Page 25 - Classification Parameter Estimation & State Estimation An Engg Approach Using MATLAB

14 DETECTION AND CLASSIFICATION

Table 2.1 Some application fields of pattern classification

Application field Possible measurements Possible classes

Object classification
Sorting electronic Shape, colour ‘resistor’, ‘capacitor’,
parts ‘transistor’, ‘IC’
Sorting mechanical Shape ‘ring’, ‘nut’, ‘bolt’
parts
Reading characters Shape ‘A’, ‘B’, ‘C’,
Mode estimation in a physical process
Classifying Tracked point features ‘straight on’, ‘turning’
manoeuvres of a in an image sequence
vehicle
Fault diagnosis in a Cylinder pressures, ‘normal operation’, ‘defect
combustion engine temperature, vibrations, fuel injector’, ‘defect air
acoustic emissions, crank inlet valve’, ‘leaking
angle resolver, exhaust valve’,
Event detection
Burglar alarm Infrared ‘alarm’, ‘no alarm’
Food inspection Shape, colour, temperature, ‘OK’, ‘NOT OK’
mass, volume

the object is qualified according to the values of some attributes of the
object, e.g. its size, shape and colour.
The sensory system measures some physical properties of the object
that, hopefully, are relevant for classification. This chapter is confined
to the simple case where the measurements are static, i.e. time inde-
pendent. Furthermore, we assume that for each object the number of
measurements is fixed. Hence, per object the outcomes of the measure-
ments can be stacked to form a single vector, the so-called measurement
vector. The dimension of the vector equals the number of meas-
urements. The union of all possible values of the measurement vector
is the measurement space. For some authors the word ‘feature’ is very
close to ‘measurement’, but we will reserve that word for later use in
Chapter 6.
The sensory system must be designed so that the measurement vector
conveys the information needed to classify all objects correctly. If this is
the case, the measurement vectors from all objects behave according to
some pattern. Ideally, the physical properties are chosen such that all
objects from one class form a cluster in the measurement space without
overlapping the clusters formed by other classes.