Page 150 - Dynamic Vision for Perception and Control of Motion

134 5 Extraction of Visual Features

After specification of search
range, mask width, and orienta-
tion to be used, the first computa-
tional step is to sum up all the
pixel values over the mask width
185.6
n w and, thus, collapse the width to
5.6
a single vector component. This
vector spans over the search range
(named PathLen in CRONOS, see
Figure 5.9).
It represents the average inten-
sity values in the direction of
Figure 5.8. Definition of edge orientation as
mask orientation; this corresponds
used in CRONOS: Starting from the horizontal
to low-pass filtering in this mask
direction to the right, angular increments are
counted clockwise direction. With more than 16-bit
processors and 8-bit intensity val-
ues for each pixel, there is no
need to divide by the number of pixels summed, thus saving computing time. (If
intensity values close to the original ones are preferred, shift operations may be
used.) In the software packages in use at UniBwM, the options for mask widths
i
are n w = 2 + 1, with preference for i = 2 to 4; this means that the smallest mask
width with only three pixels is not used.
This odd value of n w has been chosen initially to have a symmetrical distribution
of the image stripe represented around the center of the nominal pixel position,
which is convenient if no
y(i) Mask element
subpixel resolution is used.
Using subpixel resolution,
i
defining n w = 2 is the z(j)
cleaner solution for work- Vector representing low-pass
filtered values along edge
ing on different scales. n w
It is seen from Figure
5.9 that part of the search
path length is lost at the
boundaries for oblique Figure 5.9. Low-pass (high spatial frequency) filtering
mask orientations; this has orthogonal to the expected edge direction reduces the
to be taken into account
search stripe to a vector, independent of mask width n w
when specifying the search for efficient computation of correlation values
range.

5.2.1.2 Computation of Ternary Correlation Values

The vector obtained in the previous section is the basis for edge localization by ter-
nary correlation. By subtracting two consecutive vector components from each
other, gradient information in the search direction for the given angular orientation
of the mask is obtained (see Figure 5.10a, upper left). At the point where this dif-