m CE
f - focal length of the camera,
o,= a priori value of the image-coordinate
precision,
(oy, Oy, 07) are the accuracies of object-point
coordinates (X,Y,Z), and o, is the vector of
positional accuracy.
We can write eq (3) as:
gr (PEF) Zo, (4)
where S (=D/f) is the scale number, and
PEF is the positional error factor
The equations are functions of camera
parameters (f,0,) and the configuration
parameters (¢,n,D). Therefore they can be used
to:
(a) predict the accuracies which can be
expected from a given camera and
configuration, and also
(b) determine the camera and
configuration parameters which
will be required to meet a given
accuracy specification.
We use the equations within an experimental
expert system to select the camera and initial
imaging geometry that are suitable for a task.
4. THE EXPERIMENTAL EXPERT SYSTEM
4.1 Introduction
An experimental expert system has been
developed. The system uses an Expert System
shell (Expertech Xi Plus ver 3.5c2) which runs
on standard IBM PC XT and requires only 512KB
RAM. The expert system was designed in
modules. It has five integrated knowledge-
bases (modules) which interact with each other
and help advise on target design, camera
selection, imaging geometry, and data
acquisition schemes.
The system works by using the accuracy
predictors to modify design parameters until
the accuracy specified by the user is met.
In order to understand how the system works,
we shall first show how it handles the problem
of selecting camera and initial imaging
geometry; and then show an actual example of
a consultation session.
4.2 Selection of camera and initial
imaging geometry
The selection of camera and the initial
approximation of imaging geometry is a
knowledge-intensive issue. We constructed a
database which contains details of many
cameras; and lists such attributes as camera
type, focal length, format size, minimum
focusing range, as well as image coordinate
measurement precision relating to the cameras.
A second database is constructed which gives
the calculated PEF for a range of n and $. In
our system, these databases are stored in
files called cameras.dbf and optconfi.dbf (see
Fig 3). These are held independently of the
expert system. The set of data within each
database is arranged in a defined way in order
to speed retrieval and enable interrogation
within the expert system. The user supplies
information concerning the largest dimension
of the object to be measured and the accuracy
456
required. The system works by using formula 3, the
two database files, and some user-supplied
information (Fig. 3). For example, the system uses
cameras.dbf and the largest dimension of the object
to determine D and hence S. The steps involved are
depicted by the flow chart (Fig. 4), which is a
simplified version of the general procedure of
'camera selection'.
4.3 Target Design
The design of a target needs to be not only in
terms of physical characteristics (ie shape, size)
but also in terms of optical characteristics. The
determination of the size of a target that would be
appropriate for a particular measurement task is
simple. The other characteristics are more
difficult. They are determined by finding answers
to a number of questions, eg is it necessary to
provide artificial targets ?. If so should it be
contact (ie physical) or non-contact (ie optical)?
if contact, should it be a planar or non-planar ?
should the planar (or non-planar) be diffuse or
retro-reflective ? etc. The decision-making process
requires expertise, without which the desired
accuracy may not be achieved. An example is a case
(Kenefick, 1971) where the use of diffuse targets
could have yielded an accuracy that is 250% better
than that achieved with reflective targets. It is,
therefore, imperative to use the type of target
that is suitable for each measurement task.
Existing design packages leave this decision
entirely to their users.
A classification scheme for target type has been
devised (Fig. 5). Most commonly encountered targets
can be categorised into one of the types in this
classification scheme. This scheme is reasonably
well-defined, and its hierarchical structure is
compatible with the problem-solving techniques in
expert system technology based on traversing trees.
Hence target design makes a particularly good
domain for processing with expert system
technology. The classification scheme has been
converted into a decision tree/table, which in turn
is converted into IF...THEN structured rules. For
each of the target types, the series of conditions
under which it is the most suitable type is
constructed, for example:
IF image of object is required to be
invisible on the photo
THEN target type required is an active
light-reflecting(eg retro-
-reflective)
5. APPLICATION OF THE EXPERT SYSTEM: AN
EXAMPLE
A typical consultation session relating to the
selection of camera and initial imaging geometry is
shown in Fig. 6. In this session a sample problem
and the system-to-user interactions are shown. We
note that the characteristics of a camera will
determine its suitability or otherwise for a given
task. These characteristics are contained in the
database file (cameras.dbf) to which the system has
an automatic access.
When the system asks a question, the user may want
to know why such a question is being asked. This,
the user does by selecting a special function key
(«F3» in our case). The system then responds by
reporting the line of reasoning that led to that
particular question being asked. To enable the
system to evaluate a camera that is not contained
in cameras.dbf, the system provides the user with
a form (Fig.7) to fill. The information provided by
the user is used not only for the purpose of the
current consultation session, but also to update
