Qdi.j) = [Ae (9 69^
where A is the amplitude of the dot, (x,y) is its
position, (i,j) are pixel indices, and the peculiar
brackets indicate integer truncation. A commonly used
estimate of the x-axis of the centroid is given by the
"center of mass” calculation:
2iQ(,j)
ZQ(,.j)
with the summation being over all i and j between
some values (-1 and +1, say). It is obvious that if the
object's position is constrained to a square (pixel)
such as Ixl<1/2, lyl<1/2, then the x-position estimate
can have only a finite number of values. In
particular, if 1<A<2 then these values are { 0, +1/2,
+1/3 }. Due to symmetry, the same is true of the
estimate of the y-axis of the centroid. The possible
combinations of (x,y) estimates are restricted to only
13 values, which are shown as dots in figure 2. The
regions delineated in the figure are the locales which
correspond to each of the estimate values. Note that
ANY position estimator will have no more than 13
possible values, (with (x,y) constrained to the unit
square, 1<A<2, and no noise present), since there are
only 13 possible digital representation for the object.
^
x =
FIGURE 2
LOCALES WITHIN A UNIT SQUARE AND
LOCATIONS OF THE CORRESPONDING
POSITION ESTIMATES
The presence of noise will complicate the situation.
Detailed analysis of locales in the presence of noise
is beyond the scope of this presentation, but it
should be realized that noise can be incorporated into
the basic theory in an approximate manner by
defining the number of effective quantization levels
(dynamic range) to be the number of digital levels
divided by the number of levels spanned by the
additive noise. The primary applications of the theory
are the estimation of quantization uncertainty and
optimal position estimation. Both will be seen to be
robust to noise. À more detailed discussion of locales
can be found in [10].
57
The concept of locales can be easily extended from
regions of object position to higher dimensional or
more abstract parameter spaces. For example, the
position of an edge (ignoring end-points or assuming
it is infinitely long) can be expressed in terms of its
slope and distance from the origin. These two
parameters can be used to construct a locale pattern
for a straight binary or grey level edge. It is then
possible to establish bounds on the position and
orientation of the line as well as an optimal estimate
of the two parameters. Further elaboration on this
example will not be presented but the reader is
invited (challenged) to construct the locale pattern
based on the discussion in the following section.
5. GENERATION OF LOCALE
PATTERNS
The locale patterns are generated from the contours of
the target, as explained in [11]. The Gaussian dot
discussed in the previous section can be represented
by contours with unit intervals which form
concentric circles. Using the center of a pixel as the
reference origin, the contours are drawn concentric to
the origin to represent the target at position (0,0).
Displaced versions of this contour pattern are then
overlaid on the original one to get the locale map.
The displaced versions are generated by moving the
contour pattern so that it is concentric with each of
the other pixels in the analysis window. For a 3 by 3
window, 9 copies of the basic contour pattern are
overlaid to get the final locale map. This is how
figure 2 was generated, except that the resulting
overlaid contours were truncated at the boundary of
the unit pixel.
In the case of higher dimensional position spaces
(three dimensions) the same procedure is used to
generate a multi-dimensional mesh of locale volume
elements. In the case of three dimensional position
(x,y,z) for the Gaussian dot, with z along the optical
axis of the imaging camera, the basic contour pattern
is a set of concentric cones. This pattern is replicated
by translation in (x,y), then the replicas are all
merged to form the locale pattern of volume
elements.
If the coordinates are parameters (such as orientation
or size) rather than object position then the
translation of the basic "contour" pattern when
constructing the replicas is based on the position of
the centers of the image pixels in the selected
parameter space. Detailed or formal discussion of the
more abstract representation of locales is beyond the
scope of this paper, but the generality of the concept
of locales should be noted.
The basic method of generating locale patterns is
very simple. It provides an easy method of
appreciating the distribution of quantization induced
position uncertainty for any target, no matter how
complex the target is and no matter how many gray