CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation 
86 
The finding of the maxima in the correlation image with sub 
pixel accuracy by approximating the discrete correlation 
function by a continuous polynomial function is broadly 
discussed in Kraus (1996) and will not be described here. 
Figure 7 shows an example were one search image was matched 
with multiple reference images. 
Figure 7: Correlation example. The brighter the pixels, the 
higher the correlation, i.e. the greater the similarity 
between search image and reference image. 
For the practical implementation, the reference images of 
buildings are stored in a library. For each of these buildings also 
vector information (describing the building outline in the 
reference image coordinate system) is available. Thus the 
library contains multiple building types to which one image 
patch and one vector representation corresponds. 
Once a location of high correlation is found in the search image, 
the vector data of this building is transformed into the 
coordinate system of the search image. 
Figure 8: Example on image matching with two reference 
images and multiple orientations (Quickbird-subset 
of Phoenix area). Note that the computed correlation 
image is a multi-dimensional image. The number of 
dimension corresponds to the number of rotations 
(here: 120 with a rotation step of 3 degrees). 
It is often the case that buildings of the same (or similar) shape 
have different colours (grey values) in the images (e.g. due to 
different roof materials). Therefore it is advisable not to store 
image patches of the investigated buildings in the library, but 
instead, register their edges. In this case, also the search image 
has to be edge-extracted before applying the matching 
procedure. For gaining the edge information, classical operators 
like the Canny edge detector, Sobel operator, Laplacian of 
Gaussian etc. can be applied. 
Figure 8 shows an example of the image matching procedure. 
2.3 Texture Analysis 
One of the simplest ways for describing texture is to use 
statistical moments of grey level histograms of an image or a 
region. Measures of texture computed using only histograms 
suffer from the limitation that they carry no information 
regarding the relative position of the pixels with respect to each 
other. One way to bring this type of information into texture 
analysis process is to consider not only the distribution of 
intensities, but also the distribution of intensity variation 
(Gonzalez and Woods, 2002). 
For this kind of textural examination, firstly the so-called co 
occurrence matrix has to be derived for the examined area. This 
particular matrix holds e.g. information of pixel changes in 
multiple directions (usually horizontally, vertically and 
diagonally). The co-occurrence matrix’ extents are same in both 
directions and equal to the number of grey levels that will be 
considered. For example, for an 8 bit image (256 grey values) 
the co-occurrence matrix' extents would be 256 by 256. Usually 
a recoding is carried out to reduce the number of grey value 
classes (also called bins). A recoding of the original image 
down to 16 grey levels is for most of the cases satisfying (Gong 
et al., 1992). Nevertheless, during this research (on texture 
analysis), all images were recoded to 40 bins. 
Figure 9: Image and corresponding co-occurrence matrices in 
horizontal (left to right) and vertical (top to down) 
directions. 
At each position mc,r the co-occurrence matrix holds the 
number of changes from class r (=row indices) to class c 
(=column indices) (see Figure 9). This computation is carried 
out for multiple directions, meaning that one co-occurrence 
matrix is created for each direction. Figure 10 illustrates the 
creation of such matrices; here, four grey values exist and the 
co-occurrence matrices are derived for horizontal and vertical 
directions. 
The task now is to analyze a given co-occurrence matrix in 
order to categorize the region for which it was computed. 
Therefore descriptors are needed that characterize these 
matrices. Some of the most commonly used descriptors are