The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
480 
model. Chapter 3 presents the application of the 5-point 
algorithm with infrared image sequences recorded as described 
in chapter 2 and the strategy for quality comparison of the given 
building model’s façades and the estimated surfaces of the 5- 
point algorithm as well as the comparison of the measured 
camera path and the estimated camera path of the algorithm. In 
chapter 4 there are given some experimental results and quality 
measurements using Nistér’s position estimation and chapter 5 
finishes up with a conclusion. 
2. NISTÉR’S 5-POINT ALGORITHM, THE SPECIAL 
BEHAVIOUR OF INFRARED LIGHT AND CAMERAS 
2.1 A Short introduction to Nistér’s 5-point algorithm 
Nistér’s 5-point algorithm was developed as an efficient 
solution to the relative pose estimation problem of a camera 
between two calibrated views using 5 corresponding image 
points. From images only, it is possible to reconstruct only the 
relative orientation of the image pair and thus the relative 
position of the corresponding points and the cameras of the 
views can be determined. The scale of the scene cannot be 
reconstructed as well as of course the absolute positions. This 
limitation to a relative and unsealed orientation is one of the 
main problems for the integration in a given building model. 
The algorithm uses a hypothesis generator within a random 
samples consensus scheme (RANSAC) (Fischler and Bolles, 
1981). The precondition of intrinsic calibration of the camera 
given an improvement of the accuracy and robustness, 
especially for the special case, the algorithm is used for in this 
paper. The calibration of the camera minimizes problems with 
planar scenes and building façades normally appear planar. 
Without calibration the methods fails in coplanar scene points 
as there remain many correct solutions. Using not only image 
pairs but image triplets, the RANSAC scheme with the 5-point 
algorithm resolves all ambiguities. One precondition is a 
sufficient change in the observed scene between the images 
which is normally achieved by changing the camera position 
and viewing direction. A detailed mathematical description of 
the recovering of the translation and rotation of the second and 
third view corresponding to the first view, can be found in 
Nistér (2004). 
2.2 Recorded infrared image sequences 
Current IR cameras cannot reach the optical resolution of video 
cameras or even digital cameras. Like in the visible spectrum, 
the sun affects infrared records. Images in the mid-wave 
infrared are directly affected as in addition to the surface 
radiation caused by the building’s temperature the radiation of 
the sun is reflected. In long-wave infrared the sun’s influence 
appears only indirect, as the sun is not sending in the long wave 
spectrum, but of course is affecting the surface temperature of 
the building. 
Caused by the small field of view and the low optical resolution 
it was necessary to record the scene in oblique view to be able 
to record the complete facades of the building from the floor to 
the roof and to get an acceptable texture resolution. The image 
sequences were recorded with a frequency of 50 frames per 
second. The viewing angle related to the along track axis of the 
van was constant. Figure 1 shows a set of images from the 
sequence. The position of the camera was recorded with GPS 
with an accuracy of 2-5 meters and, for quality measurements 
from tachymeter measurements from ground control points. 
Fig. 1 : Images of one test sequence showing the angular view 
and the camera movement. 
2.3 Description of the given 3d building model 
The information extracted from the infrared image sequences 
has be assigned to the corresponding building in a GIS database. 
To link extracted façade textures and GIS database, the given 
polygonal building model stored in the database is taken. This 
model is given in LOD 2 and represents the façades as one 
polygonal surface with the vertices in world coordinates. 
3. AUTOMATED EXTRACTION OF SURFACES AND 
TEXTURES 
3.1 Application of the 5-point algorithm on infrared image 
sequences 
Nistér’s 5-point pose estimation can be used to extract point 
clouds from image triplets and a relative camera path. Those 
point clouds can then be used to estimate surfaces in a scene 
observed from several images. Mayer (Mayer 2007) has 
introduced an approach for wide-baseline image sequences. In 
this approach, Fôrstner points (Fôrstner and Gülch, 1987) are 
matched via cross-correlation. RANSAC is used with the 
RANSAC scheme of Chum et al. (2003) for the estimation of 
the fundamental matrix F and trifocal tensor T of the image 
triplet. The found inliers are used for a robust bundle 
adjustment (Hartley and Zisserman 2003). To orient the whole 
image sequence, the triplets are linked based on homographies 
and already known 3d points of the already oriented sequence 
part. For the reconstruction of planes from the point clouds, 
vanishing points are detected for groups of images. Because 
building façade are often vertical, the medians in x- and y- 
direction can be takes as the vertical direction. Planes are 
searched defining a maximum distance of a point to a plane. 
The best plane is the plane with the smallest distance to a 
hypothesized plane. From the plane parameters and projection 
matrices homographies are computed between the planes and 
the images. 
All images are recorded from a moving vehicle with the same 
viewing direction. This means, that the camera is only moving 
along a path. In a first glance, this seems to be a simplification. 
But, although the angle between the camera and the moving 
direction is constant, there are changes in the viewing direction 
caused by the movement of the vehicle. So the viewing 
direction cannot be seen as fixed. For the reconstruction of the