Sergey Zheltov
Due to knowledge of relative position of cameras, detection of the objects can be implemented as a matched filter.
Fig.6(a) shows the main idea of the method for two arbitrary orthophoto points. In each point (X, Y) a special mask Mxy
consisting of three regions (S,, S, S,) is generated. The regions have characteristic dimensions a, b, which can be
selected so that to ensure robust detection. The parameterb is selected from (7) when the minimum admissible height of
an object is given. If the object height is unknown then the parameter b can be selected by iterations. The angle ¢ is
determined by geometric constraints. Mask pixels in S. region have values equal to —1, pixels in S, regions have values
equal to 1, and all other values are 0. The result of the matched filter denoted by P(X, Y) is obtained by a convolution of
the orthophoto difference image with mask My:
P(X)Y) = Mxy * D(X,Y) (8)
Fig. 6(d,e) show two examples of matched filter applying.
i - " ue
; # { area
Orthopthoto area
Bee eeeeeeee usen unte un annuo deeft
Left i Right (d) (e)
camera camera (a)
Figure 6. Detection of 3D-objects by the matched filtering.
(a) Configuration of the masks for each orthopthoto points (on an example of two arbitrary points P,, P:).
(b),(c) Two examples of orthophoto difference images, (b) — one object, (c) — four objects.
(d), (e) Result of matched filtering for (b),(c)
6. RESULTS
This section shows the detection method application in more complex cases. As described in the previous section for the
primary detection of the object the matched filter is used. Object selection can be performed by the filtering result
P(X, Y) analysis. Different algorithms such as local maximum searching can be used for this purpose. For validation of
detected objects the criterion based on statistical decision theory was developed. It uses local features of the filter
output to decide about object presence. Then the information from the original images is used for final confirmation of
object presence.
Fig.7 shows the examples of the method applying. Detection results are shown on the left image of a stereopair. Each
detected object edge is shown as a rectangle. Below the rectangle two parameters of the object are shown. The
parameters are distance to the object in meters and object height in centimeters.
1046 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
ui
th
ir
al
th