The 4 steps in the search process are:
1) In Image 1, feature selection by use of interest operator
This step deals with the question which point or feature should be
selected for measurement in image 1, while the other steps attempt to
find it in the image 2.
2) Using existing neighbourhood parallaxes, approximation to the
search area in image 2 using weighted interpolation
The objective here is to limit the search area in 3), but insure that
the correct point still lies inside this area. A good approximation
increases speed and reliability, since the chances of gross errors are
reduced. The search area is computed by assuming a parallax for the
current point, which is derived from parallaxes at points already
fixed in the neighbourhood. These existing points may come from the
current grid and / or the previous one. In the example given this
search area generally lay within +/- 7 pixels of the final value.
This depends on grid spacing variation of elevation.
3) In image 2, coarse matching with cross correlation
A standard cross correlation procedure with additional parabolic
interpolation, is used to position the window within the search area
at this stage. This step 3) is required in order to insure stable
convergence in the following step 4). As already explained, epipolar
geometry enables the search to be made in one direction only.
Optimized techniques are then possible which more than double the
search speed. Generally, positioning at this stage is better than 1
pixel (often 0.1 - 0.2 pixel).
4) In image 2, fine matching using least squares adjustment
Final determination of position and slope at the chosen feature is
carried out by least squares adjustment /1/. The following sum is
minimized:
(gtx) - g28,0 )* ,
where g1, g2 are grey levels in photo 1 and 2 respectivly. g2 is
determined according to the following functional model:
HO + H1 ( g1(u,y) )
AO + A1 x + A2 y
g2(x,y)
u
where: HO, H1
A0, M, A2
unknowns of radiometric transformation
unknowns of geometric transformation
The geometric transformation is assumed to be affin in the x-direction
only (no variation in y due to the epipolar condition), i.e. the
surface is assumed to be flat around the feature selected. Iteration
1s necessary because the system is non-linear (linearization by Taylor
expansion).
-7 566 -