it
m
nd
ng
re
te
ch
nt
y
n
IS
id
er et À
AM o D —"- AJ)
fine changes in intensity, filters of different sizes are used. Five different
filters have been detected in the human visual system with filter widths
of 4, 9, 17, 35, 63 pixels.
Realizing the problems with area correlation, methods Using
characteristic points have been proposed and implemented (/1/, /8/).
Characteristic points represent a high gray value variance in all
directions, and they can be determined with interest operators. Due to the
high degree of nonlinearity in the mathematical model with a small radius
of convergence, the approximations must be very good, on the order of |
pixel size. This raises the question on how the human visual system copes
with the problem of approximations. We are obviously capable of setting
the fixation point at an average disparity level of that part of the the
object at which we are looking. The depth information obtained from the
coarsest channel is used in the next finer channel to prevent the matching
‘of zero crossings that are not conjugent points since this would lead to a
point in the object space that is not on a surface (false target problem).
Hence in a multi- channel implementation no approximations are needed.
Stereo Matching with Zero Crossings
In the original implementation of the model of the human visual system as
developed by Marr and Poggio, Grimson determinded the zero crossings in
the filtered image along epipolar lines (see, eg,/3/) This requires
resampling the original image such that the horizontal scanlines are in
fact epipolar lines. The matching process then proceeds in an iterative
manner, from coarse to fine channels. The search space for corresponding
zero crossings is reduced in this approach from the entire image to one
scan line and within the scan line to the range of twice the channel width,
We, under consideration. With an average disparity, do. obtained from the
computations from the coarser channel, the zero crossings are found
within the interval [x + d, : Wes yl. In case several zero crossings are
found within this interval, a disambiguation procedure considering surface
continuity determines the most probable match. Recent modifications of
the original implementation extend the search space to include several
scan lines (/2/). This compensates for local distortions caused Dy the
camera.
-605.-