Data Process

BUILDING EXTRACTION FROM MULTIPLE DATA SOURCES: A DATA FUSION

FRAMEWORK FOR RECONSTRUCTION OF GENERIC MODELS

K. Khoshelham

Dept. of Land Surveying and Geo-Informatics, The Hong Kong Polytechnic University, Hong Kong

*******.*@*****.***.**

Commission III, WG III/4

KEY WORDS: Building Extraction, Fusion, Automation, Modelling, Aerial image, Laser scanning, GIS

ABSTRACT:

Automated building extraction from multi-source data has attracted great attention in recent years. This paper presents an approach to

automated building extraction by fusion of image data, height data and 2D ground plans. In this approach buildings are detected

using ground plans and height data. A split-and-merge process is applied to fuse image and height data and derive the parametric

forms of roof planes. Vegetation regions are identified and discarded using the image information in red and infrared channels. Walls

are reconstructed as vertical planes upon the ground plan. The model planar faces are finally intersected and resulting plane patches

are assembled together to form a generic polyhedral model. Results of the experimental testing indicate the promising performance of

the proposed approach in automatic detection and reconstruction of buildings.

1. INTRODUCTION from noise, low contrast, shadow and occlusion; hence, features

extracted from images are incomplete and uncertain. Height

Automated building extraction has been a challenging problem data is of relatively low resolution, which makes the extraction

in the past two decades. Automated approaches that work solely of building boundaries difficult. These complexities have led

based on a single source of data, suffer from the lack of the research efforts toward methods that combine data from

robustness due to complexities in data as well as in buildings. multiple sources. In recent years a number of methods have

Therefore, in recent years research efforts have been focused on been developed for fusion of image and height data (Ameri,

automated approaches that make use of data from multiple 2000; Cord et al., 2001; Jaynes et al., 2003; Rottensteiner and

sources. This paper presents a framework for fusion of available Jansa, 2002), although height data in some of these methods has

data sources that can be used in an automated system for been generated from image data using matching techniques.

extraction of building objects. Fusion of image data and 2D ground plans has also appeared in

a number of works (Haala and Anders, 1996; Jibrini et al., 2000;

Various types of data from different sources have been used for Pasko and Gruber, 1996; Suveg and Vosselman, 2004). In

automated extraction of buildings. Aerial images are the most another fusion strategy, Haala and Brenner (1998) used DSM

widely used data. Single aerial images have been used for and 2D ground plans in an approach to automated building

automated detection and reconstruction of buildings with simple extraction.

models (Huertas et al., 1993; Lin et al., 1995; Lin and Nevatia,

1996; Nevatia et al., 1997; Shufelt and McKeown, 1993). For Despite the great deal of research that has been carried out on

reconstruction of more complex buildings, stereo and multiple automated building extraction, still the role of multi-source data

overlap aerial images have attracted greater attention (Baillard and fusion strategies has not been completely explored. The

et al., 1999; Bignone et al., 1996; Dang et al., 1994; Fischer et objective of this paper is to develop a framework for fusion of

al., 1998; Fua and Hanson, 1991; Henricsson, 1998; Henricsson available data sources that can be used for automated extraction

and Baltsavias, 1997; Herman and Kanade, 1986; Jaynes et al., of buildings. The proposed fusion framework combines aerial

2003; Kolbe, 1999; Moons et al., 1998). Image data from other images in colour and infrared channels, DSM and DTM from

sources have not been suitable for building reconstruction. aerial laser scanner and 2D ground plans from a GIS database in

Remotely sensed images from satellites are of relatively low an approach to automated extraction of buildings. In this

ground resolution and, therefore, can only be used for detection approach ground plans are used to detect buildings in the scene

of buildings. Close range images are, on the other hand, too and reconstruct the walls. Image data in red and infrared

detailed and can be used to map textures onto final channels are used to identify and remove vegetation regions.

reconstructed models. Roof planes are reconstructed by exploiting information from

image, DSM and DTM. Generic polyhedral models are finally

Height data is another widely used type of data. Digital surface formed by assembling reconstructed walls and roof planes.

models (DSMs) from aerial laser scanning systems have been

used in a number of approaches (Brunn and Weidner, 1997; The paper is structured in 6 sections. In section 2 an overview

Maas, 1999; Vosselman, 1999; Weidner and Forstner, 1995). of the proposed fusion strategy is presented. Reconstruction of

Range data from terrestrial laser scanners, however, has not walls and roof planes are described in section 3. Section 4

been proved useful for automated building extraction. discusses the reconstruction of generic models using a plane

patch reconstruction technique. Experiments and results are

Automated building extraction from both image and height data shown in section 5. Conclusions are made in section 6.

encounters a number of complexities. Image data often suffers

2. A FRAMEWORK FOR FUSION OF IMAGE DATA, 3.1 Localization of buildings using ground plans and height

HEIGHT DATA AND 2D GROUND PLANS data

The first step in a building extraction system is the detection of A 2D ground plan is usually stored as a polygon with an array

buildings in the scene. In this work, buildings are detected by of corner points with X and Y coordinates in the world

projecting ground plans to image data. For this purpose, the coordinate system. The footprint of each building is localized in

third dimension (height) of the ground plans is interpolated in the image by interpolating the height of every corner point of

the height data. the ground plan in DTM and projecting the resulting 3D corner

points to the image. Interpolation of heights in DSM with the

Having buildings detected in the scene, the reconstruction part same procedure helps to find roof boundaries of the building in

is based on finding parametric forms of the model planar faces. the image assuming that walls are vertical and there is no eave

These planar faces are then intersected and resulting plane overshooting. Concatenation of these two polygons (footprint

patches are assembled together to form a generic polyhedral and roof boundary) defines the actual area where the building

model. Roof planes are reconstructed by fusion of image and appears in the image.

height data through a split-and-merge process. This process

starts with segmenting the image data within the localized areas. 3.2 Reconstruction of roof planes using image and height

Height points from the DSM are projected to extracted image data

regions and a robust regression method is employed to fit planar

Reconstruction of roof planes is based on image regions

faces to height points belonging to each image region. Regions

extracted in areas where building candidates are detected.

in which more than a single plane is detected are split and

Extraction of image regions is carried out using watershed

neighbouring regions whose planes are coplanar are merged.

segmentation algorithm (Vincent and Soille, 1991). Extended

Vegetation regions are identified and discarded by computing

minima transform (Soille, 1999) is employed to control

an NDVI measure derived from red and infrared channels.

excessive oversegmentation.

Every planar face is attributed based on its slope and height

over the DTM. A planar face is attributed as non-roof if its

While a desirable segmentation is a partitioning of the image

height over the DTM is smaller than a minimum tolerance;

into regions where each region corresponds to a single planar

otherwise it is attributed as flat roof if its largest slope is smaller

face in object space, segmentation algorithms often result in

than a slope threshold or as slanted roof if the largest slope is

undergrown and/or overgrown regions. The purpose of the split-

larger than the slope threshold.

and-merge process is to refine the result of initial segmentation

by making use of clues derived from the DSM. For this purpose,

Wall faces are obtained by reconstructing a vertical wall over

height points are projected from DSM to extracted image

every line segment of the 2D ground plan. The average terrain

regions and a robust regression method is used to fit planar

height, derived from DTM, defines the planar surface that lies

faces to height points belonging to each image region. This

beneath the building. After the parametric forms of all planar

method is based on random selection of a finite set of samples

faces of the building are computed, every three planar faces are

from data (trial estimates) (Fischler and Bolles, 1981). Least

intersected and the resulting vertex is verified to make sure it is

median of squared residuals (Rousseeuw and Leroy, 1987) is

a correct model vertex. Verified vertices of each planar face are

used to find the best sample and also outlier points. Each sample

sorted in order to form a planar patch. Planar patches form the

contains three data points randomly selected from the DSM.

final generic polyhedral model that can be visualized using a

These points define a plane. For other points a residual value is

graphical engine.

calculated as to how they fit into this plane. The sample with the

least median of squared residuals is selected for outlier

The basic assumption in this approach is that buildings are

detection. Outliers are detected as points with residuals larger

formed by planar faces and that walls are vertical. In addition,

than a predefined tolerance and are treated as a new dataset to

building roofs are assumed to be one of the following three

determine whether they fit into a new plane. The plane fitting

types: flat roof, gable roof and hipped roof. In other words this

process is iterated until no more planes can be fitted to data

approach aims for reconstructing simple building types using a

points.

boundary-representation (B-Rep) modelling scheme.

Nevertheless, more complex buildings such as buildings with

After planar faces are detected in each image region, the

cross-gabled roofs can still be reconstructed by adopting a

segmented image is searched for regions in which more than

Constructive Solid Geometry (CSG) modelling scheme. In this

one plane is detected. Those regions are overgrown regions;

way, similar to the method developed by Suveg and Vosselman

hence, they are split into two or more regions depending on the

(2004), the 2D ground plan is first partitioned into rectangular

number of detected planes. To detect and merge undergrown

parts where each part is reconstructed using the plane patch

regions, first a region adjacency graph is constructed by

reconstruction method described above. These building parts are

tracking region boundaries in the segmented image. Plane

then combined together to form the final generic model.

parameters of every two neighbouring regions enter a

coplanarity check and the two neighbouring regions are merged

3. RECONSTRUCTION OF PARAMETRIC FORMS OF

if their associated planes are coplanar.

THE MODEL PLANAR FACES

Buildings are localized in the image using ground plans and An example of the performance of the split-and-merge process

height data. A split-and-merge process is applied to fuse image is demonstrates in figure 1. As can be seen in figure 1(B), the

and height data in the localized areas and derive the parametric initial segmentation results in an overgrown and an undergrown

forms of roof planes. Walls are reconstructed by finding the region in the right part of the roof. The result of the split-and-

parametric forms of vertical planes built on the ground plan. merge process is shown in figure 1(C) where the overgrown

The following sections describe the above processes in more region is split and two undergrown regions are merged to form a

details. correct roof region.

Therefore, each line segment of the ground plan with endpoints

p1 = ( x1, y1 ) T, p2 = (x2, y2 )T can solve for a 1, a 2 and define a

vertical plane as follows:

a1 x + a 2 y + a 3 z + a 4 = 0

where :

Figure 1: The split-and-merge process: A. The original image; a1 = y1 y (5)

B. The initial segmentation; C. The result of the

a 2 = x 2 x1

split-and-merge process.

a3 = 0

a 4 = x1 y 2 x 2 y1

The result of the split-and-merge process is more likely to be a

correct partitioning of the image where each region associates

with a single surface in object space. However, planar faces Parameters of the reconstructed wall faces are stored along with

might have been detected also in vegetation regions. To avoid two endpoints used in the calculations. Wall faces and planar

the influence of vegetation, image data in red and near infrared roof faces computed in the split-and-merge process enter the

channel is used to identify and discard vegetation regions. For plane patch reconstruction procedure described below.

this purpose a normalized difference vegetation index (NDVI) is

computed for each pixel from the following equation: 4. RECONSTRUCTION OF PLANE PATCHES

NIR RED

NDVI = A plane patch is defined as a planar polygon in 3D space. So

(1)

NIR + RED far, the computed model faces are represented in parametric

form. For graphical visualization, however, one requires plane

where NIR and RED denote pixel values in near infrared and red patches, which are represented by their vertices. Reconstruction

channels respectively. NDVI is related to the proportion of of plane patches is carried out in three steps: plane intersection,

photosynthetically absorbed radiation and its value varies from verification of vertices and sorting of vertices. In the following

1 to +1. Vegetation is characterized by high NDVI value and a these steps are described in more details.

region is identified as a vegetated region if at least 70% of its

pixels are vegetation pixels. 4.1 Plane intersection

To determine the final roof planes among the remaining regions Two planes in 3D space intersect in a line if they are not parallel

the difference between DSM height and DTM height for height or coplanar. This line will intersect a third plane in one point if

points belonging to each region is used. A planar surface is it is not parallel to it and does not lie in it. Therefore, in regular

attributed as non-roof if the difference between its average case three planes in 3D space intersect in a point if they are not

DSM height and DTM height is smaller than a minimum in a special relation to each other. In algebraic form, a system of

threshold; otherwise it is attributed as flat-roof if its largest three equations of the form denoted in eq. 2 has exactly one

slope is smaller than a slope threshold or as slanted-roof if the solution if the equations are linearly independent. More

largest slope is larger than the slope threshold. precisely if normal vectors of the three planes are linearly

independent then equations of the planes form a regular system

3.3 Reconstruction of vertical walls upon ground plans of three equations and three unknowns as denoted in eq. 6:

The parametric form of a plane passing through a point a11 x + a12 y + a13 z = k 1

and perpendicular to a normal vector

p = (x, y, z )T

a 21 x + a 22 y + a 23 z = k 2

p p p

(6)

n = ( a1, a 2, a3 ) T can be written as:

a 31 x + a 32 y + a 33 z = k 3

a1 x + a 2 y + a 3 z = k where aij are plane parameters and k i are constants. In order

(2)

to verify the linear independence of normal vectors, let:

where:

a11 a12 a13

= a 21 a 22 a 23

k = a1 x p + a2 y p + a3 z p (7)

(3)

a 31 a 32 a 33

For a vertical plane a 3 = 0 and the plane equation becomes:

The normal vectors of the three planes are linearly independent

if 0 in which case the set of equations 6 has a unique

a1 x + a 2 y = k solution that can be calculated using Cramer s rule (Pedoe,

(4)

1963):

where:

k = a1 x p + a 2 y p

x = x y = y z = z cross products are non-zero and have the same sign. Otherwise

if cross products are non-zero with different signs, the test point

where :

is outside the polygon. If any of the cross products is zero then

k1 a12 a13 the test point is determined to fall on a polygon side or on the

x = k 2 a 22 a 23 extension of a polygon side. The distinction between the two

cases is made by checking the dot product of the corresponding

k 3 a32 a 33

two vectors (which have a zero cross product). For a point on a

side of the polygon, one of the cross products is zero and the

a11 k1 a13 (8)

corresponding dot product is negative, while for a point on the

y = a 21 k 2 a 23 extension of the polygon side the dot product is positive (figure

3).

a31 k 3 a 33

a11 a12 k1

z = a 21 a 22 k 2

a31 a32 k 3

Every three planar faces of the model are intersected and the

intersection point, if there is one, is stored as a vertex for each

of the three faces.

r r

4.2 Verification of vertices

B C D

A

V1 V2 rr rr

rr rr rr rr

V1 V2 V1 V2 V1 V2 V1 V2

V1 o V2 V1 o V2

Intersection of model faces may generate incorrect vertices.

Figure 2(B) shows an example of an incorrect vertex generated

from the intersection of three model faces. All generated + - - - + +

P1 P2

vertices, therefore, have to be verified, in order to identify and

+ + 0 + 0 -

P2 P3

remove incorrect ones. To identify incorrect vertices, two

constraints are used as follows: + + + + + +

P3 P4

+ + + + + +

P4 P1

Constraint 1: A valid vertex lies either in or under any roof

plane.

: Cross Product

Recall that each building part is assumed to have one of the

o

three presumed roof types: flat, gable and hipped. For these : Dot Product

building types, all model vertices lie either in or under any roof

Figure 3: Four possible positions of a point with respect to a

plane. This property allows us to verify model vertices and

polygon. A. Point inside the polygon: all cross

identify invalid ones. The verification is carried out by

products are non-zero and have the same sign; B.

evaluating the function of the roof plane with the coordinates of

Point outside the polygon: cross products are non-

the vertex of interest. This will result in zero if the vertex lies in

zero but have different signs; C. Point on the

the roof plane, otherwise the sign of the resulting value and the

extension of a polygon side: a zeros cross product

direction of the plane normal vector determines whether the

with a positive dot product of the corresponding

vertex point is above or under the roof plane. Every vertex is

vectors; D. Point on the polygon side: a zero cross

verified against all roof planes and is removed if it is higher

product but the dot product of the corresponding

than any of roof planes.

vectors is negative.

A correct vertex is one that satisfies both constraints. An

incorrect vertex will fail to satisfy one or both constraints and

will be removed from corresponding planar faces.

4.3 Sorting of vertices

For graphical visualization of the reconstructed model, vertices

of each planar face must be given in order. To sort vertices a

simple algorithm is used, which is based on forming vectors

from the centre of gravity of vertices to each vertex and finding

Figure 2: Intersection of model faces may generate the angle between each vector and a starting vector (figure 4).

r

r

incorrect vertices. A. A correct vertex; B. An The angle between vectors u and v is given by the dot

incorrect vertex. product:

rr

Constraint 2: A valid vertex projects on or inside the u ov

cos = r r (9)

polygon of the ground plan.

uv

This constraint is based on the assumption that roof eaves do

not overshoot walls. Figure 3 illustrates the procedure to verify

Since cosine function returns the same value for, the sign of

whether a test point is on or inside a polygon. First vectors are

the cross product between the two vectors is used to determine

formed from the test point to every polygon vertex and then

the direction of the angle. The algorithm starts with an arbitrary

cross product of every two adjacent vectors are computed. The

vertex and sorts other vertices with respect to the angle of their

test point is determined inside the polygon, if all computed

vectors with the starting vector. The details of the algorithm is

as follows:

Algorithm: Sort vertices

Select an arbitrary vertex as the starting vertex;

Find the center of gravity of remaining vertices;

Form vectors from the center of gravity to each

vertex;

Compute the angle between each vector and the

starting vector (eq. 9);

Find the sign of the cross product of each vector and

the starting vector;

Figure 5: Colour aerial image of the selected scene.

For any vector with a negative sign of the cross

product change the angle to 2 .

Sort the vertices with respect to their angles.

Sorted vertices of model faces form the planar patches. The

generic polyhedral model is reconstructed once all its planar

patches are formed.

Figure 6: The DSM of the selected scene.

Figure 4: sorting the vertices of a planar face.

5. EXPERIMENTAL RESULTS

Figure 7: The DTM of the selected scene.

The image data used in the experiment is an orthorectified aerial

image of 0.5m ground resolution acquired in RGB and NIR

channels. Figure 5 shows the RGB image of the selected scene.

Height data consists of a last echo DSM and a DTM in regular

grid format with 1.0m resolution acquired using LIDAR (light

detection and ranging) system. Figure 6 and figure 7 depict the

DSM and the DTM respectively in grey levels where a brighter

level stands for a higher altitude. Ground plans of the buildings

in the scene were manually digitised from the image data and Figure 8: The ground plans manually digitised from

are shown in figure 8. Using the information in red and near the image.

infrared channels the NDVI measure was computed for each

image pixel. The resulting NDVI map is shown in figure 9

where high NDVI values (vegetation pixels) are depicted in

yellow and red colours.

Buildings were localized in the image and the split-and-merge

process was applied to localized areas to reconstruct the roof

planes. Image regions with NDVI values higher than 0.35 were

identified as vegetation and were removed from the process.

Figure 10 shows the results of the split-and-merge process as

applied on the images of the three buildings in the scene. As can

be seen in figure 10(B), the initial segmentation results in

undergrown and overgrown regions in all cases. Figure 10(C),

Figure 9: The NDVI map computed from image data in

however, shows that these undergrown and overgrown regions

red and infrared channels.

are successfully split and merged respectively.

The plane patch reconstruction technique was applied to approach, however, detection of buildings using ground plans

parametric forms of roof planes computed in split-and-merge and height data is straightforward and guarantees that other

process and wall planes reconstructed upon ground plans. objects will not be reconstructed as buildings.

Figure 11 shows the generic polyhedral models reconstructed by

assembling plane patches. The focus in the present work was on the reconstruction of

simple roof types. Future research will target more complex

roof types by adopting a CSG modelling scheme. This will

concern finding all possible ways to partition the ground plan

and select the correct one based on the number, parameters and

attributes of reconstructed roof planes. Building parts will then

be reconstructed using the approach described in this paper.

ACKNOWLEDGEMENT

The work described in this paper was supported by a grant from

the Hong Kong Polytechnic University (Project No. G-W122).

Author would also like to thank TopoSys Topographische

Systemdaten GmbH for providing the dataset used in the

experiment.

REFERENCES

Ameri, B., 2000. Automatic recognition and 3D reconstruction

of buildings from digital imagery. PhD Thesis, Stuttgart

University, Stuttgart, 110 pp.

Baillard, C., Schmid, C., Zisserman, A. and Fitzgibbon, A.,

1999. Automatic line matching and 3D reconstruction of

buildings from multiple views, ISPRS conference on automatic

extraction of GIS objects from digital imagery, IAPRS Vol.32,

Part 3-2W5, pp. 69-80.

Figure 10: Results of the split-and-merge process for Bignone, F., Henricsson, O., Fua, P. and Stricker, M., 1996.

reconstruction of roof planes. A. Original Automatic extraction of generic house roofs from high

images; B. Initial segmentations; C. resolution aerial imagery, ECCV '96: 4th European Conference

Reconstructed roofs shown as image regions. on Computer Vision, Cambridge, UK, April 15-18, pp. 85-96.

Brunn, A. and Weidner, U., 1997. Extracting buildings from

digital surface models, ISPRS Workshop on 3D Reconstruction

and Modelling of Topographic Objects, Stuttgart, pp. 27-34.

Cord, M., Jordan, M. and Cocquerez, J.-P., 2001. Accurate

building structure recovery from high resolution aerial imagery.

Computer Vision and Image Understanding, 82: 138-173.

Dang, T., Jamet, O. and Maitre, H., 1994. Applying perceptual

grouping and surface models to the detection and stereo

reconstruction of building in aerial imagery. In: H. Ebner, C.

Heipke and K. Eder (Editors), ISPRS Symposium on Spatial

Information from Digital Photogrammetry and Computer

Vision, Munich, pp. 165-172.

Figure 11: The reconstructed models.

Fischer, A. et al., 1998. Extracting buildings from aerial images

using hierarchical aggregation in 2D and 3D. Computer Vision

6. CONCLUSION and Image Understanding, 72(2): 185-203.

A framework for fusion of image data, height data and 2D Fischler, M.A. and Bolles, R.C., 1981. Random sample

ground plans was presented in this paper. This fusion consensus: a paradigm for model fitting with applications to

framework was used in an approach to automated detection and image analysis and automated cartography. Communications of

reconstruction of buildings. The relative simplicity of this the ACM, 24(6): 381-395.

approach suggests that fusing data from multiple sources can

Fua, P. and Hanson, A.J., 1991. An optimization framework for

greatly reduce the complexity of the problem. Using generic

feature extraction. Machine vision and applications(4): 59-87.

models is advantageous since a larger number of buildings can

be modelled. Nevertheless, successful reconstruction of generic

Haala, N. and Anders, K., 1996. Fusion of 2D GIS and image

models heavily depends on the results of detection. This is

data for 3D building reconstruction, International Archives of

because these models are not sufficiently specific and other

Photogrammetry and Remote Sensing, vol. XXXI, Part B3, pp.

objects might as well be reconstructed as buildings. In this

285 290.

Haala, N. and Brenner, C., 1998. Interpretation of urban surface Conference on Computer Vision, Freiburg, Germany, pp. 410-

425.

models using 2D building information. Computer Vision and

Image Understanding, 72(2): 204-214.

Nevatia, R., Lin, C. and Huertas, A., 1997. A system for

Henricsson, O., 1998. The Role of Color Attributes and building detection from aerial images. In: A. Gruen, E.

Similarity Grouping in 3-D Building Reconstruction. Computer Baltsavias and O. Henricsson (Editors), Automatic extraction of

Vision and Image Understanding, 72(2): 163-184. man-made objects from aerial images (II). Birkhauser Verlag,

Basel, pp. 77-86.

Henricsson, O. and Baltsavias, E., 1997. 3-D building

reconstruction with ARUBA: A qualitative and quantitative Pasko, M. and Gruber, M., 1996. Fusion of 2D GIS data and

evaluation. In: A. Gruen, E. Baltsavias and O. Henricsson aerial images for 3D building reconstruction, International

(Editors), Automatic extraction of man-made objects from aerial Archives of Photogrammetry and Remote Sensing, vol. XXXI,

images(II). Birkhauser Verlag, Basel, pp. 65-76. Part B3,, pp. 461-468.

Herman, M. and Kanade, T., 1986. Incremental reconstruction Pedoe, D., 1963. A geometric introduction to linear algebra.

Wiley, New York, 224 pp.

of 3D scenes from multiple complex images. Artificial

Intelligence(30): 289-341.

Rottensteiner, F. and Jansa, J., 2002. Automatic extraction of

Huertas, A., Lin, C. and Nevatia, R., 1993. Detection of buildings from LIDAR data and aerial images, International

buildings from monocular views of aerial scenes using Archives of Photogrammetry and Remote Sensing, Volume

XXXIV / 4, pp. 569-574.

perceptual organization and shadows, ARPA image

understanding workshop, Washington, DC, pp. 253-260.

Rousseeuw, P.J. and Leroy, A.M., 1987. Robust regression and

outlier detection. John Wiley & Sons, New York, 329 pp.

Jaynes, C., Riseman, E. and Hanson, A., 2003. Recognition and

reconstruction of buildings from multiple aerial images.

Shufelt, J. and McKeown, D., 1993. Fusion of monocular cues

Computer Vision and Image Understanding, 90(1): 68-98.

to detect man-made structures in aerial imagery. CVGIP: Image

Understanding, 57(3): 307-330.

Jibrini, H., Paparoditis, N., Pierrot-Desseilligny, M. and Ma tre,

H., 2000. Automatic building reconstruction from very high

Soille, P., 1999. Morphological image analysis: principles and

resolution aerial stereopairs using cadastral ground plans, 19th

applications. Springer, Berlin, 316 pp.

ISPRS Congress, Amsterdam, The Netherlands.

Suveg, I. and Vosselman, G., 2004. Reconstruction of 3D

Kolbe, T.H., 1999. Identifikation und Rekonstruktion von

building models from aerial images and maps. ISPRS Journal of

Geb uden in Luftbildern mittels unscharfer Constraints. PhD

Photogrammetry and Remote Sensing, 58: 202-224.

Thesis, University of Vechta, Vechta, 149 pp.

Vincent, L. and Soille, P., 1991. Watersheds in digital spaces:

Lin, C., Huertas, A. and Nevatia, R., 1995. Detection of

an efficient algorithm based on immersion simulations. IEEE

buildings from monocular images. In: A. Gruen, O. Kuebler and

transactions on pattern analysis and machine intelligence, 13(6):

P. Agouris (Editors), Automatic extraction of man-made objects

583-598.

from aerial and space images. Birkhauser Verlag, Basel, pp.

125-134.

Vosselman, G., 1999. Building reconstruction using planar

faces in very high density height data, ISPRS Conference on

Lin, C. and Nevatia, R., 1996. Buildings detection and

Automatic Extraction of GIS Objects from Digital Imagery,

description from monocular aerial images, ARPA Image

Munich, pp. 87-92.

Understanding Workshop, Palm Springs, CA.

Weidner, U. and Forstner, W., 1995. Towards automatic

Maas, H.G., 1999. Closed solutions for the determination of

building extraction from high resolution digital elevation

parametric building models from invariant moments of airborne

models. ISPRS Journal of Photogrammetry and Remote

laserscanner data, ISPRS Conference on Automatic Extraction

Sensing, 50(4): 38-49.

of GIS Objects from Digital Imagery, Munich, pp. 193-199.

Moons, T., Frere, D., Vandekerckhove, J. and Van Gool, L.,

1998. Automatic modeling and 3D reconstruction of urban

house roofs from high resolution aerial imagery, European

Contact this candidate