak E., Al-Durgham M., (2011), Model-Based
Automatic 3D Building Model Generation by Integrating LiDAR
and Aerial Images, Archives of Photogrammetry, Cartography
and Remote Sensing, 22, pp. 187-200
Kersting A. P. B., (2012), Quality Assurance of Multi-Sensor
Systems, PhD Thesis , Department of Geomatics Engineering,
University of Calgary, Canada, pp. 15-160.
Lari Z., Habib A., Kwak E., (2011), An Adaptive Approach for
Segmentation of 3D Laser Point Cloud, ISPRS Workshop Laser
Scanning 2011, Calgary, Canada, 29 – 31 August 2011.
Lari Z., Habib A., (2012), A new approach for segmentation of
multi-resolution and multi-platform LiDAR data, Proceedings
of Global Geospatial Conference , Québec City, Canada, 14-17
May 2012.
Renaudin E., Habib A., and Kersting A. P., (2011), Featured-Based
Registration of Terrestrial Laser Scans with Minimum Overlap
Using Photogrammetric Data. ETRI Journal, 33(4), pp. 517527, doi: 10.4218/etrij.11.1610.0006.