Professional Documents
Culture Documents
net/publication/260905377
Article in Proceedings of SPIE - The International Society for Optical Engineering · December 2004
DOI: 10.1117/12.586294
CITATIONS READS
68 3,692
3 authors:
Antonio Vettore
University of Padova
104 PUBLICATIONS 1,082 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Fabio Remondino on 11 April 2014.
ABSTRACT
Photogrammetry has dealt since many years with the 3D reconstruction of objects from images. It provides for accurate
sensor calibration and object modeling using analog or digital imageries, it is very portable and many commercial
software is available for image processing and 3D modeling. On the other hand, laser scanning technology and all the
related reverse engineering software are becoming a very promising alternative for many kind of surveying and
modeling applications. Laser scanners allow to acquire very quickly a huge amount of 3D data which can be often
combined with color high-resolution digital images.
Among the plenty of works so far presented, in particular on the use of laser scanning for cultural heritage survey, some
modeling and accuracy related issues have been not yet solved and discussed in details. In this contribution we report
about two case studies realized with photogrammetry and laser scanner and we provide some advices and suggestions
about the more suitable 3D modeling method for a given object, taking into account its size and shape complexity, the
required accuracy and the target application.
Key words: Adjustment, Reconstruction, Modeling, Laser Scanner, Registration
1. INTRODUCTION
2. TECHNOLOGY AND INSTRUMENTS
2.1. Laser Scanner
Laser scanning is an active sensor approach able to recover directly 3D measurements of the scanned scene. Structured,
coded or laser light can be used and on the market different commercial solutions are available, based on triangulation,
time-of-flight, continuous wave or reflectivity measurements. An overview of 3D scanning instruments is presented in
Boehler et al.5 or Blais4. The scanner instruments are very expensive, but they provide for a huge amount of data (3D
point clouds) in few moments. The often need special or dedicated software to read the output data file. The
manufacturers usually provide for this software while other packages9, 16, 17 are also available for the modeling,
registration, editing and texturing of the 3D data.
For our case studies, two different systems were used: the Leica HDS 2500 (former Cyrax 2500)14 and the NRC-BIRIS
laser scanner2 (Figure 1).
Figure 1: The two laser instruments used in the presented case studies: Leica HDS 2500 (left), NRC-BIRIS (right).
The Leica HDS 2500 system is a TOF-based terrestrial laser scanner, providing a quite high accuracy for the depth
measurement (4mm up to 50m) with a beam diameter divergence of 6 mm at 50m. Given the limited Field Of View
(FOV= 40∞ H x 40∞ V) and the absence of a rotating head, several stations are normally required to capture the whole
geometry. Besides 3D coordinates, the laser is also able to register the reflected beam intensity and color information,
through a built-in low resolution CCD camera.
The BIRIS laser scanner is a non-contact measuring system suited for close range applications. It is based on the
triangulation and defocusing principle, by which 3D object coordinates can be measured knowing the baseline between
a laser diode and the built-in CCD camera and the position on the CCD image plane where the reflected beam intersects.
The employed model features a depth accuracy of 50µm when the laser-object distance is kept between 20cm and 40
cm. The laser wavelength is about 650nm, i.e. in the visible red field.
2.2 Image-based modeling
Photogrammetry, by definition, obtains reliable measurements and 3D models by means of photographs. It deals since
many years with the 3D reconstruction of objects from images. The classical close-range photogrammetric pipeline
consists of camera calibration and orientation, image point measurements, 3D point cloud generation, surface generation
and texturing. Even if precise calibration and orientation procedures are mostly required, reliable packages are now
available1, 12, 15, 19. Some packages allow the complete scene modeling, from single or multi images, while others
perform only sensor calibration and orientation. Generally a sensor calibration is performed with a perspective
geometrical model by means of a photogrammetric bundle method. It requires corresponding points between the images
and, as non-linear algorithm, it needs initial approximations for the unknownsí parameters.
The image measurement step can be performed with manual or automatic procedures. Automated photogrammetric
matching algorithms can produce very dense point clouds, but mismatches, irrelevant points and missing parts could be
present in the results, requiring a post-processing check of the data. Moreover, the automated procedures usually do not
take into consideration the geometrical conditions of the surface's object and mainly work with smoothing constraints:
therefore is often quite difficult to turn randomly generated point clouds into correct polygonal structures of high quality
and without losing important information or features. On the other hand, if the measurements are done in manual or
semi-automatic mode, there is a higher reliability of the measures but a smaller number of points that describe the
measured object.
For our case studies, two digital cameras are employed: a Nikon COOLPIX 5700 and a LEICA DIGILUX 1. The
NIKON COOLPIX is a 5 Million pixel camera with a 2/3" CCD sensor size and a maximum image resolution of 2560 x
1920 pixels, corresponding to 3.4 µm pixel size. The LEICA camera has a 1/1.76î CCD sensor, with 4 Million pixel.
The maximum image resolution is 2240x1680 pixel, corresponding to 3.21 µm pixel size.
3. 3D MODELING OF A CHURCH
The first case study is the ancient church of Pozzoveggiani, located in the surrounding of Padua (Italy). The church
(Figure 2) is 7 x 16 m large in the planar dimensions while the height is between 8 m (main roof) and 17 m (the tower).
This little church boasts a geometry composed by both planar and curved surfaces, allowing to compare easily the
results about the application of different modeling techniques. A clear view of the church is almost available all around
it, thus removing the problem of shadowing effects due to occlusions. All these features made the church a good
candidate for our first test. Some ground control points, used in the photogrammetric process and as common points for
the accuracy comparison between the two modeling approaches, were measured with a Leica TCR 705 total station. The
instrument has a distancer with two laser beams: one, invisible, for conventional measurements to prism and a second
one, visible, for reflectorless surveys. The control points on the church were surveyed in reflectorless mode and about
170 control points were measured, with an accuracy of ca 5 mm.
Figure 2: Some images of the Pozzoveggiani church.
3.1 Image-based reconstruction and modeling
The digital images of the Pozzoveggiani church were acquired with the digital camera NIKON COOLPIX 5700. A set
of ca 80 images was acquired with a fix focal length, with a resolution of 1024 x 768 pixel. For the processing only 22
images were selected. The pixel size of the images is ca 8.6 micron while in object space it is ca 0.95 cm.
The orientation of the images had to be performed in order to extract precise 3D information and measurements. The
exterior orientation approximations were achieved with a space resection algorithm using some control points measured
semi-automatically in the images by means of a least squares template matching. About 200 image correspondences
were then automatically extracted using an approach based on interest operator, least squares matching and the epipolar
geometry18. Some wrong correspondences not related to the church were manually deleted. The correspondences, as
well as 29 control points, were imported in a self-calibrating bundle adjustment for the recovering of the camera interior
and exterior parameters, as well as the 3D coordinates of the tie points. The a posteriori standard deviation was 0.9 pixel
(σ0,priori = 1 pixel) while the 2 additional parameters for tangential lens distortion (P1 and P2) could not be reliable
determined. The theoretical precision of the computed object coordinates was 0.017m, 0.027 m and 0.022 m in X, Y and
Z direction respectively. Figure 3 shows the recovered cameras poses and the 3D objects coordinates of the used tie and
control points. Afterwards, lines, curves and surfaces were manually extracted from the images and the complete 3D
model reconstructed. The results of the modeling are presented as wireframe and colored shaded model in Figure 4 and
as textured model in Figure 5.
Figure 3: Recovered cameras poses with used tie and control points.
An analysis of the recovered object coordinates with respect to some check points measured with the total station was
also performed. 21 check points well distributed all over the church were compared with the recovered object
coordinates and the differences are reported in Table 1.
Figure 4: Wireframe model with points and lines manually measured as well as the recovered camera poses (upper image).
Colored shaded 3D model from different view points (central and right images).
Figure 5: Views of textured 3D model of the Pozzoveggiani church, displayed in VRML format.
3.2 Laser scanner measurement and modeling
The church was surveyed with a Leica HDS 2500 terrestrial laser scanner. The capturing of the geometry of the whole
building required the acquisition of 23 range images, along with their corresponding intensity data. The average scan
resolution (i.e. point spacing) was set to 0.7mm, resulting in global point cloud of about 11 Million points. As a first
stage of the processing workflow, a contrast stretching algorithm, developed in Matlab, was applied to the intensity data,
in order to enhance the viewing quality of the point clouds when rendered with color information from intensity values.
This step was required in order to be able to better recognize matching points needed both for the alignment and the 3D
model georeferencing procedures.
Indeed, during export of acquired of range data from Cyclone software (v. 5.1) in PTX format (readable by
Polyworks16) the original intensity values were shrink in a more little interval, decreasing the overall contrast level, as
shown in Figure 6 (left side). By applying a contrast stretching algorithm, natural features on the wall of the church
were much more recognizable (Figure 6, right side).
Figure 6: Result of contrast stretching algorithm applied to the bell tower of the church.
After that, the scans were imported in Polyworks for the alignment/registration process. This procedure requires the
manual selection of homologues points between the scans while an ICP-based global registration algorithm3 is
afterwards applied in order to get a composite point cloud.
More specifically, a two-steps registration procedure was carried out. In the first stage, range data groups were
interactively aligned pair by pair, through manual selection of N matching points on adjacent scan groups. As a result,
an approximate transformation matrix for each scan pair was obtained and then used in the second stage as starting point
for the refined global alignment. In both steps a scan group was locked in order to define the reference frame of the
model. The global registration algorithm provided us with a residual error of 4.3 mm, obtained as RMS of the standard
deviations of the image pair alignments computed by ImAlign. This value confirms the goodness of the registration
procedure implemented in Polyworks: the residual error is due to the inherent accuracy of the laser scanner only.
Figure 7: Two views of the decimated 3D model, rendered as a point cloud with color information from intensity data.
In order to compare the two models (photogrammetry and laser scanning), the aligned point cloud was simplified (i.e.
decimated), deleting most of the overlapping points, which lead to a redundant information that is only useful for the
alignment and the subsequent triangulation step, not for the model georeferencing. In this way, the model size has been
reduced, preserving its geometry and resulting in a smaller and more manageable data file. As a result a global point
cloud of about 9 Million points was obtained. Figure 7 shows the simplified model.
Afterwards, the simplified range data model has been registered in the reference system defined by the total station
survey. A set of control points measured with the total station was identified on the 3D model, by visual inspection of
the operator. At this step, only 58 points could be well recognized, mainly due to the following reasons: 1) most of the
points measured with the total station were placed on corners or edges of the walls, i.e. at critical locations for the laser
scanner, therefore they could not be reliably extracted form the 3D model; 2) some digital images were tilted (i.e. no
vertical imagery), therefore due to the perspective projection it was difficult or impossible to identify the corresponding
points on the model. In practice, the matching procedure was applied only on points lying on the wall surface, apart
from corners or edges easily recognizable. After that, 8 out of the selected 58 control points were used to compute a 3D
similarity transformation (7 parameters), which was then applied to the remaining 50 points. Finally, their transformed
coordinates were compared with the corresponding point coordinates measured by the total station (i.e. check points),
yielding the residuals along the X, Y and Z axes as listed in Table 2.
It should be noted that both the georeferencing and the comparison stages has been performed only on the decimated
point cloud instead of the triangulated model: due to the inherent ì natureî of the meshing process, points are substituted
by triangles and further simplifications of the cloud can occur, i.e. points can be discarded leading to a change respect
with the original data.
4. 3D MODELING OF A SMALL STATUE
The second case study is performed using a small wooden statue (ca 30x40x20 cm), as shown in Figure 8. The object
presents many small details, of approximately 2 cm width. In this case, the study is mainly performed to investigate how
the two techniques can correctly reconstruct the small details of the statue.
Figure 8: The small wooden Buddha statue and two closer views on the detailed headís structures and folds (right)
4.1. Image-based reconstruction and modeling
For this case, the LEICA DIGILUX 1 digital camera was used. Five images, mainly in front of the object were acquired,
with a resolution of 1200x1600 pixel. The pixel size of the images is 0.0045 mm, while in object space is ca 1 mm. For
the orientation of the images, semi-automated image measurements are performed, based on template least squares
matching. The testfield in the background is used as reference, for the datum definition and to get the approximations of
the camera positions. The camera parameters are recovered with a bundle adjustment while the camera interior
parameters were already known from a calibration protocol.
As next step, a least squares matching algorithm is applied on triplets of images to recover a dense and robust set of
corresponding image points7. Some homologues seeds points are manually selected in the images and then the matcher
automatically finds the correspondences in both images. The central image is used as template, the others as search
images. The matcher searches the corresponding points in the two search images independently and at the end of the
process, the data sets are merged to become triplets of matched points. In our application, two triplets of images are used
and approximately 94 910 and 95 190 points are respectively extracted.
Afterwards the three-dimensional coordinates of the matched points are computed by forward intersection, using the
results of the orientation process. Some manual editing was required to eliminate blunders and wrong matches. The
achieved accuracy of the recovered object points is approximately 1.4 mm, 1.3 mm and 1.9 mm in X, Y and Z direction
(depth) respectively. The generated point cloud (Figure 9, center) does not contain all the details of the object, in
particular on the head. These smoothing effects are caused by the automated matching process. In fact, the image
patches of the least squares matching are assumed to correspond to planar object surface patches: along the small folds
of the dress and on the head, this assumption is not valid anymore, therefore these features are smoothed out. Smaller
image patches could theoretically avoid or reduce the smoothing effects, but may be not suitable for the determination
of the matching reshaping parameters, because a small patch may not include enough image signals content (Figure 9,
left).
The recovered point cloud was afterwards triangulated and a surface generated, as shown in Figure 9, right. The holes
are due to lack of texture in the object, leading to incorrect or failed matching. Probably a random pattern projection or
the use of image enhancement filters (e.g. Wallis) could lead to slightly better results.
4.2 Laser scanner measurements and modeling
For the reconstruction of the small Buddha statue, the BIRIS instrument was used. The scanner acquired 700 vertical
profiles (with ca 256 points/profile) with a viewing angle of ca 35 degrees. 54 scans were performed, and ca 4.2 Million
points acquired. The registration of the different point clouds was performed with Geomagic Studio9 software (v. 6.0).
The alignment is based on ICP algorithm3, selecting manually the corresponding points. After the registration process
Figure 9: The matching image patches in one triplet: image patches larger than the object will smooth out the feature (left). The point
cloud recovered with the image-based approach (center) and the generated surface model (right).
and some manual editing, ca 2 Million points were triangulated to generate a surface model of the small object. The
final 3D model, shown in Figure 10, contains approximately 3.4 Million triangles. All the small detail on the head and
the folds of the dress are clearly visible and reconstructed. Compared to Figure 9, right, it is clear the advantage of using
a laser scanner instruments for this kind of application.
Figure 10: Two scans of the statue (left) and the final 3D model displayed in shaded mode (right)
5. CONCLUSIONS
In this paper we reported about the 3D modeling of a church and a small wooden statue using photogrammetry and laser
scanner techniques. In the case of the architectural object, photogrammetry recovered an accurate 3D model, even if the
object edges had to be measured manually. Automated procedures, which can correctly extract object's edges, are not
yet reliable and manual measurements recover always better results. An important issue was the image acquisition: in
our case, the images did not cover all the church, in particular the lower part (near the ground) and the roof, leading to
some holes in the final 3D model. Furthermore, if automated procedures for tie point extraction are used, the baseline
between the images should not be too large. The overall photogrammetric processing time (image acquisition,
processing, modeling and texturing) took about 10 hours. On the other hand, the 3D model obtained from the range data
shows that the use of natural targets as checkpoints for model georeferencing does not guarantee to get highly precise
results, though a good scan alignment can be obtained. The final result is indeed affected by several error sources:
accuracy of the laser scanner, good intensity response from reflected laser beam, grid spacing of the range data set
respect with the digital image resolution, operator skill during identification of model check points, availability of well
signalized points on as much as possible vertical imagery. Given all these factors, it is advisable the use of artificial
targets, when possible, to increase the accuracy on check points selection. Furthermore, such targets could be surveyed
at higher resolution compared to natural corners. However, it should be noted that some laser scanners can show an
unexpected behavior when dealing with retro-reflective targets, which leads to macroscopic measurement errors.
Therefore great care must be taken in choosing the appropriate kind of target. The overall modeling process took 7-8
hours, as we decided to deal with point clouds only: the triangulation step was not considered in this work.
Concerning the second case study (a small Buddha statue), the automated image-based modeling showed problems in
the correct and accurate measurement and reconstruction of the small details, which were smoothed out by the least
squares matching procedure. Similar effects have been discussed in Gruen et al10. Probably higher resolution images
could avoid some of these effects. On the other hand, the laser scanner modeling of the statue could achieve satisfactory
results, even if more that 15 hours were dedicated to the clouds registration, editing and triangulation. The final model
contains more than 3 Million triangles, bringing also problems for the interactive visualization and texturing of the
model.
In conclusion, the choice of the reconstruction and modeling technique depends on different factors and it is often
strictly related to the budget of the project/application. Both modeling processes have their advantages and
disadvantages, as demonstrated in the presented examples. A good solution could be the combination of the methods11,
as each one has attributes and elements that complement one another. Anyway, for both approaches, the modeling part
(from point cloud to surface) is still rich of problems and often the most time consuming.
In general, the time spent for the modeling of range data vary from model to model and it has to be taken into account
too. It depends on several factors like: number of scans to be aligned, object shape complexity, availability of intensity
data, scans provided as binary or ASCII files, computer capability and speed. Though laser scanning systems allow fast
and quite accurate surveys, depending upon the final product required by the user (aligned point cloud only or
triangulated and hole filled model), the processing time can increase considerably compared to other classic surveying
techniques. Concerning photogrammetric methods, automated and reliable procedure for the extraction of image
correspondences and photo-triangulation are already available, but most of the approaches for lines/edges extraction are
still based on manual measurements, as more reliable and accurate.
REFERENCES
1. Australis: http://www.sli.unimelb.edu.au/australis [October 2004]
2. Beraldin, J-A., Blais, F., Cournoyer, L., Rioux, M., Rodella, R., Bernier, F., Harrison, N., Digital 3D Imaging
System for Rapid Response on Remote Sites, IEEE 3DIM Proceedings, 1999
3. Besl, P.J., and McKay, N.D., A method for registration of 3D shapes, IEEE Pattern Analysis and Machine
Intelligence, 14(2), pp. 239-256, 1992
4. Blais, F., A review of 20 years of range sensor development, El-Hakim, Gruen, Walton (Eds), Videometrics VII,
Proc. of SPIE Electronic Imaging, Vol. 5013, pp. 62-76, 2003
5. Boehler, W., Marbs, A., 3D Scanning and Photogrammetry for Heritage Recording: a Comparison, Proc. of 12th
Int. Conf. on Geoinformatics, 2004
6. Bologna R., Guarnieri A., Minchilli M., Vettore A., Automatic registration of 3D views, Proceedings of the ISPRS
Comm. V Symposium ì Close Range Imaging-Long-Range Visionî , 2-6 September, Corf˘ , Greece, 2002
7. DíApuzzo, N., Surface Measurement and Tracking of Human Body Parts from Multi Station Video Sequences,
Diss., Technische Wissenschaften ETH Zurich, Nr. 15271, 2003
8. El-Hakim S. F., 3D Modeling of Complex Environments, Videometrics and Optical Methods for 3D Shape
Measurement, Proceedings of SPIE , Vol 4309, 2001
9. Geomagic Studio ñ Randrop: http://www.geomagic.com [October 2004]
10. Gruen, A., Remondino, F., Zhang, L., Photogrammetric Reconstruction of the Great Buddha of Bamiyan,
Afghanistan, The Photogrammetric Record, 19(107), pp. 177-199, 2004
11. Guidi, G., Beraldin, J-A., Ciofi, S., Atzeni, C., Fusion of range camera and photogrammetry: a systematic
procedure for improving 3D models metric accuracy, IEEE Trans. On System, Man and Cybernetics, 33(4), pp.
667-676, 2003
12. iWiteness: http://www.photometrix.com.au [October 2004]
13. Kadobayashi, R., Kochi, N., Otani, H., Furukawa, R., Comparison and evaluation of laser scanning and
Photogrammetry and their Combined Use for Digital Recording of Cultural Heritage, Int. Arch. of PRS&SIS, Vol.
35(5), Istanbul, Turkey, 2004
14. Leica Geosystems: http://www.leicageosystems.com [October 2004]
15. Photomodeler: http://www.photomodeler.com [October 2004]
16. Polyworks ñ InnovMetric: http://www.innovmetric.com [October 2004]
17. RapidForm ñ INUS Technology: http://www.rapidform.com [October 2004]
18. Remondino, F., 3D reconstruction of static human body with a digital camera, El-Hakim, Gruen, Walton (Eds),
Videometrics VII, Proc. of SPIE Electronic Imaging, Vol. 5013, pp. 38-45, 2003
19. ShapeCapture: http://www.shapecapture.com [October 2004]