generally does not respect the number of pixels per
line transmitted by the camera. The frame.grab
simply resamples the line by interpolation. These
problems have been discussed quite thoroughly in
assessments of the use of CCD imagers for
photogrammetric purposes, as in [15,3,2].
There are two basic approaches to overcoming the
loss of geometric fidelity imposed by commercial
digital image acquisition systems; signal analysis and
specialized electronic circuitry. It has been shown for
example, that by clever Fourier analysis of the image
signal, the horizontal jitter introduced by the
digitizer's resampling can be determined, hence
corrected for [14]. Small targets can thus be
positioned to an accuracy of about 1/60th of a pixel.
With special electronics, one might expect to do
better still.
3. SUB-PIXEL IMAGE PROCESSING
Empirical investigations into attainable sub-pixel
position estimation in digitally reconstructed
imagery, such as [18], have indicated that sub-pixel
measurement is realizable, but the level of
performance which can be attained is still a debatable
issue.
A difficulty with empirical results is that they are
only strictly true for the particular system
configuration used for the experiment and they are
rarely assured of being optimal in any sense.
Theoretical investigations on the other hand, tend to
be idealized and optimistic. Neither empirical nor
theoretical investigations have given a good
framework of knowledge about sub-pixel position
estimation which can be assimilated by the
practitioners of the art. While there have been some
important advances in terms of theory and practice,
([5,4,12,17], to mention a few) there is a lack of
cohesion or common basis on which they can be
purviewed. The "locales theory" introduced in [11]
may provide some of the required common ground for
the important problem of sub-pixel position
estimation. The locales theory introduces a bound for
geometric precision against which other analysis and
methodologies may be collectively compared.
It is worth emphasizing perhaps, that the issue of
sub-pixel position estimation can generally be
isolated to just two components of the overall image
processing task, namely the acquisition of the digital
image and the sub-pixel position estimation
algorithm itself. There are many other steps, as
figure 1 indicates, which do not directly involve sub-
pixel considerations. The position of a target may in
fact be estimated twice, once in the course of
detection or recognition, wherein an approximate
pixel location is determined, and again during precise
sub-pixel position estimation based on the raw gray
scale image data. In this context, it is important that
sub-pixel position estimation be clearly distinguished
from the tasks of detection and recognition. It will be
assumed throughout this paper that target position is
known to approximately one pixel; it is the task of
sub-pixel position estimation to improve upon the
rough estimate.
IMAGE ACQUISITION
STORAGE
ENHANCEMENT
SEGMENTATION
EXTRACTION
NN ELITE,
tot tata Eee
OBJECT COORDINATES
FIGURE 1.
IMAGE PROCESSING STEPS FOR
POSITION EXTIMATION
(AFTER EL-HAKIM [9])
4. LOCALES
This section will begin with a brief discussion of the
history and terminology of locales. The concept of
"locales" was developed in [11] for arbitrary targets
and at about the same time for the more restricted
case of binary line segments encoded by chain codes
[6]. The idea is simple and useful: a locale is a region
within which the object (target) may be moved
without causing any change to its digital
representation. The term "domain" used by Dorst and
Smeulders [6] refers to a region in a transformation
space of object position, but the principle is the
same as for locales. The term "feasibility region", as
adopted by Berenstein et. al. [1] for the object-
position equivalent of the "domain", is the same as
locale. Feasibility regions will be introduced in
section 8 in the context of locale construction,
whereby the former are intersected to generate the
latter.
Interest in locales arises from the fact that the locale
size determines the position uncertainty due to
quantization and the locale center is the optimal
position estimate in terms of minimizing
quantization errors.
Just what is a "locale"? Consider the following
example. A small dot might appear in a digital image
as a sampled and quantized Gaussian function Q(i,j),
56