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
IMAGE TO POINT CLOUD METHOD OF 3D-MODELING
A.G. Chibunichev, V.P. Galakhov.
Moscow State University of Geodesy and Cartography, #4, Gorokhovsky pereulok, 105064,
Moscow, Russia. agchib@mail.ru, vpgal@mail.ru
KEY WORDS: Photogrammetry, Laser scanning, Photography, Three-dimensional, Fusion, Modeling.
ABSTRACT:
This article describes the method of constructing 3D models of objects (buildings, monuments) based on digital images and a point
cloud obtained by terrestrial laser scanner. The first step is the automated determination of exterior orientation parameters of digital
image. We have to find the corresponding points of the image and point cloud to provide this operation. Before the corresponding
points searching quasi image of point cloud is generated. After that SIFT algorithm is applied to quasi image and real image. SIFT
algorithm allows to find corresponding points. Exterior orientation parameters of image are calculated from corresponding points.
The second step is construction of the vector object model. Vectorization is performed by operator of PC in an interactive mode using
single image. Spatial coordinates of the model are calculated automatically by cloud points. In addition, there is automatic edge
detection with interactive editing available. Edge detection is performed on point cloud and on image with subsequent identification
of correct edges. Experimental studies of the method have demonstrated its efficiency in case of building facade modeling.
1. INTRODUCTION
One of the most frequent tasks for land survey is high detailed
survey of building facades, monuments and objects of cultural
heritage. The purpose of this survey is to obtain detailed model
of object’s interior and exterior. Several different kinds of
survey can be used to obtain model satisfying these demands. In
this paper we’ll discuss two methods: photogrammetric
processing of digital images and terrestrial laser scanning with
further modeling. These methods could be compared in
application to facade modeling.
The process of image capturing is simple and fast, though needs
planning of total image captures and their relative positions.
Absolute and relative tie points are needed to create
photogrammetric model, therefore additional method of survey
is used to achieve coordinates of absolute tie points. Quality and
precision of modeling based on photogrammetric processing of
photographic images in high degree depends on experience and
qualification of PC operator and furthermore on his
physiological abilities.
Terrestrial laser scanning needs much more time for capturing
spatial data, because a laser scanner needs to be placed on a
tripod, prepared for scanning and only after that scanning could
be started, which is not as fast as simple image capturing.
Absolute tie points are needed for increasing of total model
precision or for referencing to external coordinate system, but
total quantity of tie points is less then for photogrammetric
model. Precision and quality of modeling by the point cloud
much less depends on qualification and experience of PC
operator. But operator should perform several different kinds of
manipulations with parameters of point cloud projection onto
2D surface of screen (rotation, translation, scaling) before
operator get concept of object in his imagination.
Practical experience shows advantages and disadvantages of
these methods (Fabris, Achilli, Artese 2009).
Cost of good enough digital camera is low comparing to
terrestrial laser scanning system and obviously camera can be
added into kit. Using combination of these two methods
complementing each other allows resulting model quality
increasing and reducing time efforts (Jansa J., Studnicka N.,
Forkert G., Haring A., Kager H., 2004).
These two methods could be combined on several different
levels.
The first, easiest, level of combination is coloring of cloud
points according to digital image colors. Naturally laser scanner
is capturing point coordinates and intensity of returned laser
beam (Chunmei Hu, Yanmin Wang, Wentao Yu, 2008).
Coloring of cloud points makes easier process of data
perception by PC operator.
The next level is creation of the digital surface model from
cloud points, orthophototransformation of digital images and
surface texturing. Efficiency of this level is higher.
Furthermore, operator can vectorize orthophoto in 2D and then
project vectors onto 3D, because the surface model can be used
as DTM for orthophoto (Ayman Zureiki, Michel Roux, 2009).
Lines, extracted automatically from orthophoto, could also be
projected onto DTM surface. But the quality of projecting
depends on quality of created surface model. In case of
terrestrial laser scanning captured point clouds contain many
“noise” points due to great number of obstructions in natural
environments. Point clouds needed to be filtered and cleaned
from “noise” points before surface modeling. Every “noise”
point leads to incorrect surface modeling and thus to incorrect
orthophototransformation of digital images. But filtering is not
enough. Breaklines should be created for accurate surface
modeling. All these operations need labor input from operator.
New methods of modeling developed to solve these problems.
Nex and Rinaudo (2009) proposed new approach of automatic
modeling using digital images and point cloud processing. On
the first step bundle block adjustment and edge extraction
performed for digital images. After that edge extraction
performed for point cloud using extracted image edges.
New method of vectorization proposed in this article. The
method is based on vectorization by single untransformed
images. Automatic image and point cloud segmentation is
performed. Segments are used for edge extraction. Extracted
edges are transposed onto images and onto quasi images
generated from point cloud. SIFT operator used for tie point
extraction. Tie points are used for calculation of exterior
orientation parameters of images. After image referencing edges
could be projected onto image plane. Matching edges are left as
vectors. Operator of PC manually creates new vectors by single