Journal of Archaeological Science 36 (2009) 100109

Contents lists available at ScienceDirect

Journal of Archaeological Science

journal homepage: http://www.elsevier.com/locate/jas

Recovering traditions in the digital era: the use of blimps

for modelling the archaeological cultural heritage
Javier Gomez-Lahoz, Diego Gonzalez-Aguilera*
Cartographic and Land Engineering Department, University of Salamanca, Hornos Caleros, 50, 05003 Avila, Spain

a r t i c l e i n f o

a b s t r a c t

Article history:
Received 10 July 2007
Received in revised form 1 July 2008
Accepted 29 July 2008

Driven by progress in sensor technology, algorithms and data processing capabilities, the recording and
3D virtual modelling of complex archaeological sites is currently receiving much attention. Nevertheless,
the problem remains the huge effort and costs that have to be invested to obtain realistic models. Besides
on-site measurements, much time is often spent in manually rebuilding the whole site with a CAD
package or a 3D-modelling tool.
In this paper, a low-cost and exible system has been devised and realized for virtual archaeological sites
modelling, starting from the acquisition system and ending with the generation of a virtual model in
three dimensions. Different efforts have been made to increase the level of automation without losing
accuracy and reliability. Finally, in order to demonstrate its capabilities some examples applied in
archaeological sites are tested and reported.
2008 Elsevier Ltd. All rights reserved.

Keywords:
Archaeology
Low-cost aerial photogrammetry
Modelling

1. Introduction
From the beginnings of photogrammetry up until now, photogrammetry has always been a technique that has provided accurate
and reliable data in a cost effective way, assigning to the background other important aspects such as 3D visualization of the
results. In fact, the most relevant visualization examples in
a photogrammetry context were dened by a rigid support, twodimensional representation for a rigorous metric analysis; for
example topographic maps with both planimetric features and
contours obtained from aerial photogrammetric cases and rectied
maps of facades from close-range applications.
Fortunately, the emergence of new technologies and disciplines
has provided that the 3D visualization of large and complex sites is
currently receiving much attention. In this way, photogrammetry is
making progresses towards two directions: to popularize the
output data, that is, to make the data available to as many users and
for as many applications as possible; to popularize the technique,
that is, to make the technique readily available in a user-friendly
environment to as many non-photogrammetrist, end-users as
possible. This last goal seems to be the most demanded nowadays,
since several software packages enable the end-user to obtain some
photogrammetric products. However, regarding the popularization
of the output data, 3D visualization together with virtual reality
constitue an alternative that offers promising perspectives for the

* Corresponding author.
E-mail addresses: [email protected], [email protected] (D. Gonzalez-Aguilera).
0305-4403/$ see front matter 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jas.2008.07.013

dissemination of the results. Particularly, in the archaeologistss

eld, new insights can be gained by immersion in ancient worlds,
inaccessible sites can be made available to global public and
different periods or building phases can coexist.
The paper presents the following structure. After this introduction, Section 2 briey reports and discusses some geomatic
techniques applied to archaeological sites modelling. Section 3
explains in detail the full low-cost methodology for the recording,
modelling and virtual visualization of archaeological sites. Some
experimental applications together with a technical discussion are
reported in Section 4. A nal section is devoted to sketch some
future activities and conclusions.
2. Geomatic techniques applied to archaeological
sites modelling
The fast, accurate, cheap modelling and visualization of
archaeological sites is a trivial demand, but far from being fullled.
The justication is twofold. Firstly, because of its complexity, with
the presence of diverse geometric and radiometric features.
Secondly, due to its conceptual interpretation: the archaeological
drawing is rst and foremost a scientic document and as such,
reproduces reality through a graphic interpretation according to
xed rules and criteria.
Traditional methods of string grids do not provide accuracy
standards and a simple survey of the site can only provide a layout
with a few accurate points connected with vectors, without any
further information. Moreover, both methods have the disadvantage of extra people working within the archaeological site for

J. Gomez-Lahoz, D. Gonzalez-Aguilera / Journal of Archaeological Science 36 (2009) 100109

a prolonged period of time, which increases the economical cost,

as well as the possibility of accidental destruction of important
ndings. Ioannidis et al. (2000) sustain that restoration, recording,
reconstruction or even the study of an archaeological site requires
accurate plots through the use of modern surveying techniques;
more recently, Finat et al. (2005) talk about a exible multi-input
and multi-output approach able to support the information arising
from different sensors or techniques and to provide different
levels of information to users with different requirements: from
users to experts; in the same line, El-Hakim et al. (2004) remark
on eight requirements to achieve a full modelling of complex sites:
high geometric accuracy, capture of all details, photorealism, high
automation level, low-cost, portability, application exibility, and
model size efciency. The order of importance of these requirements depends on the applications objective. As a result, the use
of laserscanning and close-range photogrammetry in archaeology
is becoming quite common due to its property of combining
metric characteristics together with a high level of radiometric
detail. However, until now end-users have been discouraged by
cost, time needed to processing data and the fact that the nal
result is unmanageable.
Low-cost photogrammetric systems composed by hardware and
free software devices are the key to success to offer archaeologists
all the powerful tools for fast and accurate recording and mapping
of archaeological sites. Use of low-cost platforms for aerial
photography has been reported in many cases (Miyatsuka, 1996;
Theodoridou et al., 2000; Zischinsky et al., 2000; Karras et al., 1999;
and Faustmann and Palmer, 2005). Kites, balloons, radio controlled
model helicopters, rope-way, sh rods and well buckets are only
some of the ingenious methods photogrammetrists are using for
low altitude photography. In most of these cases, the ideal layout of
the photographs is not attained. This is reported in the case of the
radio controlled model helicopter (Tokmakidis and Skarlatos, 2002)
and in the case of the balloon (Karras et al., 1999), and generally in
any case where the photographer cannot fully control the position
(kites, balloons) or he is not physically behind the camera (ropeway, sh rods, well buckets). Although the radio controlled helicopter with a radio link for transmission of the imaged object on
the ground does not seem to suffer from the aforementioned
problems, this is not the case. Therefore, the scale is not equal
between photographs and overlaps are far from the ideal. Some of
these problems have been partially solved recently; Skarlatos et al.
(2004) report several good results obtained in archaeological sites
using a radio controlled model helicopter, as a semi-metric camera
platform; Eisenbeiss (2005) presents an autonomous helicopter
system that adds a GPS/INS system providing good results.
However, both approaches continue to require highly skilled
operators to control the helicopter.
The approach that is proposed in this paper overlaps with the
work done by these researchers but also differs in several ways.
Firstly, in relation to hardware or low-cost acquisition systems, an
original and robust camera platform that guarantees stability and
quality in taking photographs has been devised and realized
allowing the user to acquire rigorous vertical (stereoscopic) and
oblique images easily. Secondly, free software and tools have been
developed to provide virtual archaeological sites modelling.

3. Methodology
The low-cost process developed (Fig. 1) involves well-known
steps: design (sensor and network geometry); data processing
(feature extraction, image matching, image orientation, geometric
constraints and 3D points); data modelling (constrained mesh
generation, vertical walls modelling) and data visualization (virtual
visualization and mapping textures).

101

Fig. 1. Low-cost methodology for archaeological sites modelling.

3.1. Design: sensor and network geometry

After some testing concerning kites and powered paragliders,
we found that blimps provided the more suitable combination of
stability and simplicity. Paramotors, besides some other drawbacks,
such as training, noise and price, are not stable enough (the same
holds for kites) to guarantee good stereoscopic images (though they
are adequate for the less restrictive geometry of oblique images).
We found also that winds above 67 km/h (specially if variable)
were a very serious problem for the blimp to achieve the normal
case geometry and so, we had to devise a platform system that
could absorb, to a certain degree, instabilities of the wind and to
guarantee that the camera attitude would remain the same
between two consecutive takings. This was achieved by the
combination of gyroscopes and the Picavet system described below.
So, the blimp was acquired with no further conditions but the
platform was designed and developed by us to house and shut the
camera, to allow for two degrees of freedom rotation and to
partially compensate unpredictable movements.
The designed sensor system is constituted by two main units:
Flight Unit (FU) and Ground Control Unit (GCU).
 FU. The ight unit is composed of four main parts: a helium
blimp with a capacity of 11 m3 that provides a lifting force of
more than 5 kg; a Picavet system (implemented in 1912 by
Pierre L. Picavet) (Fig. 2b) which consists of a rigid frame suspended, by means of a continuous string and pulleys from
another string attached to the blimp body. The result is a relatively stable self-levelling platform which maintains a xed
inclination angle and provides effective stability of the wind
movements; a reex digital camera Nikon D70 with a resolution of 6.1 mega pixels, a focal length of 14 mm and CCD
dimensions of 23.7  15.6 mm; a camera platform (Fig. 2a)
equipped with servomechanisms (endowed with gyroscopic
devices to increase stability), video and radio control which
allow us to obtain the video signal of the camera view over
a monitor in real time, as well as to control the two main
camera rotations. The most signicant contribution in this
sense is an electronic gyroscope which allows us to preserve
camera rotation independently of blimp movements.
 GCU. The ground control unit is composed of three main parts:
a monitor, to obtain a eld of view pre-visualization in real

102

J. Gomez-Lahoz, D. Gonzalez-Aguilera / Journal of Archaeological Science 36 (2009) 100109

Fig. 2. Sensors design: camera platform (a) and Picavet system (b).

time; remote controllers for controlling the shutting and

camera rotations (azimuth and tilt) and control ropes to
provide planimetric and altimetric control over blimp position.
On the other hand, with relation to the design of network
geometry, several assumptions have to be taken into account:
 The accuracy of a network increases with the increase of the
base-to-depth (B/D) ratio and by using convergent images
rather than images with parallel optical axes.
 The accuracy improves signicantly with the number of images
where a point appears. But measuring the point in more than
four images gives less signicant improvement;
 The accuracy increases with the number of measured points
per image. However, the increase is not signicant if the
geometric conguration is strong and the measured points are
well-dened (like targets) and well distributed in the image;
 The image resolution (number of pixel) inuences the accuracy
of the computed object coordinates: on natural features, the
accuracy improves signicantly with the image resolution,
while the improvement is less signicant on well-dened large
resolved targets.

combinations between scales, ight height, base and their overlap

correspondence on the ground. sXY and sZ represent the planimetric and altimetric a priori deviation, respectively.
The combination underlined is usually selected as the optimal
one in that it satises our accuracy and efciency requirements,
allowing cartographic scales between 1/500 and 1/1000.
After that, it only remains to compute the number of photographs and strips required to cover the area of interest (Fig. 3).
Moreover, a determined number of oblique images could be taken
to complement vertical stereoscopic images, providing relevant
geometric information, as well as contextual and panoramic
information.
Finally and before photography taking, several control points are
pre-signalized on the ground with special targets and surveyed
with GPS or total station, providing an external datum (reference
system) of the network. At least a minimum of three points

Considering the above assumptions and the sensor geometry

described above, it is advisable to plan a ight project before taking
aerial images. The optimal ight height as well as a photo-base
should be computed before taking images in order to provide the
overlap and scale required for the archaeological site modelling.
Considering the camera parameters mentioned above together
with an overlap of 63%, Table 1 illustrates the different

Table 1
Flight project planning
Height (m)

Photograph scale

sXY (mm)

sZ (mm)

Overlap (m)

Base (m)

14
28
42
56
70
82
96
112
126
140

1/1000
1/2000
1/3000
1/4000
1/5000
1/6000
1/7000
1/8000
1/9000
1/10000

8
16
24
32
40
48
56
64
72
80

12.8
25.6
38.4
51.2
64
76.8
89.6
102.4
115.2
128

15  15
30  30
45  45
60  60
75  75
90  90
105  105
120  120
135  135
150  150

8.7
17.4
26.1
34.8
43.5
52.2
60.9
69.6
78.3
87

Fig. 3. Design network geometry.

J. Gomez-Lahoz, D. Gonzalez-Aguilera / Journal of Archaeological Science 36 (2009) 100109

between image pairs are surveyed. This step is especially important

in archaeological sites modelling where the true prevalence of Zaxis direction provides a real altimetric analysis.
3.2. Data processing
Depending on the types of images captured from air (vertical or
oblique), two different processing approaches are performed.
3.2.1. Vertical images processing
A sequential process composed of several well-known steps and
computed through bundle adjustment is developed. LISA free
software (Jacobsen, 2001) was used to process vertical images.
Firstly, a feature extraction of interest points1 based on a Haralick
(1985) operator is performed. Haralick rst extracts windows of
interest from the image and then computes the precise position
of the point of interest inside the selected windows. The windows
of interest are computed with a gradient operator and the normal
matrix. The point of interest is determined as the weighted centre
of gravity of all points inside the window. Secondly, a pair-to-pair
image matching process is performed using the extracted interest
points. The process returns the best match in the second image for
each interest point in the rst image. At rst cross-correlation is
performed between image pairs and then the results are rened
using least square matching, LSM (Grun, 1985). The point with the
biggest correlation coefcient is used as an approximation for the
matching process. The cross-correlation process uses a small
window around each point in the rst image and tries to correlate it
against all points that are inside a search area in the adjacent image.
The search area is given considering the direction of ight sequence
and the image parallax (disparity). The nal number of possible
matches depends on the threshold parameters of the LSM and on
the disparity between image pairs; usually it is around 40% of the
extracted points. The disparity threshold between the two images
is one of the most complicated parameters to select. Incorrect
disparity leads to very few correspondences (parallax parameter
smaller than the correct one) or very long computation (bigger
parameter). Sometimes when the matching process is unstable,
a ltering of false correspondences based on the disparity gradient
concept (Klette et al., 1998) is used.
Finally, once images matching is performed, the image
parameters together with the surface 3D coordinates can be
computed automatically through bundle adjustment (Triggs et al.,
2000). As adjustments observations point-feature is used. The
mathematical basis of the bundle adjustment is the collinearity
model, that is, a point in object space, its corresponding point in the
image plane and the projective centre of the camera lie on a straight
line. Photogrammetric bundle adjustment is extended with functional constraints, in particular through the external datum of the
network or with restrictions derived from geometric or physical
characteristics of the network.

103

based modelling applied to a wide spectrum of objects and sites.

The single image-based modelling approach starts with a feature
extraction of straight lines based on Canny (1986) and Burns et al.
(1986) operators. Moreover, a clustering of straight lines according
to three main orthogonal directions of the scene or three vanishing
points is applied. This step is performed through a robust estimator,
RANSAC (Fischler and Bolles, 1981) that incorporates an original
slope analysis as voting criterion. This step also allows us to isolate
some relevant breaklines.
After that, an estimation of camera parameters (internal and
external) based on vanishing points and some a priori information
about the scene, that is a distance, is carried out. Therefore,
a complete camera model can be recovered following two steps, in
which internal and external parameters are estimated separately. In
the rst step, the intrinsic parameters, that is, the focal length, the
principal point and the radial lens distortion, are recovered automatically based on vanishing points geometry and image analysis.
The orthocentre of the triangle formed from the three vanishing
points of the three mutually orthogonal directions identies the
principal point of the camera through the cross product of the
segments of the triangle and its heights. The focal length can be
computed afterwards as the square root of the product of the
distances from the principal point to the triangles vertices. Finally,
the radial lens distortion parameters (k1, k2) are estimated by
exploiting the presence of short segments (mini-segments)
through the collinearity condition. This estimation is performed
only with line segments that satisfy the orientation constraint. In
the second step, the extrinsic parameters, that is the perspective
rotation matrix and the translation vector which describe the rigid
motion of the coordinate system xed in the camera are estimated
in a two-pass process. Firstly, the rotation matrix (camera orientation) is obtained automatically based on the correspondence
between the vanishing points and the three main object directions.
This relationship allows the cosine vectors of the optical axis to be
extracted, obtaining the three angles (axis, tilt, swing) directly.
Then, the translation vector, that is, the absolute camera pose is
estimated based on some a priori object information, for example
a distance together with a geometric constraint dened by the user.
Therefore, the reference frame for the camera pose estimation is
dened with relation to the object geometry based on a local
coordinate system. The robustness of the method depends on the
precision and reliability of vanishing points computation, so the
incorporation of robust estimators in the previous step is crucial.
Finally, in order to recover 3D information from single oblique
images and overcome the so called ill-posed2 problem, the
collinearity condition supported by geometric constraints on the
object (perpendicularity, coplanarity, parallelism, etc.) and image
invariants (distances and angles) can be used to solve the 3D
reconstruction problem from a single image (Heuvel, 1998).
3.3. Data modelling

3.2.2. Oblique images processing

Oblique images acquired from the air are processed independently exploiting a single image-based modelling approach (Aguilera and Lahoz, 2006). Therefore, this step constitutes the perfect
complement for vertical images in archaeological sites modelling,
since it provides relevant geometric information due to occlusions
or lack of features to match between images. sv3DVision (Aguilera
and Lahoz, 2006) was the software used for oblique images processing. This free software allows us to perform a single image-

The reconstruction of surfaces from point cloud is rather difcult in the case of archaeological sites since the data are usually
uneven and sparse. No regular surfaces such as planes, cylinders or
cones can be handled to render the site surface. In addition, the
presence of vertical walls demands a two steps approach: in the
rst step, the terrain surface is modelled through conventional
triangulation procedures; in the second one, the vertical walls are
modelled from simple primitives and then blended with the
triangle mesh.

1
Interest points are locations in the images where the signal content changes,
usually in two-dimensions; they are geometrically stable under different transformations and present high information content.

2
For each point, we have two equations (collinearity condition) and three
unknown coordinates (X,Y,Z). Therefore the system is underdetermined and some
more assumptions need to be introduced.

104

J. Gomez-Lahoz, D. Gonzalez-Aguilera / Journal of Archaeological Science 36 (2009) 100109

3.3.1. Surface modelling

The triangulated mesh is a powerful tool to render complex
surfaces. A collection of points with no regular distribution is
converted into the vertices of a set of irregular triangles. Best results
are achieved when the sample is dense in points and some
geometric information (breaklines) can be observed. Nevertheless
this is not always the case as the measured points may be unorganized and often noisy. Triangulation can be performed in several
ways according to the geometry of the input data, so there is a wide
spectrum of algorithms that can be applied, from the easy 2D
approaches based on Voronoi diagrams to the sophisticated 3D
approaches based on tetrahedralization methods. 3D triangulations
approaches not always provide good results, but require highly
demanding computation efforts. Thurston (1997) asserts that 2.5D
approaches require less storage space for mesh generation, being
more efcient in some cases.
In the present case, the method performed to deal with the
complexity of archaeological sites is based on a 2.5D approach of
Delaunay Triangulation (Delaunay, 1934) supported by an incremental method (Bourke, 1989) and improved with robust
assumptions such as geometric and topological constraints, that is
breaklines, length of triangles and adjacency of triangles. The so
called incremental method strategy starts, as rst approximation,
with a unique triangle which encloses the complete point cloud.
This triangle is split in recursive sets of triangles always supported
by existing points, until every point is a vertex of the set of triangles.
This process can be signicantly improved if constraints, such as
breaklines, can be added to the initial data. These features are
enforced to contain triangle edges, so none of the breaklines can be
crossed by an edge triangle. On the other hand, a topological
analysis among adjacent triangles is established based on the
number of the vertices that constitute each triangle and their
adjacent triangles. This faces topology improves whatever search
operation onto the mesh, as well as the re-triangulation if it is
required.
3.3.2. Vertical walls modelling
After this, vertical walls are extracted semi-automatically
through image processing. The basic geometric shapes that represent vertical walls are rst recovered using straight lines extracted
from images, taking advantage of constraints available from
geometric relationships such as parallelism, coplanarity and
orthogonality.
Two cases can be considered depending on the position of
vertical walls (over or under ground, that is buried):
 Vertical walls over the ground are generated automatically
projecting the extracted elements over the triangulated surface
orthogonally. An interpolation over the triangulated surface is
required to solve some indetermination in the projection of
polygonal elements, as well as in the intermediate positions
(Fig. 4).
 Vertical walls under the ground (buried) are modelled using
a re-triangulation process (Fig. 5). Firstly, the edges and their
corresponding points (big red point) that form the upper part
of a wall are identied. Then, the set of points, connected to

Fig. 4. Vertical walls modelling over ground.

Fig. 5. Vertical walls modelling under ground.

these points, and placed at a lower position are determined

(red point at the lower right side of the big red point). The
original triangles constituted by these two sets of points (dash
red line) are false triangles because they do not represent the
vertical walls and so are deleted. Secondly, a new point (empty
blue point) is created for every point of the rst set and placed
vertically under it. The Z of these points is obtained from the Z
of the nearest points of the second set. Finally, two types of new
triangles are created: on one hand, the new points are connected to the closest points of the second set to form quasi
horizontal triangles corresponding to the terrain (dash blue
horizontal line) and, on the other hand, the new points are also
connected to the rst set of points to form the vertical triangles
that render the wall.

3.4. Data visualization

With relation to data visualization, a texture mapping technique, based on grey-scale or colour images, has been implemented. For every surface element (surfel) on the object the
corresponding pixel element (pixel) on the image can be computed
easily through the collinearity condition if the inner and outer
camera parameters are known, as in the present case. Then, the
grey-scale or colour RGB pixel values can be transferred (or interpolated) to the matching surfel (and closest neighbours). Particularly, this procedure is developed by means of the so called Anchor
Points method (Kraus, 1993). Only the corresponding pixel of
every vertex of the triangles is determined and the grey-scale value
or RGB value is transferred from the rst to the second. Then
a linear (afne) transformation is used to interpolate the grey (RGB)
value of every surfel within every triangle from the values of the
vertices.
In this case, the visualization step is supported by the VRML
(Virtual Reality Modelling Language) standard and is performed
with the aid of the software sv3DVision. Surface geometry and
simple topology are transformed into a world of objects that
exhibit the following features:
 Geometry, expressed by simple primitives (mainly boxes for
the vertical walls), or by sophisticated ones (so called Indexed
Line/Face Sets for the terrain).
 Appearance, expressed by a so called material implementation
that species the surface response to different light sources
and also by a texture or pattern representation of the surface.
 Referencing in a local or global datum dened by its origin, axis
direction and scale factor.
 Metadata information, containing specic information about
the object: number of faces, number of lines, surface, length,
accuracy, etc.
 Sensing activity: sensors can be attached to objects, that detect
user movements through the scene or when the user interacts
with some input device.

J. Gomez-Lahoz, D. Gonzalez-Aguilera / Journal of Archaeological Science 36 (2009) 100109

105

Fig. 6. Design step: network geometry.

 Movement: the previous sensing activity can trigger different

responses such as motion or reshaping.
 Decision taking (intelligence): scripts can be implemented in
java or javascript code to analyze diverse circumstances and,
consequently, adopt the better response.
 Prototyping: objects can be grouped, encapsulated and reused
and so, integrated into recursive and complex frames.
4. Experimental applications
The approach described above has been tested over two
emblematic archaeological settlements located in Castile (Spain):
the Roman city of Clunia (Burgos) and the Celtic settlement of Las
Cogotas (Avila).
Roman city of Clunia. Colonia Clunia Sulpicia, capital of Conventus Cluniensis in the Province of Tarraconensis in Hispania, was
the largest Roman city in the Iberian Peninsula. Situated on
a limestone plain at an altitude of 1000 m it reached an extension of
130 ha. It was founded ex-novo in the Augustean/Tiberian epoch as
a sinecism of two pre-existent celtic-iberian settlements: the celticiberian nucleus of Clunioq (which gave its name to the Roman city)
and the settlement known as Arauzo de Torre. The city underwent
a remarkable economic development during the 1st and 2nd
centuries A.D. and various archaeological expeditions have brought
to light the forum, the basilica, Roman baths, an abattoir and other
blocks occupied by private houses as well as the theatre which
exploits natural features of the terrain and has been a popular
subject for prints since the 19th century.
The target was to model the part of the city known as House I.
From a geometric point of view this site may be characterized as
follows:
 The area to be covered with stereoscopic overlap is a square of
about 100 m  100 m.
 The terrain on which the house rises is almost at. The larger
relieves are due to the excavated zones.
 The house walls display a regular frame. The heights of these
walls range from only a few centimetres up to 3 m. The regularity helps to identify the singular points and lines that
support the house outline but the vertical walls introduce
plenty of occlusions.
Under these conditions the following decisions were assumed. A
ground sample distance (GSD) of about 4045 mm (leading to

a mapping scale of about 1/5000) would be adequate for the model.

The stereoscopic height precision related to it, 70 mm, is also
adequate. Due both to the object regularity and the occluded zones,
a dense oblique image network would improve greatly the results,
both in precision and reliability and for both dimensions, planimetry and altimetry. Therefore, the following steps were undertaken.
The ight plan (Fig. 6) consisted of three strips with three
vertical photographs each at a ying height of 75 m. Two strips
with two photographs would have been enough to achieve simple
stereoscopic conditions but, as stated above, the occlusions related
to vertical walls made up our mind to increase photographic
coverage. In addition, oblique images from a variety of positions
were also taken. The datum was dened by the GPS observation of
12 targets (red rectangles). In addition nine bigger targets (in red
also) were placed on the terrain to ease the identication of the
projected principal points of each vertical image to be taken.
Vertical image processing was performed automatically through
image correspondences with cross-correlation and LSM (Fig. 7).
Table 2 shows the statistics obtained in vertical image processing.
Vanishing point detection and computation, assuming orthogonality conditions between the walls, main directions, were applied
to estimate the inner and outer camera parameters. The addition of
some geometric constraints led to a partial 3D reconstruction of the
scene, for example vertical walls. After image processing, a semiautomated point cloud of 4700 points with several breaklines and
vertical walls was obtained. Firstly, for the conversion of the point
cloud to a triangular surface mesh, a 2.5D constrained Delaunay
triangulation was applied (Fig. 8) considering only surface breaklines. Secondly, an automated vertical walls modelling using simple
elements like polygonal boxes was applied projecting the extracted
elements over the triangulated surface orthogonally (Fig. 8).
Occluded areas were complemented and reconstructed semiautomatically through oblique image-based modelling.
Finally, a full textured model is generated (Fig. 9). Six vertical
images were used to map textures over the surface while four
Table 2
Number of extracted points and achieved theoretical precisions s
Images per strip
Tie points
Points in two images
Points in three images
sX (m)
sY (m)
sZ (m)

3
531
330
201
0.03
0.02
0.04

106

J. Gomez-Lahoz, D. Gonzalez-Aguilera / Journal of Archaeological Science 36 (2009) 100109

Fig. 7. Stereoscopic pair processing through matching technique.

oblique images applied texture to vertical walls. Nevertheless, in

both cases the texture mapping was applied through the Anchor
Points method considering previously a visibility surface
determination.
Celtic settlement of Las Cogotas. Las Cogotas is another large
site (15 ha), defended by two walled enclosures. The last excavations carried out in the southwest of the second enclosure (RuizZapatero and Alvarez-Sanchs, 1995) have revealed an area with
abundant archaeological material and various specialized areas;
a large communal midden, a stone pavement of complex interpretation connected with the fortication wall, and a potters

workshop. The wheel-made pots with painted decoration found at

the site are dated to the second century BC. The observed stratigraphy is also important: the existence of a midden under the wall
demonstrates that before the potters workshop and the fortications were built, artisanal activities were already being carried out
in this area.
In this case, the main target, the modelling of the main and
central part of the Celtic settlement was enriched with the goal of
the idealized reconstruction of some of the original houses on the
virtual model. From a geometric point of view this target may be
characterized in the following way:

Fig. 8. Surface modelling through Delaunay triangulation and vertical walls modeling.

J. Gomez-Lahoz, D. Gonzalez-Aguilera / Journal of Archaeological Science 36 (2009) 100109

107

Fig. 9. Textured and virtual 3D model.

 The area to be covered with stereoscopic overlap can be

enclosed in a rectangle of about 120 m  220 m.
 The settlement lies on a rocky hill and therefore, the terrain is
rather complicated. The slopes are larger at the outer part of
the walls. It must be highlighted that the reconstruction of the
original houses location is highly based on breaklines that can
be clearly appreciated on the place and thus, that must be
precisely modelized.
 The walls have been partially rebuilt and exhibit a curved
display.
Under these conditions a GSD of about 2530 cm (corresponding to a mapping scale of about 1/3500) with a related height
precision of 45 mm was decided. Therefore, the following steps
were undertaken:
The ight plan (Fig. 10) consisted of three strips with six vertical
photographs each, at a ying height of 50 m. In addition, oblique
images from a variety of positions were also taken to obtain relevant information of the vertical sides of the wall. The datum was
dened by the GPS observation of 15 targets (red rectangles). In
addition, 18 bigger targets (in red also) were placed on the terrain to
ease the identication of the projected principal points of each
vertical image to be taken.
Vertical image processing was performed automatically through
image correspondences with cross-correlation and LSM (Fig. 11).
Table 3 shows the statistics obtained in vertical image processing.

Fig. 10. Design step: network geometry.

Oblique image processing through single image-based modelling performs as a complement, providing some relevant geometric
information in the two walled enclosures such as geometric
constraints (green planes) (Fig. 12).
After image processing, a semi-automated point cloud of 7500
points with several breaklines and vertical walls was obtained.
Firstly, for the conversion of the point cloud to a triangular surface
mesh, a 2.5D constrained Delaunay triangulation was applied
considering only surface breaklines. Secondly, an automated
vertical walls modelling using simple elements like polygonal
boxes was applied projecting the extracted elements over the
triangulated surface orthogonally. Occluded areas were complemented and reconstructed semi-automatically through oblique
image-based modelling.
Vertical images were used to render the main surface while
oblique images were used to render the vertical walls. In addition
several Celtic houses were reconstructed and placed on the terrain
following the historical data available (Fig. 13).
4.1. Technical discussion
In this section the nal results are discussed, as well as the
different considerations of the proposed process.
Regarding the data capture we can remark that the wind in the
case of Clunia, about 810 km/h, made it difcult to navigate the
blimp and led to additional time (about 50%) devoted to this
purpose. In the case of Las Cogotas the difculties arose from the
natural conditions of the site: steep slopes, huge granite rocks, trees
and tall weeds. This hindered considerably the task, not only for the
difculty of moving around the place carrying the blimp, controlling the camera set and measuring the targets, but also because of
the difculty of placing with enough accuracy (less than 3 m) the
projected principal points of the vertical images. Very probably this
task can be improved, whenever these natural conditions are
found, if the coordinates of these points are computed previously
and then, the GPS is used to nd the targeted location. Note that the
regular disposition of the walls in Clunia facilitated greatly this
work.
In relation to data processing, the most remarkable aspects to be
discussed are relative to the automatic reconstruction of the walls.
In the case of Clunia, these difculties concern the points (corners)
where orthogonal walls met. Additional manual work was needed
to identify these points accurately before the upper part of the walls
was projected on the ground. Further developments should aim at
the automation of this item. In the case of Las Cogotas the most

108

J. Gomez-Lahoz, D. Gonzalez-Aguilera / Journal of Archaeological Science 36 (2009) 100109

Fig. 11. Stereoscopic pair processing through matching technique.

crucial aspects are related to the curved nature of the walls. In this
case, also, some editing work was needed to cluster adequately
isolated segments before the wall generation step.
Finally, concerning the data visualization, it should highlight the
qualitative change that represents the inclusion of houses in the
virtual model. This issue not only requires special care for the
material implementation of roofs, walls, doors, etc. It implies a new
visualization concept which incorporates time as an extra dimension towards the 4D worlds.

5. Conclusions and future perspectives

 The development of the aerial low-cost platform constitutes

a non-destructive technique for archaeological sites modelling
that enables the access to the geometry (shape, size and
dimensions) and radiometry (texture and material) of the
object even if it is inaccessible.
 The owline implemented represents an original alternative to
get low-cost aerial orthophotos with large mapping scales.
 The nal results have been well accepted by professional
archaeologists, since the usually obtained metric measurements are based on classical approaches which are whether
limited in accuracy or time consuming. Particularly, the accuracy attained with this approach is around 0.03 m for (X,Y)
coordinates and 0.05 m for Z coordinates.

The presented paper has investigated and tested the viability

and possibilities opened by the use of blimps for archaeological
sites modelling. A consistent and reliable full process pipeline has
been developed and presented. It was demonstrated with different
practical examples tested through our own tools and software.
In relation to the most relevant aspects of the proposed
approach, we could remark the following:

Table 3
Number of extracted points and achieved theoretical precisions s
Images per strip
Tie points
Points in two images
Points in three images
sX (m)
sY (m)
sZ (m)

5
337
225
112
0.01
0.03
0.05

Fig. 12. Geometric constraints (planes) applied to walled enclosure modeling.

J. Gomez-Lahoz, D. Gonzalez-Aguilera / Journal of Archaeological Science 36 (2009) 100109

109

Fig. 13. Textured and virtual 3D model with the ideal reconstruction of Celtic houses.

 The virtual 3D model has enabled archaeologists to extract

whatever metric measurement interactively, as well as to get
remote views from whatever viewpoint.
With relation to the eld work and its logistic, the initial operation of inating the balloon and putting it in place consumes about
4560 min and after this, the operations of changing the batteries
or downloading the images from the memory card are somewhat
fast (no more than ten minutes). So the larger the ground area, the
more cost efcient is the methodology, reaching a productivity of
3 ha/h approximately. The system is simple enough to be handled
by two persons: one to drive the blimp and another one to control
the image taking.
Finally, regarding future perspectives, the 3D virtual model
provides for free and exible travel in space, but does not provide
a travelling in time. Therefore, the development of a process able to
analyze and superpose the excavations for every stratigraphic layer
would allow us to dispose of its evolution over time. Another midterm challenge could be the adaptation of this low-cost system to
remote sensing applications through new lters (ultraviolet and
infrared) incorporated into the camera platform, so multispectral
images could provide relevant information about the location and
classication of occluded areas.
Acknowledgements
The authors would like to thank Amaya Rubio and Henar Zapico
who, as students at Salamanca University, supported and put in
practice some of the results and tools obtained at the Las Cogotas
and Clunia settlements.
References
Aguilera, D.G., Lahoz, J.G., 2006. sv3DVision: didactical photogrammetric software
for single image-based modeling. International Archives of Photogrammetry,
Remote Sensing and Spatial Information Sciences 36 (6), 8 pp. on CD-ROM.
Bourke, P., 1989. An algorithm for interpolating irregularly-spaced data with
applications in terrain modelling. Pan Pacic Computer Conference, Beijing,
China, 6 pages.
Burns, J.B., Hanson, A.R., Riseman, E.M., 1986. Extracting straight lines. IEEE Transactions on Pattern Analysis and Machine Intelligence 8 (4), 425455.
Canny, J.F., 1986. A computational approach to edge detection. IEEE Transactions on
Pattern Analysis and Machine Intelligence 8 (6), 679698.
Delaunay, B., 1934. Sur la sphere vide. A la memoire de Georges Voronoi. Izvestia
Akademii Nauk SSSR, Otdelenie Matematicheskikh i Estestvennykh Nauk.
(Bulletin of Academy of Sciences of the USSR) 7, 793800.
Eisenbeiss, H., 2005. Applications of photogrammetric processing using an autonomous model helicopter. Proceedings of the XX CIPA Symposium. Torino, Italy, 6 pp.

El-Hakim, S.F., Beraldin, J.-A., Picard, M., Godin, G., 2004. IEEE Computer Graphics
and Applications 24 (3), 2129.
Faustmann, A., Palmer, R., 2005. Wings over Armenia: use of a paramotor for
archaeological aerial survey. Antiquity 79 (304), 402410.
Finat, J., Santamara, M.A., Martnez, J., Fernandez, J.J., San Jose, J.I., 2005. The Roman
Theatre of Clvnia: hybrid strategies for applying virtual reality on laser scanning
3d les. Proceedings of the ISPRS Working Group V/4 Workshop. 3D-ARCH
2005, Mestre-Venice, Italy. 8 pp.
Fischler, M.A., Bolles, R.C., 1981. Random sample consensus: a paradigm for model
tting with applications to image analysis and automated cartography.
Communications of the ACM 24 (6), 381395.
Grun, A., 1985. Adaptive least squares correlation: a powerful image matching
technique. South African Journal of Photogrammetry, Remote Sensing and
Cartography 14 (3), 175187.
Haralick, R.M., 1985. Second directional derivative zero-crossing detector using the
cubic facet model. IEEE Computer Vision and Pattern Recognition, 672677.
Heuvel van den, F.A., 1998. 3D reconstruction from a single image using geometric
constraints. ISPRS Journal of Photogrammetry and Remote Sensing 53 (6), 354
368.
Ioannidis, C., Potsiou, C., Soile, S., Badekas, J., 2000. Detailed 3D representations of
archaeological sites. International Archives of Photogrammetry and Remote
Sensing 33 (B5), 642649.
Jacobsen, K., 2001. PC-based digital photogrammetry, Workshop on Data Analysis
and Image Processing Techniques, Damascus, GORS, 10 pp.
Karras, G., E., Mavrommati, D., Madani, M., Mavrelis, G., Lymperopoulos, E., Kambourakis, A., Gesadis, S., 1999. Digital orthophotography in archaeology with
low-altitude non-metric images. Proceedings of the ISPRS Workshop in
Photogrammetric measurement, object modeling and documentation in
architecture and industry, Thessaloniki, Greece, 22(5), 811.
Klette, R., Koschan, A., Schluns, K., 1998. Computer Vision: Three-dimensional Data
From Images. Springer-Verlag Singapore Pte. Limited.
Kraus, K., 1993. Advanced Methods and Applications. Vol. 2. Fundamentals and
Standard Processes. Vol. 1. Institute for Photogrammetry Vienna University of
Technology. Ferd. Dummler Verlag, Bonn.
Miyatsuka, Y., 1996. Archaeological real time photogrammetric system using digital
still camera. International Archives of Photogrammetry and Remote Sensing 31
(B5), 374378.
Ruiz-Zapatero, G., Alvarez-Sanchs, J.R., 1995. Las Cogotas: Oppida and the roots of
urbanism in the Spanish Meseta. In: Cunliffe, B., Keay, S.J. (Eds.), Social
Complexity and the Development of Towns in Iberia: From the Copper Age to
the Second Century AD. Oxford University Press, Oxford, pp. 209236.
Skarlatos, D., Theodoridou, S. and Glabenas, D., 2004. Archaeological surveys in
Greece using radio-controlled helicopter. Proceedings of FIG Working Week,
Athens, Greece, 15 pp.
Theodoridou, S., Tokmakidis, K., Skarlatos, D., 2000. Use of radio-controlled model
helicopters in archaeology surveying and in building construction industry.
International Archives of Photogrammetry and Remote Sensing 33, 6 pp.
Thurston, W.P., 1997. Three-Dimensional Geometry and Topology, vol. 1. Princeton
University Press, New Jersey.
Tokmakidis, K., Skarlatos, D., 2002. Mapping excavations and archaeological sites
using close range photos. Proceedings of the ISPRS Commission V, Corfu,
Greece, pp. 459462.
Triggs, B., McLauchlan, P., Hartley, R., Fitzgibbon, A., 2000. Bundle adjustment
a modern synthesis. LNCS. In: Vision Algorithms: Theory and Practice. Springer
Verlag, pp. 298375.
Zischinsky, T., Dorffner, L., Rottensteiner, F., 2000. Application of a new model
helicopter system in architectural photogrammetry. International Archives of
Photogrammetry and Remote Sensing 33 (B5), 7.