The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
984
2. PRESENTING A NOVEL APPROACH TO DEVELOP
LSM IN ORDER TO BUILDING CHANGE
DETECTION
2.1 Object decomposition
For the purpose of our approach, a building is expressed
through a wireframe representation based on prior vector
information (Habib, and Kelley, 2001). For decomposition of
the lines, a new element is introduced, the Minimum Spatial
Element (MSE). The MSE describes the resolution of spatial
change that the user is interested in. Using this information, we
perform a segmentation of outlines, and lines are essentially
substituted the corresponding points along the outline (Figure.
1).
Figure 1: object decomposition
f(x,y)-g(x,y)=e(x,y) (2)
With e(x,y) being the error vector.
In a typical least squares matching method, observation
equations can be formed relating the gray values of
corresponding pixels, and they are linearized as(Grien, and
Akca, 2005):
fix, y) - e(x, y) = g° (x, y) +
dg°jx,y)
dx
dx +
dy
(3)
The derivatives of the image function in this equation express
the rate of change of gray values along the x and y directions,
evaluated at the pixels of the patch. Depending on the type of
edge, the geometric relationship describing the two windows
may be as complex as an affine transformation, or as simple as a
simple shift and/or rotation. To facilitate solutions we can
resample template and/or actual image to have edges lying
along one of the major coordinate axes. For example, we can
have resampled edges oriented along the y axis of the
corresponding windows. In this case, the above equation may be
reduced to:
f(x, y) - e(x, y) = g° (x, y) + dg ^ dx ( 4 )
ox
Where the two patches are geometrically related through a shift
along the x direction:
2.2 Providing template
In order to provide template, first the image coordinate of the
point along the building edges is computed for each image by
colinearity equation. The image coordinates is considered as the
centre of the template. Determining the size of the templates is
the second step. This allows us to optimize computational
performance without compromising accuracy potential. From an
accuracy and reliability point of view, small image patches, e.g.
3x3 pel 2 are not suitable for LSM due to small redundancy
number. An optimal size of matched window also differs with a
pixel size and scale of the original images. A size of 15 x 15
pel 2 for good quality images with a lot of details and at least
21x21 pel 2 for noisy images or when significant differences
in brightness between a template and a search area occur are
recommended in (Kraus, 1997). So, the appropriate template
size for our images used in this research is 21 x 21 pel 2 .
2.3 Development of least square template matching
A window depicting an edge pattern is introduced as a reference
template that is subsequently matched to digital image patches
in the vicinity of actual edge segments. The concept behind the
method is simple yet effective: by matching the edge template
window to an image window, we can identify edge locations in
the image as conjugate to the a priori known template edge
positions. Assuming f(x,y) to be the reference edge template and
g(x,y) to be the actual image patch, a matching correspondence
is established between them when
f(x,y)=g(x,y) (1)
However, considering the effects of noise in the actual image,
the above equation becomes
XNEW ~ XOLD + dX (5)
The dX parameter is the unknown that allows the repositioning
of the image window to a location that displays better
radiometric resemblance to the reference template. Regardless
of the choice of geometric transformation, the resulting
observation equations are grouped in matrix form as
-e = Ax - l; P (6)
Where, / is the observation vector, containing gray value
differences of conjugate pixels. The vector of unknowns x
comprises the shift at the x direction, while A is the
corresponding design matrix containing the derivatives of the
observation equations with respect to the parameters, and P is
the weight matrix. A least squares solution allows the
determination of the unknown parameters as
x = (A T PAy l A T Pl (7)
Based on the formulation of the above model, we have to face
the major challenge about formulation of the weight matrix
described in the next section.
In order to better handle noise through the analysis of local edge
information within the matching windows, we have to allow
edge pixels to influence the solution more than the rest of the
template. This can be performed by manipulating the
corresponding weight matrix P. Indeed, by assigning high
weight values to certain pixels, we allow them to influence the
solution more than pixels with low weight values and the effects
of undesired variations on our solution are kept at minimum.