isprsarchives XXXIX B3 297 2012

A SEMIAUTOMATIC APPROACH FOR GENERATION OF SITE MODELS FROM

CARTOSAT-2 MULTIVIEW IMAGES

Archana Mahapatra, Sumit Pandey, D Sudheer Reddy, P S Subramanyam, B. K. Das, P V Radhadevi, J Saibaba, Geeta Varadan Advance Data Processing Research Institute, Department of Space,

203 Akbar Road, Secunderabad-500009, INDIA,

archaana@adrin.res.in

Commission III, WG III/4, III/5

KEY WORDS: Site Model, Physical Sensor Model, Relative orientation, conjugate points, edge extraction, normalized DSM ABSTRACT:

In the last decade there has been a paradigm shift in creating, viewing and utilizing geospatial data for planning, navigation and traffic management of urban areas. Realistic, three-dimensional information is preferred over conventional two dimensional maps. The paper describes objectives, methodology and results of an operational system being developed for generation of site model from Cartosat-2 multiview images. The system is designed to work in operational mode with varying level of manual interactivity. A rigorous physical sensor model based on collinearity condition models the ‘step n stare’ mode of image acquisition of the satellite. The relative orientation of the overlapping images is achieved using coplanarity condition and conjugate points. A procedure is developed to perform digitization in mono and stereo modes. A technique for refining manually digitized boundaries is developed. The conjugate points are generated by establishing a correspondence between the points obtained on refined edges to analogous points on the images obtained with view angles ±26 deg. It is achieved through geometrically constrained image matching method. The results are shown for a portion of multi-view images of Washington City obtained from Cartosat-2. The scheme is generic to accept very high resolution stereo images from other satellites as input.

1. INTRODUCTION

A site model represents both the natural and man-made features such as terrain, road, and buildings to sufficient level of detail so that it can be considered as a true model of the area under consideration. Buildings are important objects of any 3D city model. The algorithms for fully automatic extraction of buildings are not matured enough as a result many researchers in this field opt for semi automatic methods (Gruen 99). Building extraction requires representation of roof structures. A finer model represents roof details, overhangs and includes realistic texture of the building (Ulmand Poli, 2006). It can also have architectural details of the building. In a coarse model, detection of the buildings is only possible. A DSM rendered with texture may be considered as coarse city model (Kraub, 2008).

The inputs required for generating finer details includes aerial images, terrestrial images, LIDAR measurements and building plans (Vosselman et al, 2001). The geometric resolution and radiometric quality of the images are important as it should be possible to identify and delineate features.

Aerial images have the advantage of very high geometric resolution and low noise levels. In the last few years, there has been considerable improvement in the technology to provide better resolution images from space platforms. The images from IKONOS, Quickbird, Worldview-1/2, Geoeye-1, and Cartosat-2 have geometric resolution less than one meter. These geometric resolution levels are not sufficient to generate finer models, however extraction and representation of major features are still possible.

The objective of this work is to develop a semiautomatic system to generate site models from satellite stereo images/multiview images in operational mode. At present the focus is mainly on

the extraction and representation of building flat roof tops. It is not envisaged to generate a photorealistic site model, reconstruction of complex building roof structures and individual representation of trees.

The paper describes in brief the imaging geometry of Cartosat-2 and physical sensor model developed, followed by details of computation of relative orientation parameters and comparative analysis of rational function model with physical sensor model. A method of refining digitized edges with the help of canny edges is briefed. The semiautomatic mode of data capture and image matching techniques are described subsequently. A section is devoted to explain generation of DSM and DTM. Finally results are presented for a portion of the Washington city.

2. METHODOLOGY

2.1 Cartosat-2 Imaging Geometry

India’s highest resolution imaging satellite, Cartosat-2, was launched in 2007. The spatial resolution of better than one meter in panchromatic band is achieved through ‘step n stare’ mechanism of image acquisition. In this image acquisition process, the effective ground velocity is reduced by continuously steering camera in the direction opposite to spacecraft motion. Reduction of ground trace velocity in turn improves the spatial resolution of the image in along-track direction. To reduce the pixel cross talk and improve the image quality, the CCD arrays have staggered configuration in the focal plane. Even and odd pixels of a line are recorded by the two detector arrays that are separated by 35µ in the focal plane. It means that the adjacent pixels of a line are not imaged at same instant. If all odd numbered pixels of a line are imaged at instant t, the even numbered pixels will be imaged at instant represented by t + 5×integration time.

International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 XXII ISPRS Congress, 25 August – 01 September 2012, Melbourne, Australia

The spacecraft has three modes of image modes are termed as spot, paintbrush and mu spot mode covers a swath of 9.6 km with s from 6 to 290 km. The paintbrush mode combined swath by imaging adjacent strips multiview mode, the same area is imaged different view angles from the same orbit This mode is useful for computation of he However, due to continuous variation of pi track resolution and the base to height ratio acquisition varies for each imaged line. D modelling the imaging geometry of s complex. A physical sensor model is develo account the dynamic nature of imaging proce

Fig. 1: Cartosat-2 Multiview Imag 2.2 Physical Sensor Model for Cartosat-2 The Cartosat-2 spacecraft is equipped with system, star sensors and gyros to provid orientation information at regular time inte sensor model utilizes this information in coherent manner. The model does not appro the orbit. The osculating nature of the orb converting the position and velocity par varying Keplerian elements, which are interp position at the time of imaging. The orienta available as set of quaternions, which are angles. The residual orientation error is mod pitch, and yaw over a short segment of precise control points are not available the is improved using control points identifi Thematic Mapper (ETM) orthoimages and S The model is based on well known collinear states that the object position, image perspective centre lie on a straight line at the Equation (1) represents the collinear mathematical form.

. . In equation (1), x and y represent the image is the effective focal length of the imaging co-ordinates of the object point and Xp, Yp, of the perspective centre at the time of imag denoted by s. M is the transformation ma image and the object space. The matrix multiplying a series of rotations connectin ordinate systems.

2.3 Relative Orientation of Multiview Im The residual orientation error in the multivi acquisition is highly correlated for the ove

age acquisition; these multiview modes. The th strip length varying ode provides a wider ps from same orbit. In ed from two or three it as shown in Fig. 1. height of the objects. pitch rate, the along-io of multiview image

Due to these factors spacecraft becomes eloped that takes into ocess of Cartosat-2.

aging Mode 2

th satellite positioning ide the position and tervals. The physical in a systematic and roximate the shape of orbit is accounted by parameters to slowly terpolated to know the ntation information is re converted to Euler odelled as bias in roll, f imaging. In case, he positional accuracy tified from Enhanced

SRTM DEM. earity condition which

e position and the the time of imaging.

earity condition in

(1) ge plane coordinates, f ng system, X, Y, Z are , Zp are co-ordinates aging. Scale factor is matrix connecting the rix M is formed by cting intermediate

co-Images

tiview mode of image verlapping images as

these images are acquired within a the same orbit. Thus, it is possible orientation of these images by orientation error as unknown. This advantages; first, it is easy to identi overlapping images and secondly points are not available, relatively for computation of relative heights The developed approach to rela images is based on coplanarity perspective centres and the objec plane. Mathematically the coplanar

, , where , , represents scalar triple vectors joining object point and t first and second images respectiv connecting two perspective centre with respect to differential roll, p pairs of conjugate points are suffi correction. The position of the orientation information is p onboard/ground processed measu images.

2.4 Computation of Rational Po Over the past decade, rational func alternate to physical sensor models fact that physical sensor model information about the camera geom of image acquisition process. O function models are easy to implem commercial satellite imagery p rational function model in place of the system truly sensor independen quantify the results acquired with p rational function model.

Rational polynomial coefficients independent approach (Tao, 2002) computes the orientation param explained in previous section. The obtained for the given object s linearized form of equation (1) (M image positions for a given set o object space co-ordinates are es sensor model. The set of object poi positions are used to compute coefficients. The derived set of rat are used for relating image and ob polynomial coefficients are used possible to use commercially avai from satellites such as Geoeye-1, W for site model generation.

2.5 Image Matching Techniques Digital image matching technique digital surface model and automat points. Digital image matching is c ill posed problem. This problem posed one by imposing regulariz technique is to reduce the dom introducing geometric constraints problem line segments are automat using geometric and photometric c

in a short interval of time from ible to perform analytic relative y considering the differential his approach has two significant ntify the conjugate points in the dly, if precise ground control ly oriented images can be used

ts of the objects.

elatively orient the multiview y condition. It states that the ject point lie on the epipolar narity condition is expressed as 0 (3) ple product and , are the d the perspective centre of the tively, represents the vector tres. Equation (3) is linearized l, pitch and yaw values. Three fficient to compute differential e perspective centres and the primarily obtained from asurements supplied with the

Polynomial Coefficients nction models are being used as els. This is primarily due to the dels are complex; they need ometry and good understanding On the other hand, rational lement and supported by major providers. Moreover, using of physical sensor model makes ent. However, it is important to h physical sensor model and the

ts are computed using terrain 2). The physical sensor model rameters as per the method The image space co-ordinate is t space co-ordinate using the Mahapatra et al, 2004). The t of uniformly spaced grid of estimated using the physical points and corresponding image ute the rational polynomial rational polynomial coefficients object space. Since the rational ed for further processing, it is vailable stereo images obtained , Worldview-1/2, and IKONOS

ues

ues are used for extraction of atic identification of conjugate is considered as mathematically em can be transformed to well rizing constraint. One possible omain of probable match by s. In building reconstruction atically extracted and matched ic constraints (Baillard C, Park, International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012

S, Schmid,C.). Some researchers propos information for edge extraction and lin matching[Ok, A.O, Scholze, S.], our approa to extract and match line segments for two r rooftops are modelled which means that sam to multiple points, thus, it is sufficient to m edge of the building to compute height o Building boundaries are digitized, refined nadir image. These boundaries along information analytically transferred to the relation between geometrically uncorrecte image and ortho image is established using th

If two overlapping images are relatively disparity between two conjugate points is du relief. Based on the geometry either the matched in original stereo pair or the other i to form set of epi-polar images. In case of C angle about roll axis is significant; this mean due to terrain variation will not be restricted Image matching strategy has to take into acquisition process. The disparity map resolution is utilized to guide the matching i cases geometric constraint is used to limit th dense DSM is generated using area based m performed on epi-polar images, while featu matches the edges of buildings in raw as w images using the geometric constraints. F matching techniques, normalized cross corr similarity measure. In both the matching a size is increased dynamically. The sim normalized cross correlation initiates with a two thresholds, termed as noise and acceptan normalized cross correlation coefficient for t than or equal to the acceptance correlation is checked for forward and reverse matching match point if the correlation coefficient acceptance threshold for both the direction coefficient is less than the acceptance thresh the noise threshold, the template size for re space is increased and correlation coeffici The noise threshold is selected as 0.4 and a is selected as 0.9. Initial window size of the by 13 which is increased in steps of two pix repeated for at least three different sized win is repeated for each level of image pyram points are transferred to the next level using interpolation for unmatched point is not do original resolution images.

2.6 Computation of Normalized DSM A Digital Terrain Model (DTM) is the elev landscape which does not include above gro other hand, a Digital Surface Model (DSM) with their heights above the ground as well The man-made objects with different heigh can be detected by applying a threshold Model. The DTM is estimated using mathem The morphological operators help in bring terrain from the DSM. The above ground o using the DSM and morph output. T operators, namely “opening” and “closing” The size of the window depends upon the The normalized DSM is generated by su from the DSM. Segmentation and area buildings of desired size. Finally the build

posed use of colour line segment stereo roach does not attempt o reasons (1) only flat ame height is assigned o match points on the t of the building. (2) ed and connected on g with the height e ortho image as the ed image, epi-polar g the sensor model.

ly oriented then the due to the topographic the set of points are er image is re-sampled f Cartosat-2, often the eans that the disparity ted to one dimension. to account the image p obtained in lower g in next step; In both t the search space. The d matching techniques ature based matching s well as in epi-polar . For both the image orrelation is used as a g approachs, template imilarity measure of a small template and tance threshold. If the r the template is more n threshold, the point ing. It is accepted as a ent is more than the ons. If the correlation eshold but higher than r reference and search ficient is recomputed. d acceptance threshold the search image is 13 pixels. The process is windows. This process amid .The unmatched ing interpolation. The done for matching of

levation model of the ground objects. On the ) includes the objects ell as the topography. ights over the terrain to Digital Surface ematical morphology. nging the background d objects are detected Two morphological ” are used iteratively. e size of the building. subtracting the DTM ea filters detect the ilding outline can be

constructed using the neighbourin definition of the building, the gr gray image is also used. The DTM which is used as an input for co height.

Fig. 6 Derived DSM Fig 2.7 Development of 2-D and 3-D Automatic detection of edges an meaningful entity has been an area Attempts to find a completely auto for these problems is an active digitization option is used to ma outlines in one image. These manu further refined using the Canny op are found only in the neighbourho edge to ensure better localization. other images using the geome matching procedure. In case the i not able to find corresponding digitization is done in 3-D viewing in Z direction and place each v position in depth during digitiza Positions of cursors in left and Options are available to draw poin shows digitized building boun Washington image.

Fig. (2): Digitized buil The process improves edge loca effort of manual identification of ed 2.8 Refinement the Edges Digitization of buildings is the maj fully automatic methods for buildi matured as a result we prefer method. The precision of edges digitized edges to the nearest re manual digitization. This is perform 1. Digitize manually in the proxim 2. Refine the edges that are man densifying and constructing a these points and searching fo image.

uring gradients. For the better gradient information from the TM provides the ground height computing the building/object

Fig. 7 Normalized DSM D Digitization Tool

and grouping them to form a rea of research for a long time.

tomatic and successful solution ive area of research. A 2-D manually digitize the building anually digitized boundaries are operator. In this case, the edges rhood of the manually digitized . These edges are matched in metrically constrained image e image matching procedure is ing edges in another image, ng mode. User can move cursor vertex of the feature at any tization and fuse the cursors. nd right images are recorded. oints, lines and polygons. Fig. 2

undaries for a portion of

uilding boundaries

ocalization and minimizes the f edges precisely.

ajor step in 3D site modelling, lding extraction are not enough r semi automatic digitization es is attained by refining the real edges which are lost in ormed in the following steps, ximity of the intended building.

anually digitized in step-1 by a neighbourhood of each of for real edges in canny edge International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012

3. Join the disconnected edges guided by th A simple case of digitizing (inner and outer e building) a portion of image is shown in t orange boundaries show the manually digiti edge of the building. The upper and lower c figure 3(b) are the annular region defi threshold; the edges shown in green are obta operator [Canny, 1986]. The boundaries Canny operator provide better accura digitization.

Fig. 3: (a) Digitized boundaries, (b) Ref 2.9 Matching of Edges and DSM Generat The DSM is obtained through the matchi images at interval of four pixel units. Fig. 4 DSM. Fig. 5 shows the normalized DSM from subtracting the derived DTM from th that the majority of the buildings can be det DSM. It is intended to use the DSM as boundary extraction at later stage.

The positions of refined edges are known in shown in Fig. 6(a), and the corresponding p in the other image using the image to gro image transformation as shown in figure displays matched points.

The position in one image is matched aro position of the other image. The correlation as 0.9. About 30 % of the points get matc computed for all these points which event height of the building edge. The variations at different points of the same roof top selec of 1m. The average height of all these po height of the object assuming roof to be a building height with respect to ground subtracting ground height derived from DTM

Fig. 8(a) points on refined Fig. 8(b) Estim edges in nadir image on imag

26 deg v

Figure 8(c) Ma edges of the im

26 deg

the digitized edges. er edges of the circular n the figure 3(a). The itized inner and outer r circles in magenta of efined with suitable btained through Canny ies obtained through uracy than manual

Refined boundary ration

ching of the epipolar 4 shows the obtained M which is obtained the DSM. It is clear detected in normalized as cue for building in near nadir image as g points are estimated ground and ground to ure 6(b). Figure 6(c) around the estimated on threshold is chosen atched. The height is entually represent the ns of calculated height lected are of the order points is assigned as a plane surface. The und is obtained by TM.

timated points age acquired with g view angle

Matched points on the image acquired with eg view angle

2.10Site Model Generation Proce The flow chart of site model gene figure 7. The basic inputs are at le images of the area of interest. information is available in anci physical sensor model, the relativ estimated. The rational polynomial terrain independent mode. The ep for these views. A dense DSM normalized DSM and Digital Te buildings are delineated by 2-D procedures. The points on the ed image. The remaining unmatched and refined in 3D viewing mode. T matched edge pairs. The ground lev the building height. The height, de Terrain model are inputs for objec software.

Fig 7: Block diagram of site m

3. RESULTS AND

3.1 Results of Relative Orientati Table 1 shows the results of relat images. Fourteen conjugate poin overlapping images. Five points w residual orientation parameters. remaining conjugate points. Startin near nadir image, the image pos estimated in the second image. compared against the actual posit

C ar t os a t- 2 m u ltiv ie w im ag e a n d an c illar y

info rm ation

S en s or m od elin g an d g en er at io n o f r atio na l po lyn o m ial co eff ic ien ts

G e n er ation o f e pip ola r im a ge s

2- D D igit iz at io n of o utlin e o f b u ildin gs

R efin ing d igit ize d e dg es u s in g C an ny op er at o r

G e om etric a lly c on st r ain ed m a t c hin g

o f ed ge p o int

3 -D d igitiz atio n of u nm atc he d ed ge s

R e fin ing d igitiz ed e dg es u s ing C an n y op er at o r

C o m pu t ing the he igh t of b u ild in g

O b jec t M od elin g an d vis u aliz at io n

ocess

eneration process is depicted in least two Cartosat-2 multiview st. The attitude and position ncillary data files. Using the tive orientation parameters are ial coefficients are computed in epipolar images are generated SM is used for generating Terrain Model. The edges of D digitization and refinement edges are matched in another ed edges are manually digitized The height is computed for the level height is subtracted to get delineated buildings and digital ject modelling and visualization

e model generation system. D DISCUSSION ation

lative orientation of multiview oints were identified on the s were used for computation of s. The results are shown on rting with the image position in position of conjugate point is . The estimated positions are sitions, the difference between

G e ne r ation o f de ns e D S M

G en e ra tio n of n D S M a nd

D T M

International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 XXII ISPRS Congress, 25 August – 01 September 2012, Melbourne, Australia

the actual and estimated position is shown direction. The achieved average value is 0.0 in line and pixel directions respectively. The is 1.392 and 0.99 in line and pixel direction r Table 1 Results of Relative orientation of M Cartosat-2 images (Washington)

Actual line no

Actual pixel no

Estimated

Line no Pixel no Alo

tra

18000.6 154.037 18000.6 153.967

18164.3 509.96 18166.1 507.942

18234.6 507.42 18234.2 508.151

18216.2 875.47 18213.5 874.369

18235.6 1043.02 18234.8 1043.31

18247.2 1134.07 18248.7 1134.07

18285.1 1429.53 18285 1429.88

18528.4 1966.17 18529.6 1967.36

18518.4 2135.51 18517.7 2136.15

std dev Average

3.2 Comparison between rational poly and physical sensor model

The ‘step n stare’ mode of image acquisition mode of imaging. It is observed that in th acquisition, the difference between the sen image position and Rational Polynomial Co image position for same object point is o pixels.

Third order rational polynomial coefficients independent mode (Tao, 2002). In this mode of a grid of equally spaced ground co-ordi using the physical sensor model. At least 200 varying from minimum to maximum value are computed. These points are used t polynomial coefficients. Image position of an points which do not have any point comm points used for fitting the rational polynom computed using the physical sensor model a model. The plot of the difference for near na in Fig. 8(a). The RMS error is 0.1244 and line and pixel direction respectively for near The plot for image with a view angle of 26 Fig. 8(b). The blue line represents the residu and the red line represents the residuals in RMS error is 0.6910 and 0.5043 pixel un direction respectively. These values are hig residuals error between physical sensor derived positions for satellite like IRS-P6 imaging is synchronous (Nagasubramaniam,

Fig. 8(b) Plot for nadir image

wn in line and pixel 0.02 and -0.001 pixels he standard deviation n respectively. Multi-view images of

Residual error (pixel units) Along track

Across track

0 0.07

-1.8 2.018

0.4 -0.731

2.7 1.101

0.8 -0.29

-1.5 0

0.1 -0.35

-1.2 -1.19

0.7 -0.64

1.392 0.99 0.022 -0.001

olynomial coefficient

ion is an asynchronous this mode of image sensor model derived Coefficient computed of the order of 0.5

s are fitted in terrain ode the image position rdinates are computed 200 points with height e in suitable step size to fit the rational f another set of ground mmon with the set of omial coefficients are l and rational function nadir image is shown nd 0.090 pixel unit in

ar nadir image. 26 degree is shown in iduals in line direction in pixel direction. The unit in line and pixel high compared to the or derived and RPC where the mode of , 2007, Liang, 2006)

Fig. 8(b) Plot for image w Fig. 8 Plot of difference between ra derived and physical sensor model 3.3 Height Accuracies, Object M The derived heights along with th and orthoimage are used as inpu visualization. Open source software triangular mesh and adding texture is synthetic and symbolic; it may or structure of the building. Fig. 9 sh site model. The height accuracies to LIDAR data available from observed that the average error is 0 1.8m.

Fig. 9 A view of generated site mo city

4. CONCLU

The paper presents an overall sche of site model from spaceborne m results are shown for Cartosat-2 im procedure obviates the need of prec measurements are suitable for th system is designed in a generic w from other similar satellites if rati are available. At present the ma starting point for capturing the o process too can be automated as t represents detection of buildings edges with Canny operator impr Geometrically constrained imag

e with 26 view angle

rational polynomial coefficient el derived image positions

Modelling and Visualization the building outlines are used nput for object modelling and are Blender is used for creating ure to the building. The texture or may not represent the actual shows a view of the generated es are evaluated with reference m Google Earth website. We s 0.8m and standard deviation is

model of portion of Washington

LUSION

hema and results for generation multiview/stereo images. The image. The relative orientation recise ground control if relative the desired application. The c way to accommodate images rational polynomial coefficients manual digitization process is outline of the buildings, this s the derived normalized DSM gs fairly well. Refinement of proves the edge localization. age matching procedure is International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012

successful in finding conjugate points on edges. The system is primarily designed for spaceborne images, and the available resolution of Cartosat-2 and commercially available images permits extraction and representation of man-made structures. Detail roof reconstruction and representing complex roof structure may not be possible. A good photogrammetric framework provides the desired accuracy. Generated site models represent the area under consideration fairly well and can be used for various planning and other applications.

REFERENCES

Baillard C.,Schmid C., Zisserman A. and Fitzgibbon A.,1999.Automatic line matching and 3D reconstruction of buildings from multiple views, in ISPRS conference on automatic extraction of GIS objects from digital imagery, IAPRS volume 32, Part 3-2W5 , pp. 69-80

Blender 2.6.2, www.blender.org, last accessed on 20 April 2012 Canny J F, 1986, Computational; approach to edge detection, IEEE transactions on Pattern Analysis and Machine Intelligence 8, pp 679-698..

Gruen A , Xinhua Wang, 1999. Cyber City Modeler, a tool for interactive 3-D city model generation. In: German Photogrammetry Week, August, 1999, Stuttgart, Germany GoogleEarth,www.google.com/earth/explore/showcase/3dbuild ings.html, last accessed on 10 January 2012

Kraub Thomas, Manfred Lehner, Peter Reintarz, 2008. Generation of coarse 3-D models of urban areas from high resolution stereo satellite images. In: International archives of the photogrammetry, remote sensing and spatial information sciences, vol XXXVII, part B1, pp 1091-1098, Beijing

Liang-Chien Chen, Tee-Ann TEo, Chien-Liang Liu, May 2006, The Geometrical Comparisions of RSM and RFM for Formosat –2 Satellite Images. Photogrammetric Engineering and Remote Sensing Vol.72 No:5 , 573-579.

Mahapatra Archana, R Ramchandran, and R Krishnan, 2004. Modeling the Uncertainty in Orientation of IRS-1C/1D with A rigorous Photogrammetric Model.PE & RS, 70(8), pp 939-946 Nagasubramanian V. , Radhadevi P.V. , Ramachandran R. , and Krishnan R. 2007. Rational function model for sensor orientation of IRS-P6 Liss-4 imagery. The Photogrammetric Record, 22(120), pp 309-320.

Ok, A.O.; Wegner, J.D.; Heipke, C.; Rottensteiner, F.; Soergel, U.; Toprak, V. (2010): A new straight line reconstruction methodology from multi-spectral stereo aerial images, In: Proceedings of Photogrammetric Computer Vision and Image Analysis Conference, IntArchPhRS vol. XXXVIII, part 3A, Paris, 2010, pp. 25-30

Park, S., Lee, K., and Lee, S., 2000. A Line Feature Matching Technique Based on an Eigenvector Approach, Computer Vision and Image Understanding, 77, pp. 263-283.

Scholze, S., Moons, T., Ade, F., Van Gool, L., 2000. Exploiting Color for Edge Extraction and Line Segment Stereo Matching. In: IAPRS, 815–822.

Schmid, C. and Zisserman, A., 1997. Automatic Line Matching Across Views. In: Proceedings of CVPR, pp. 666–671.

Tao C.V. and HU, Y., July 2002. 3-D Reconstruction methods based on the Rational Function Model, Photogrammetric Engineering and Remote Sensing Vol. 68, No: 7, 705-714.

Ulm Kand Daniela Poli. 3D city with Cybercity modeller. In:1 st EARSeL Workshop of the SIG Urban Remote Sensing Humboldt-Universität zu Berlin, 2-3 March 2006.

Vosselman G, Dijkman D, 2001. 3D building model reconstruction from point clouds and ground plans. In: International Archives of Photogrammetry and Remote Sensing, Vol XXXIV-2/W4 Annapolis, MD

International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 XXII ISPRS Congress, 25 August – 01 September 2012, Melbourne, Australia

A SEMIAUTOMATIC APPROACH FOR GENERATION OF SITE MODELS FROM

CARTOSAT-2 MULTIVIEW IMAGES

Archana Mahapatra, Sumit Pandey, D Sudheer Reddy, P S Subramanyam, B. K. Das, P V Radhadevi, J Saibaba, Geeta Varadan Advance Data Processing Research Institute, Department of Space,

203 Akbar Road, Secunderabad-500009, INDIA,

archaana@adrin.res.in

Commission III, WG III/4, III/5

KEY WORDS: Site Model, Physical Sensor Model, Relative orientation, conjugate points, edge extraction, normalized DSM ABSTRACT:

In the last decade there has been a paradigm shift in creating, viewing and utilizing geospatial data for planning, navigation and traffic management of urban areas. Realistic, three-dimensional information is preferred over conventional two dimensional maps. The paper describes objectives, methodology and results of an operational system being developed for generation of site model from Cartosat-2 multiview images. The system is designed to work in operational mode with varying level of manual interactivity. A rigorous physical sensor model based on collinearity condition models the ‘step n stare’ mode of image acquisition of the satellite. The relative orientation of the overlapping images is achieved using coplanarity condition and conjugate points. A procedure is developed to perform digitization in mono and stereo modes. A technique for refining manually digitized boundaries is developed. The conjugate points are generated by establishing a correspondence between the points obtained on refined edges to analogous points on the images obtained with view angles ±26 deg. It is achieved through geometrically constrained image matching method. The results are shown for a portion of multi-view images of Washington City obtained from Cartosat-2. The scheme is generic to accept very high resolution stereo images from other satellites as input.

1. INTRODUCTION

A site model represents both the natural and man-made features such as terrain, road, and buildings to sufficient level of detail so that it can be considered as a true model of the area under consideration. Buildings are important objects of any 3D city model. The algorithms for fully automatic extraction of buildings are not matured enough as a result many researchers in this field opt for semi automatic methods (Gruen 99). Building extraction requires representation of roof structures. A finer model represents roof details, overhangs and includes realistic texture of the building (Ulmand Poli, 2006). It can also have architectural details of the building. In a coarse model, detection of the buildings is only possible. A DSM rendered with texture may be considered as coarse city model (Kraub, 2008).

The inputs required for generating finer details includes aerial images, terrestrial images, LIDAR measurements and building plans (Vosselman et al, 2001). The geometric resolution and radiometric quality of the images are important as it should be possible to identify and delineate features.

Aerial images have the advantage of very high geometric resolution and low noise levels. In the last few years, there has been considerable improvement in the technology to provide better resolution images from space platforms. The images from IKONOS, Quickbird, Worldview-1/2, Geoeye-1, and Cartosat-2 have geometric resolution less than one meter. These geometric resolution levels are not sufficient to generate finer models, however extraction and representation of major features are still possible.

The objective of this work is to develop a semiautomatic system to generate site models from satellite stereo images/multiview images in operational mode. At present the focus is mainly on

the extraction and representation of building flat roof tops. It is not envisaged to generate a photorealistic site model, reconstruction of complex building roof structures and individual representation of trees.

The paper describes in brief the imaging geometry of Cartosat-2 and physical sensor model developed, followed by details of computation of relative orientation parameters and comparative analysis of rational function model with physical sensor model. A method of refining digitized edges with the help of canny edges is briefed. The semiautomatic mode of data capture and image matching techniques are described subsequently. A section is devoted to explain generation of DSM and DTM. Finally results are presented for a portion of the Washington city.

2. METHODOLOGY

2.1 Cartosat-2 Imaging Geometry

India’s highest resolution imaging satellite, Cartosat-2, was launched in 2007. The spatial resolution of better than one meter in panchromatic band is achieved through ‘step n stare’ mechanism of image acquisition. In this image acquisition process, the effective ground velocity is reduced by continuously steering camera in the direction opposite to spacecraft motion. Reduction of ground trace velocity in turn improves the spatial resolution of the image in along-track direction. To reduce the pixel cross talk and improve the image quality, the CCD arrays have staggered configuration in the focal plane. Even and odd pixels of a line are recorded by the two detector arrays that are separated by 35µ in the focal plane. It means that the adjacent pixels of a line are not imaged at same instant. If all odd numbered pixels of a line are imaged at instant t, the even numbered pixels will be imaged at instant represented by t + 5×integration time.

International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 XXII ISPRS Congress, 25 August – 01 September 2012, Melbourne, Australia

The spacecraft has three modes of image modes are termed as spot, paintbrush and mu spot mode covers a swath of 9.6 km with s from 6 to 290 km. The paintbrush mode combined swath by imaging adjacent strips multiview mode, the same area is imaged different view angles from the same orbit This mode is useful for computation of he However, due to continuous variation of pi track resolution and the base to height ratio acquisition varies for each imaged line. D modelling the imaging geometry of s complex. A physical sensor model is develo account the dynamic nature of imaging proce

Fig. 1: Cartosat-2 Multiview Imag 2.2 Physical Sensor Model for Cartosat-2 The Cartosat-2 spacecraft is equipped with system, star sensors and gyros to provid orientation information at regular time inte sensor model utilizes this information in coherent manner. The model does not appro the orbit. The osculating nature of the orb converting the position and velocity par varying Keplerian elements, which are interp position at the time of imaging. The orienta available as set of quaternions, which are angles. The residual orientation error is mod pitch, and yaw over a short segment of precise control points are not available the is improved using control points identifi Thematic Mapper (ETM) orthoimages and S The model is based on well known collinear states that the object position, image perspective centre lie on a straight line at the Equation (1) represents the collinear mathematical form.

. . In equation (1), x and y represent the image is the effective focal length of the imaging co-ordinates of the object point and Xp, Yp, of the perspective centre at the time of imag denoted by s. M is the transformation ma image and the object space. The matrix multiplying a series of rotations connectin ordinate systems.

2.3 Relative Orientation of Multiview Im The residual orientation error in the multivi acquisition is highly correlated for the ove

age acquisition; these multiview modes. The th strip length varying ode provides a wider ps from same orbit. In ed from two or three it as shown in Fig. 1. height of the objects. pitch rate, the along-io of multiview image

Due to these factors spacecraft becomes eloped that takes into ocess of Cartosat-2.

aging Mode 2

th satellite positioning ide the position and tervals. The physical in a systematic and roximate the shape of orbit is accounted by parameters to slowly terpolated to know the ntation information is re converted to Euler odelled as bias in roll, f imaging. In case, he positional accuracy tified from Enhanced

SRTM DEM. earity condition which

e position and the the time of imaging.

earity condition in

(1) ge plane coordinates, f ng system, X, Y, Z are , Zp are co-ordinates aging. Scale factor is matrix connecting the rix M is formed by cting intermediate

co-Images

tiview mode of image verlapping images as

these images are acquired within a the same orbit. Thus, it is possible orientation of these images by orientation error as unknown. This advantages; first, it is easy to identi overlapping images and secondly points are not available, relatively for computation of relative heights The developed approach to rela images is based on coplanarity perspective centres and the objec plane. Mathematically the coplanar

, , where , , represents scalar triple vectors joining object point and t first and second images respectiv connecting two perspective centre with respect to differential roll, p pairs of conjugate points are suffi correction. The position of the orientation information is p onboard/ground processed measu images.

2.4 Computation of Rational Po Over the past decade, rational func alternate to physical sensor models fact that physical sensor model information about the camera geom of image acquisition process. O function models are easy to implem commercial satellite imagery p rational function model in place of the system truly sensor independen quantify the results acquired with p rational function model.

Rational polynomial coefficients independent approach (Tao, 2002) computes the orientation param explained in previous section. The obtained for the given object s linearized form of equation (1) (M image positions for a given set o object space co-ordinates are es sensor model. The set of object poi positions are used to compute coefficients. The derived set of rat are used for relating image and ob polynomial coefficients are used possible to use commercially avai from satellites such as Geoeye-1, W for site model generation.

2.5 Image Matching Techniques Digital image matching technique digital surface model and automat points. Digital image matching is c ill posed problem. This problem posed one by imposing regulariz technique is to reduce the dom introducing geometric constraints problem line segments are automat using geometric and photometric c

in a short interval of time from ible to perform analytic relative y considering the differential his approach has two significant ntify the conjugate points in the dly, if precise ground control ly oriented images can be used

ts of the objects.

elatively orient the multiview y condition. It states that the ject point lie on the epipolar narity condition is expressed as 0 (3) ple product and , are the d the perspective centre of the tively, represents the vector tres. Equation (3) is linearized l, pitch and yaw values. Three fficient to compute differential e perspective centres and the primarily obtained from asurements supplied with the

Polynomial Coefficients nction models are being used as els. This is primarily due to the dels are complex; they need ometry and good understanding On the other hand, rational lement and supported by major providers. Moreover, using of physical sensor model makes ent. However, it is important to h physical sensor model and the

ts are computed using terrain 2). The physical sensor model rameters as per the method The image space co-ordinate is t space co-ordinate using the Mahapatra et al, 2004). The t of uniformly spaced grid of estimated using the physical points and corresponding image ute the rational polynomial rational polynomial coefficients object space. Since the rational ed for further processing, it is vailable stereo images obtained , Worldview-1/2, and IKONOS

ues

ues are used for extraction of atic identification of conjugate is considered as mathematically em can be transformed to well rizing constraint. One possible omain of probable match by s. In building reconstruction atically extracted and matched ic constraints (Baillard C, Park, International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012

S, Schmid,C.). Some researchers propos information for edge extraction and lin matching[Ok, A.O, Scholze, S.], our approa to extract and match line segments for two r rooftops are modelled which means that sam to multiple points, thus, it is sufficient to m edge of the building to compute height o Building boundaries are digitized, refined nadir image. These boundaries along information analytically transferred to the relation between geometrically uncorrecte image and ortho image is established using th

If two overlapping images are relatively disparity between two conjugate points is du relief. Based on the geometry either the matched in original stereo pair or the other i to form set of epi-polar images. In case of C angle about roll axis is significant; this mean due to terrain variation will not be restricted Image matching strategy has to take into acquisition process. The disparity map resolution is utilized to guide the matching i cases geometric constraint is used to limit th dense DSM is generated using area based m performed on epi-polar images, while featu matches the edges of buildings in raw as w images using the geometric constraints. F matching techniques, normalized cross corr similarity measure. In both the matching a size is increased dynamically. The sim normalized cross correlation initiates with a two thresholds, termed as noise and acceptan normalized cross correlation coefficient for t than or equal to the acceptance correlation is checked for forward and reverse matching match point if the correlation coefficient acceptance threshold for both the direction coefficient is less than the acceptance thresh the noise threshold, the template size for re space is increased and correlation coeffici The noise threshold is selected as 0.4 and a is selected as 0.9. Initial window size of the by 13 which is increased in steps of two pix repeated for at least three different sized win is repeated for each level of image pyram points are transferred to the next level using interpolation for unmatched point is not do original resolution images.

2.6 Computation of Normalized DSM A Digital Terrain Model (DTM) is the elev landscape which does not include above gro other hand, a Digital Surface Model (DSM) with their heights above the ground as well The man-made objects with different heigh can be detected by applying a threshold Model. The DTM is estimated using mathem The morphological operators help in bring terrain from the DSM. The above ground o using the DSM and morph output. T operators, namely “opening” and “closing” The size of the window depends upon the The normalized DSM is generated by su from the DSM. Segmentation and area buildings of desired size. Finally the build

posed use of colour line segment stereo roach does not attempt o reasons (1) only flat ame height is assigned o match points on the t of the building. (2) ed and connected on g with the height e ortho image as the ed image, epi-polar g the sensor model.

ly oriented then the due to the topographic the set of points are er image is re-sampled f Cartosat-2, often the eans that the disparity ted to one dimension. to account the image p obtained in lower g in next step; In both t the search space. The d matching techniques ature based matching s well as in epi-polar . For both the image orrelation is used as a g approachs, template imilarity measure of a small template and tance threshold. If the r the template is more n threshold, the point ing. It is accepted as a ent is more than the ons. If the correlation eshold but higher than r reference and search ficient is recomputed. d acceptance threshold the search image is 13 pixels. The process is windows. This process amid .The unmatched ing interpolation. The done for matching of

levation model of the ground objects. On the ) includes the objects ell as the topography. ights over the terrain to Digital Surface ematical morphology. nging the background d objects are detected Two morphological ” are used iteratively. e size of the building. subtracting the DTM ea filters detect the ilding outline can be

constructed using the neighbourin definition of the building, the gr gray image is also used. The DTM which is used as an input for co height.

Fig. 6 Derived DSM Fig 2.7 Development of 2-D and 3-D Automatic detection of edges an meaningful entity has been an area Attempts to find a completely auto for these problems is an active digitization option is used to ma outlines in one image. These manu further refined using the Canny op are found only in the neighbourho edge to ensure better localization. other images using the geome matching procedure. In case the i not able to find corresponding digitization is done in 3-D viewing in Z direction and place each v position in depth during digitiza Positions of cursors in left and Options are available to draw poin shows digitized building boun Washington image.

Fig. (2): Digitized buil The process improves edge loca effort of manual identification of ed 2.8 Refinement the Edges Digitization of buildings is the maj fully automatic methods for buildi matured as a result we prefer method. The precision of edges digitized edges to the nearest re manual digitization. This is perform 1. Digitize manually in the proxim 2. Refine the edges that are man densifying and constructing a these points and searching fo image.

uring gradients. For the better gradient information from the TM provides the ground height computing the building/object

Fig. 7 Normalized DSM D Digitization Tool

and grouping them to form a rea of research for a long time.

tomatic and successful solution ive area of research. A 2-D manually digitize the building anually digitized boundaries are operator. In this case, the edges rhood of the manually digitized . These edges are matched in metrically constrained image e image matching procedure is ing edges in another image, ng mode. User can move cursor vertex of the feature at any tization and fuse the cursors. nd right images are recorded. oints, lines and polygons. Fig. 2

undaries for a portion of

uilding boundaries

ocalization and minimizes the f edges precisely.

ajor step in 3D site modelling, lding extraction are not enough r semi automatic digitization es is attained by refining the real edges which are lost in ormed in the following steps, ximity of the intended building.

anually digitized in step-1 by a neighbourhood of each of for real edges in canny edge International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012

3. Join the disconnected edges guided by th A simple case of digitizing (inner and outer e building) a portion of image is shown in t orange boundaries show the manually digiti edge of the building. The upper and lower c figure 3(b) are the annular region defi threshold; the edges shown in green are obta operator [Canny, 1986]. The boundaries Canny operator provide better accura digitization.

Fig. 3: (a) Digitized boundaries, (b) Ref 2.9 Matching of Edges and DSM Generat The DSM is obtained through the matchi images at interval of four pixel units. Fig. 4 DSM. Fig. 5 shows the normalized DSM from subtracting the derived DTM from th that the majority of the buildings can be det DSM. It is intended to use the DSM as boundary extraction at later stage.

The positions of refined edges are known in shown in Fig. 6(a), and the corresponding p in the other image using the image to gro image transformation as shown in figure displays matched points.

The position in one image is matched aro position of the other image. The correlation as 0.9. About 30 % of the points get matc computed for all these points which event height of the building edge. The variations at different points of the same roof top selec of 1m. The average height of all these po height of the object assuming roof to be a building height with respect to ground subtracting ground height derived from DTM

Fig. 8(a) points on refined Fig. 8(b) Estim edges in nadir image on imag

26 deg v

Figure 8(c) Ma edges of the im

26 deg

the digitized edges. er edges of the circular n the figure 3(a). The itized inner and outer r circles in magenta of efined with suitable btained through Canny ies obtained through uracy than manual

Refined boundary ration

ching of the epipolar 4 shows the obtained M which is obtained the DSM. It is clear detected in normalized as cue for building in near nadir image as g points are estimated ground and ground to ure 6(b). Figure 6(c) around the estimated on threshold is chosen atched. The height is entually represent the ns of calculated height lected are of the order points is assigned as a plane surface. The und is obtained by TM.

timated points age acquired with g view angle

Matched points on the image acquired with eg view angle

2.10Site Model Generation Proce The flow chart of site model gene figure 7. The basic inputs are at le images of the area of interest. information is available in anci physical sensor model, the relativ estimated. The rational polynomial terrain independent mode. The ep for these views. A dense DSM normalized DSM and Digital Te buildings are delineated by 2-D procedures. The points on the ed image. The remaining unmatched and refined in 3D viewing mode. T matched edge pairs. The ground lev the building height. The height, de Terrain model are inputs for objec software.

Fig 7: Block diagram of site m

3. RESULTS AND

3.1 Results of Relative Orientati Table 1 shows the results of relat images. Fourteen conjugate poin overlapping images. Five points w residual orientation parameters. remaining conjugate points. Startin near nadir image, the image pos estimated in the second image. compared against the actual posit

C ar t os a t- 2 m u ltiv ie w im ag e a n d an c illar y

info rm ation

S en s or m od elin g an d g en er at io n o f r atio na l po lyn o m ial co eff ic ien ts

G e n er ation o f e pip ola r im a ge s

2- D D igit iz at io n of o utlin e o f b u ildin gs

R efin ing d igit ize d e dg es u s in g C an ny op er at o r

G e om etric a lly c on st r ain ed m a t c hin g

o f ed ge p o int

3 -D d igitiz atio n of u nm atc he d ed ge s

R e fin ing d igitiz ed e dg es u s ing C an n y op er at o r

C o m pu t ing the he igh t of b u ild in g

O b jec t M od elin g an d vis u aliz at io n

ocess

eneration process is depicted in least two Cartosat-2 multiview st. The attitude and position ncillary data files. Using the tive orientation parameters are ial coefficients are computed in epipolar images are generated SM is used for generating Terrain Model. The edges of D digitization and refinement edges are matched in another ed edges are manually digitized The height is computed for the level height is subtracted to get delineated buildings and digital ject modelling and visualization

e model generation system. D DISCUSSION ation

lative orientation of multiview oints were identified on the s were used for computation of s. The results are shown on rting with the image position in position of conjugate point is . The estimated positions are sitions, the difference between

G e ne r ation o f de ns e D S M

G en e ra tio n of n D S M a nd

D T M

International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 XXII ISPRS Congress, 25 August – 01 September 2012, Melbourne, Australia

the actual and estimated position is shown direction. The achieved average value is 0.0 in line and pixel directions respectively. The is 1.392 and 0.99 in line and pixel direction r Table 1 Results of Relative orientation of M Cartosat-2 images (Washington)

Actual line no

Actual pixel no

Estimated

Line no Pixel no Alo

tra

18000.6 154.037 18000.6 153.967

18164.3 509.96 18166.1 507.942

18234.6 507.42 18234.2 508.151

18216.2 875.47 18213.5 874.369

18235.6 1043.02 18234.8 1043.31

18247.2 1134.07 18248.7 1134.07

18285.1 1429.53 18285 1429.88

18528.4 1966.17 18529.6 1967.36

18518.4 2135.51 18517.7 2136.15

std dev Average

3.2 Comparison between rational poly and physical sensor model

The ‘step n stare’ mode of image acquisition mode of imaging. It is observed that in th acquisition, the difference between the sen image position and Rational Polynomial Co image position for same object point is o pixels.

Third order rational polynomial coefficients independent mode (Tao, 2002). In this mode of a grid of equally spaced ground co-ordi using the physical sensor model. At least 200 varying from minimum to maximum value are computed. These points are used t polynomial coefficients. Image position of an points which do not have any point comm points used for fitting the rational polynom computed using the physical sensor model a model. The plot of the difference for near na in Fig. 8(a). The RMS error is 0.1244 and line and pixel direction respectively for near The plot for image with a view angle of 26 Fig. 8(b). The blue line represents the residu and the red line represents the residuals in RMS error is 0.6910 and 0.5043 pixel un direction respectively. These values are hig residuals error between physical sensor derived positions for satellite like IRS-P6 imaging is synchronous (Nagasubramaniam,

Fig. 8(b) Plot for nadir image

wn in line and pixel 0.02 and -0.001 pixels he standard deviation n respectively. Multi-view images of

Residual error (pixel units) Along track

Across track

0 0.07

-1.8 2.018

0.4 -0.731

2.7 1.101

0.8 -0.29

-1.5 0

0.1 -0.35

-1.2 -1.19

0.7 -0.64

1.392 0.99 0.022 -0.001

olynomial coefficient

ion is an asynchronous this mode of image sensor model derived Coefficient computed of the order of 0.5

s are fitted in terrain ode the image position rdinates are computed 200 points with height e in suitable step size to fit the rational f another set of ground mmon with the set of omial coefficients are l and rational function nadir image is shown nd 0.090 pixel unit in

ar nadir image. 26 degree is shown in iduals in line direction in pixel direction. The unit in line and pixel high compared to the or derived and RPC where the mode of , 2007, Liang, 2006)

Fig. 8(b) Plot for image w Fig. 8 Plot of difference between ra derived and physical sensor model 3.3 Height Accuracies, Object M The derived heights along with th and orthoimage are used as inpu visualization. Open source software triangular mesh and adding texture is synthetic and symbolic; it may or structure of the building. Fig. 9 sh site model. The height accuracies to LIDAR data available from observed that the average error is 0 1.8m.

Fig. 9 A view of generated site mo city

4. CONCLU

The paper presents an overall sche of site model from spaceborne m results are shown for Cartosat-2 im procedure obviates the need of prec measurements are suitable for th system is designed in a generic w from other similar satellites if rati are available. At present the ma starting point for capturing the o process too can be automated as t represents detection of buildings edges with Canny operator impr Geometrically constrained imag

e with 26 view angle

rational polynomial coefficient el derived image positions

Modelling and Visualization the building outlines are used nput for object modelling and are Blender is used for creating ure to the building. The texture or may not represent the actual shows a view of the generated es are evaluated with reference m Google Earth website. We s 0.8m and standard deviation is

model of portion of Washington

LUSION

hema and results for generation multiview/stereo images. The image. The relative orientation recise ground control if relative the desired application. The c way to accommodate images rational polynomial coefficients manual digitization process is outline of the buildings, this s the derived normalized DSM gs fairly well. Refinement of proves the edge localization. age matching procedure is International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012

successful in finding conjugate points on edges. The system is primarily designed for spaceborne images, and the available resolution of Cartosat-2 and commercially available images permits extraction and representation of man-made structures. Detail roof reconstruction and representing complex roof structure may not be possible. A good photogrammetric framework provides the desired accuracy. Generated site models represent the area under consideration fairly well and can be used for various planning and other applications.

REFERENCES

Baillard C.,Schmid C., Zisserman A. and Fitzgibbon A.,1999.Automatic line matching and 3D reconstruction of buildings from multiple views, in ISPRS conference on automatic extraction of GIS objects from digital imagery, IAPRS volume 32, Part 3-2W5 , pp. 69-80

Blender 2.6.2, www.blender.org, last accessed on 20 April 2012 Canny J F, 1986, Computational; approach to edge detection, IEEE transactions on Pattern Analysis and Machine Intelligence 8, pp 679-698..

Gruen A , Xinhua Wang, 1999. Cyber City Modeler, a tool for interactive 3-D city model generation. In: German Photogrammetry Week, August, 1999, Stuttgart, Germany GoogleEarth,www.google.com/earth/explore/showcase/3dbuild ings.html, last accessed on 10 January 2012

Kraub Thomas, Manfred Lehner, Peter Reintarz, 2008. Generation of coarse 3-D models of urban areas from high resolution stereo satellite images. In: International archives of the photogrammetry, remote sensing and spatial information sciences, vol XXXVII, part B1, pp 1091-1098, Beijing

Liang-Chien Chen, Tee-Ann TEo, Chien-Liang Liu, May 2006, The Geometrical Comparisions of RSM and RFM for Formosat –2 Satellite Images. Photogrammetric Engineering and Remote Sensing Vol.72 No:5 , 573-579.

Mahapatra Archana, R Ramchandran, and R Krishnan, 2004. Modeling the Uncertainty in Orientation of IRS-1C/1D with A rigorous Photogrammetric Model.PE & RS, 70(8), pp 939-946 Nagasubramanian V. , Radhadevi P.V. , Ramachandran R. , and Krishnan R. 2007. Rational function model for sensor orientation of IRS-P6 Liss-4 imagery. The Photogrammetric Record, 22(120), pp 309-320.

Ok, A.O.; Wegner, J.D.; Heipke, C.; Rottensteiner, F.; Soergel, U.; Toprak, V. (2010): A new straight line reconstruction methodology from multi-spectral stereo aerial images, In: Proceedings of Photogrammetric Computer Vision and Image Analysis Conference, IntArchPhRS vol. XXXVIII, part 3A, Paris, 2010, pp. 25-30

Park, S., Lee, K., and Lee, S., 2000. A Line Feature Matching Technique Based on an Eigenvector Approach, Computer Vision and Image Understanding, 77, pp. 263-283.

Scholze, S., Moons, T., Ade, F., Van Gool, L., 2000. Exploiting Color for Edge Extraction and Line Segment Stereo Matching. In: IAPRS, 815–822.

Schmid, C. and Zisserman, A., 1997. Automatic Line Matching Across Views. In: Proceedings of CVPR, pp. 666–671.

Tao C.V. and HU, Y., July 2002. 3-D Reconstruction methods based on the Rational Function Model, Photogrammetric Engineering and Remote Sensing Vol. 68, No: 7, 705-714.

Ulm Kand Daniela Poli. 3D city with Cybercity modeller. In:1 st EARSeL Workshop of the SIG Urban Remote Sensing Humboldt-Universität zu Berlin, 2-3 March 2006.

Vosselman G, Dijkman D, 2001. 3D building model reconstruction from point clouds and ground plans. In: International Archives of Photogrammetry and Remote Sensing, Vol XXXIV-2/W4 Annapolis, MD

International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 XXII ISPRS Congress, 25 August – 01 September 2012, Melbourne, Australia