Fusion in Signal Processing - Information Fusion in Signal and Image Processing - page 28

Thus, we have the following questions:
- how long do we have to wait for the missing data?
- when they reach the processing center, should they be integrated to the fusion
process?
- how does this data contribute to the quality tracking operation?
- how does the quantity of information change as it grows older?
As with missing data of contextual origin, the method can rely on a decision tree,
whose final action is to choose between gaining information by using the delayed data,
or gaining time by discarding this same data. There are necessarily several criteria
to this choice because it involves the content of the data delayed by evaluating the
expected information gain for each track, but the complexity of the resulting situation
without this data plays a role and requires evaluating the risk of confusing tracks if
these tracks are close, as well as evaluating the risk of losing the track.
When a multi-sensor system is operating, for scene surveillance, often the quality
of the result depends mostly on time. In a multi-target environment, recognizing and
identifying the objects in the scene has to be done in as short a time as possible, with-
out waiting for all of the data that can be provided by the sensors. Because accessing
the data takes time, one of the solutions is to operate in the pull mode, or data request
mode, and to choose the smallest (the most discriminating) set of attributes for differ-
entiating objects from one another.
For airspace surveillance, there are many types of models and data available for
classification and identification. Thus, physical attributes that do not change, such as
the wingspan or the length of a plane, can sometimes help in directly recognizing
the target, or at least in obtaining a measurement and therefore formulate hypotheses
on the target's class, which will later be confirmed or proven wrong based on other
attribute measurements. Furthermore, many types of sensors can provide a wide vari-
ety of information, depending on what mode is selected and thus be used to measure
different attributes. The variety of a priori information that can be obtained about a tar-
get when trying to classify it can be used to think in advance. This way, it is possible to
recognize an object faster by taking into consideration the discriminatory capabilities
of the different attributes conditionally to each class.
If, for each attribute, we have at our disposal the membership function for each
class, the choice of the quickest attribute for characterizing an object is the one that
leads to the result with as little ambiguity as possible. We then have to define the
membership functions that represent the different classes. The degree of separation
between two membership functions [CON 01, CON 02] can then be used to imple-
ment a mechanism for making it easier to choose the first attributes to search for (the
selection of the sensor and of its mode), corresponding to each attribute. The a pri-
ori information is obtained either from learning or by constructing a database and is