In: Paparoditis N., Pierrot-Deseilligny M„ Mallet C.. Tournaire O. (Eds). 1APRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France, September 1-3, 2010 
data are used to estimate an approximated camera position. In 
the second step 2D points in the image are extracted and 
matched with 3D points of the model using Hough transform 
and generalized M-estimator. Then the Lowe’s (1987) 
algorithm is applied for refinement of the camera parameters. 
This approach yields very good results for in downtown area; 
however, it fails at residential region because of not enough 
extracted vertical edges. 
In texture mapping using thermal images the properties specific 
to the IR spectrum should be taken into consideration. First of 
all, IR images have lower contrast and lower resolution than 
images in visible spectrum. Consequently, matching with 3D 
building model based on edge matching (Frueh et al., 2004) or 
on vertical vanishing points (Ding & Zakhor, 2008) could be 
difficult. Stilla et al. (2000) proposed a method for matching of 
low resolution IR images based on intersection points of roof 
edges. 
For texture mapping a visibility analysis is necessary. Generally, 
there are two groups of methods for checking of the visibility: 
(i) variations of depth-buffer (depth image) approach (Frueh et 
al., 2004; Hoegner & Stilla, 2007; Karras et al., 2007) and (ii) 
polygon-based hidden area detection (Kuzmin et al., 2004). In 
the polygon-based method proposed by Kuzmin et al. (2004) all 
polygons are projected onto image plane and intersected. This 
procedure is appropriate for nadir view images, because of 
small number of intersections. However, using oblique view 
this method would be very time consuming and could cause 
many small polygons. 
The depth-buffer method is a basic method removing hidden 
surfaces adopted from computer graphics. The depth-buffer is a 
matrix storing for every pixel the distance from projection 
centre to the model surface. This method was often proposed in 
some variations. Karras et al. (2007) tries to generalize the 
problem of orthorectification and texture mapping. He proposes 
a method for visibility checking based on depth image. Every 
triangulated 3D mesh is projected onto projection plane and for 
every triangle occupied pixels get identity number (ID) of the 
triangle. For pixels with more IDs the closest one is chosen. 
Frueh et al. (2004) used a modified depth-buffer storing 
additionally the product of a triangle’s normal vector with the 
camera viewing direction at each pixel. Using information about 
vector product not occluded edges can be detected. Abdelhafiz 
& Niemeier (2009) integrate digital images and laser scanning 
point clouds. They use a Multi Layer 3GImage algorithm which 
classifies the visibility on two stages: point stage and surface 
stage. The visible layer and back layers are applied. Occluded 
vertexes are sent to a back layer, while visible vertexes appear 
on the visible layer. An image is used for texture mapping of a 
mesh, if all three vertexes of it are visible in this image. 
Abdelhafiz & Niemeier discuss also the problem of extrinsic 
(un-modelled) occlusions caused by such objects as traffic 
signs, trees and street-lamps. They propose a Photo Occlusion 
Finder algorithm which checks textures from many images for 
one mesh. When the textures of one mesh are not similar an 
occlusion occurred. 
Objects taken by image sequences with a high frame rate from a 
flying platform appear in multiple frames. In this case textures 
with optimal quality have to be taken for texturing. Lorenz & 
Doellner (2006) introduced a local effective resolution and 
discuss it on example of images from a High Resolution Stereo 
Camera (HRSC) due to its special projection of line scanners 
(perspective and parallel). Frueh et al. (2004) uses a focal plane 
array. He determines optimal textures taking into account 
occlusion, image resolution, surface normal orientation and 
coherence with neighbouring triangles. He proposes to accept 
textures with few occluded pixels instead textures with very low 
resolution taken from extremely oblique view. This quality 
calculation is focused on texturing with optical images and 
good user perception. 
In this paper we propose a texture selection method for thermal 
inspection of buildings using a weighted quality function. This 
approach allows reducing or increasing the influence of 
occlusions, distance and viewing directions on the texture 
quality. In Chapter 2 a necessary for texturing camera 
calibration, positioning and orientation is described. In chapter 
3 a concept for texture mapping is introduced. Moreover, the 
influence of oblique view imagery on texture resolution is 
discussed. The equation for weighted quality measure is 
presented. Finally, in Chapter 4 experiments with some 
exemplary textures are shown and discussed in chapter 5. 
2. SYSTEM CALIBRATION 
In most cases, how already mentioned, GPS/INS data do not 
refer to the projection centre. Consequently, boresight and 
leverarm parameters are required. In addition, camera 
parameters, such as focal length, principle point, and 
distortions, need to be determined. As the solution we propose a 
system calibration using an extended bundle adjustment with 
camera self calibration which is described by Kolecki et al. 
(2010). In this method in few images of the sequence ground 
control points (GCP) need to be measured and all parameters of 
exterior and interior orientation as well as boresight and 
leverarm corrections should be estimated. Parameters obtained 
in the adjustment should be applied for projection onto all 
images of the sequence. 
3. A CONCEPT FOR TEXTURE MAPPING 
The region within an IR frame corresponding to a face of the 
3D model can be determined by projection of the polygons into 
the image. In datasets captured by moving cameras with high 
frame rate most polygons of the model appear many times in the 
images with different aspect angles. This advantage allows 
choosing the texture captured from the best pose. Additionally 
in some cases and bridge the problem of occlusions can be 
resolved. The quality of the textures extracted from different 
frames belonging to the same plane varies depending on 
viewing direction, distance to the camera, and partial 
occlusions. For selecting the best texture a quality measure has 
to be defined and the selection procedure has to be 
implemented. A flowchart of this procedure is depicted in Fig. 
1. 
Starting from the first frame for each face a projection is carried 
out. If the face lies within the frame a partial occlusion Oq 
(Chapter 3.1) and quality measure (Chapter 3.3) are 
calculated. In case that the quality q tj of is higher than the 
quality of the currently stored texture, new texture ty is created 
and the current texture is overwrite with ty. 
3.1 Occlusions 
Every face is projected into the image and pixels occupied by 
this plane get the ID of this plane and its distance from the 
projection centre. A pixel is considered as occupied if its centre