479 
CASE STUDY OF THE 5-POINT ALGORITHM FOR TEXTURING EXISTING 
BUILDING MODELS FROM INFRARED IMAGE SEQUENCES 
Hoegner Ludwig, Stilla Uwe 
Technische Universitaet Muenchen, GERMANY - Ludwig.Hoegner@bv.tu-muenchen.de, stilla@tum.de 
Commission III, WG III/4 
KEY WORDS: Urban, Image, Acquisition, Texture, Extraction, Thermal, Infrared 
ABSTRACT: 
Today, thermal inspections of buildings are normally done in recorded single infrared images directly. Thus, no 3d references of 
found objects and features like i.e. heating pipes or leakages is possible. In computer vision several techniques for the extraction of 
building surfaces and surface textures from optical images have been developed during the last years. Those algorithms like i.e. 3- 
point matching, surface estimation via homography or the 5-point algorithm introduced by Nister are specialised for optical images 
with their strong edges and high resolution. In this paper, the 5-point algorithm introduced by Nister is adopted for the extraction of 
textures from infrared images sequences for an already given building model. Special problem caused by the physical behaviour of 
the infrared spectrum and the technical limitations of the cameras will be discussed including their influence on the usability of the 
matching algorithm. 
1. INTRODUCTION 
One focus in today’s discussion of global warming and climate 
lies on thermal inspection of single buildings on the one hand 
and urban environment on the other hand. With ground cameras 
the irradiation of building facades can be investigated (Klingert, 
2006) analyzing infrared images. Airborne IR-systems are 
applied for vehicle detection (Hinz and Stilla, 2006, Stilla and 
Michaelsen, 2002) or exploration of leakages in district heating 
systems (Koskeleinen, 1992). Satellite images are used for the 
analysis of urban heat islands (Lo and Quattrochi, 2003). Janet 
Nichol (Nichol and Wong, 2005) first introduced a method to 
integrate 3d information and infrared images. Satellite IR data 
are combined with simplified block models of building in a 3d 
city models. The satellite images however only allow to assign 
a building roof a temperature but not to search for structures on 
the roof. Façades remain almost invisible. 
To inspect and analyze the thermal behaviour of building 
façades in detail, it is necessary to record them with ground 
based cameras. In difference to airborne and satellite images, 
ground images normally do not contain a complete building in a 
single image. Therefore, it is necessary to combine several 
images to extract the complete texture for a façade. This 
combination needs the knowledge of the parameters of the 
camera used for the record to correctly project the images into 
the scene. 
Techniques for position estimation, matching and scene 
reconstruction have been in use in image processing of optical 
images for a couple of years. The estimation of exterior 
orientation from a single image works with at least 3 
correspondences (3-point algorithm) between image and model 
(Haralick et al, 1994). Techniques for 4- and 5-point estimation 
are elicited by Quan (Quan and Lan, 1999) and Triggs (1999). 
For 6 and more correspondence points the Direct Linear 
Transformation (DLT) can be applied (Triggs, 1999). For 
homogenious façade structures that approximately form a plane, 
homography can be adopted to detect planes in image pairs and 
the relative exterior orientation of the camera in relation to 
these planes (Hartley and Zisserman, 2000). Another popular 
strategy working on image pairs is Nistér’s 5-point position 
estimation (Nistér 2004). This algorithm searches for pairs of 
points of interest in image pair like the homography, but can 
handle several planes visible in the image pair. Due to the small 
field of view, the low spatial resolution of the IR images and 
the low level of detail of the given building model, only few 
point correspondences between IR image and 3D model can be 
identified. Strategies based on the orientation of the image 
sequence itself like homography and Nistér are more useful for 
the given scenario. 
This link between 3d building models and infrared image 
sequences allows dealing with the analysis big building 
complexes that cannot be observed in one single image. By 
integrating the infrared image data and the 3d model data, it 
becomes possible to assign infrared information to a building 
and store them together in a GIS database. Images taken with 
different aspect ratio, from different IR bands or taken at 
different time can be combined for analysis. Several effects of 
warming and cooling of façades can be described including 3d 
model information, i.e. shadows caused by occlusion. 
This paper focuses on the usability of the 5-point algorithm for 
image sequences with constant viewing direction, low contrast 
and low resolution and the integration of a given building 
model. Surface hypotheses are not used to create a building 
model from the images and image sequences, but are used to 
match the relative oriented scene generated by the 5-point 
algorithm with a given building model from a GIS database. We 
will concentrate on the evaluation of the quality of the relative 
estimated orientation of cameras and estimated extracted 
building façades according the given building model of the GIS 
database and the measured path of the recording infrared 
camera. 
In chapter 2, the 5-point position estimation is briefly described 
and the special behaviour and problems with recording building 
façades in the infrared spectrum are introduced together with 
the a short look at the used camera as well as the given building