SEMI AUTOMATIC REGISTRATION OF LASER SCANNER DATA The follow
observatior
parameters
S.T. Dijkman, F.A. van den Heuvel the parame
section des
; : . In th
Delft University of Technology, Department of Geodesy scams WC
Thijsseweg 11, 2629JA Delft, The Netherlands
Email: {S.T.Dijkman, F.A.vandenHeuvel} @citg.tudelft.nl 3. FITT
CommissionV, WG V/2 Before the
need to be
KEYWORDS: laser scanning, industrial photogrammetry, registration, augmented reality, orientation, reconstruction es
designed in
ABSTRACT Ba i
distances fi
Laserscanning is a technique that is becoming very popular for the acquisition of 3D models, especially in industrial envionments. adjusting t
As industrial sites are in many cases complex environments, scans from multiple viewpoints have to be taken to be able to model the
entire site. Before modelling the site can start, registration of all the scans is required. This paper discusses he semi automatic grit Po
registration of laser scanner data. Starting point for the registration are object models measured in two scans taken from different caleulate h
positions. The fitting algorithms to find the parameters of the models are based on nonlinear least square adjustments. The method offer hi : fe
used is not bound to the fitting of a particular model but can be used for different kind of models. The parameters are estimated from The specifi
the distances from the points to the model. As the equations are nonlinear initial values for the parameters are needed. A method to
find these for cylinders and planes is discussed. Based on the parameters of the models measured in different scans the registration is
performed. Assuming neighbouring laser scans to be taken from one sit of an objæt no approximated values are needed.
different m
with respec
Section 2.1
, p plane. As tl
1. INTRODUCTION To be able to model sites from a combination of range data for the pose
) and visual imagery these data sets have to be registered. In this section: as
In the process industry there is a need to obtain accurate three- paper a technique is presented that is capable of doing this. equations 1
dimensional models. Those models are for example needed in The method uses corresponding objects measured in different method to
case of extensions of the existing infrastructure or for scans to determine the transformation parameters. In case parameters
Augmented Reality (AR) applications. AR allows the user to objects are measured in images, it is possible to register the
create realistic simulations of the real world by blending real images as well. In the first step planes or cylinders are fitted in
and virtual environments. Potential applications of AR are different laser scans. Secondly an operator assigns 2.1 Plane
training, documentation, planning and maintenance correspondences between the measured objects. The last step
applications. The research at the Delft University of involves the registration of the scans. In close rar
Technology (DUT) focuses on the 3D-reconstruction of fact that pla
industrial sites using a combination of close range imagery and Much work has already been done on the registration of laser parameteris
range data. The research is part of an interdisciplinary
European Union project called STAR (Services and Training
scan images. The iterative closest point (ICP) algorithm (Besl,
McKay, 1992) establishes a set of corresponding points from
be seen as 2
through Augmented Reality, www.realviz.com/STAR). two data sets. From this set of comesponding points a Is importan
transformation is computed. Drawbacks of this method are that perpendicul
In recent years DUT developed a method to measure three it isn't obvious how to handle multiple data sets and that one provides a
dimensional industrial sites from a collection of close range point set is assumed to be a subset of the other. That is it plane can b
images (Ermes, 1999). This method assumes a predefined
library of parameterised object models defined by
Constructive Solid Geometry (CSG) trees is used to describe
the site. Each particular object model is projected onto the
images. In the first step an operator alters the pose and shape
parameters of the model in order to obtain approximate values
for the object parameters. The projected CSG model is aligned
manually with the object as seen in the image.
Due to technical improvements of close range laser scanners it
is nowadays feasible to obtain point clouds with accuracies
better than 5mm with distances of up to 50 meters. Using laser
scanning alone is not feasible in case very high accuracies are
desired. Edges cannot be precisely located because of the
distance between neighbouring laser points. Images are used to
overcome this problem. The grey values in images provide
information about the locations of discontinuities in a scene.
Furthermore visual imagery can be mapped onto reconstructed
3D models to create a visually realistic model (Stamos &
Allen, 2000).
assumes the same points are measured in both scenes, which is
not the case with laser scanning furthermore occlusions cannot
be dealt with.
Zhang (Zhang, 1992) exploits the same idea as ICP but can
also handle data sets that are not subsets of one another. This
method detects wrong correspondences by dynamically
adjusting the distance threshold between points throughout the
iterations. Later, algorithms were not designed to establish
point-to-point relationships, but to establish the relationship
between points in one set to locations on the surfaces
represented by the points of the other data set (Chen and
Medioni, 1992; Dorai et al., 1997). The techniques described
all assume the registration of two images whereas (Eggert et
al., 1998) is designed to register more scans simultaneously.
None of the techniques described above, however, is designed
to register range data and image data simultaneously, although
some work has been done on the registration of 2D data and
3D data. (Phong et al., 1995; Kumar & Hanson, 1994)
=12-
signed dista
Figure 1. Ps
€
