Fusion in Signal Processing - Information Fusion in Signal and Image Processing - page 29

stored in the form of membership functions. At each iteration, the result is a list of
attributes, arranged in a particular order, and the supervisor in charge of allocating the
system's entire resources then provides a measurement corresponding to one of the
attributes. The performance evaluation requires a simulation with the introduction of
noise and a statistical validation over a large number of trials, whose quality criterion
is the number of requests necessary to obtain an object's class. The order used for
arranging these requests can be pre-defined, random or based on a designed mecha-
nism. Finally, the number of requests needed to obtain the right classification of the
object can bring improvements up to, for example, three out of eight total requests,
without affecting non-recognition and poor recognition rates. This overall orientation
of suggesting modular algorithmic subdivisions, whose output is no longer a single
decision but instead an ordered list (in this case the requests), is increasingly sought
after. This is because it naturally authorizes the resource allocation stage to operate in
a more serene way, thus allowing it to fully play its part, with a significant decrease in
conflicts over resources.
In the end, many research fields involve the problem of organizing a processing
chain generally consisting of modules with sequential subdivisions and other compet-
ing subdivisions. What is specific to the field of fusion is the fact that it is necessary to
evaluate in different places the amount of redundant or complementary information in
a sub-module's input in order to know how this particular data can be integrated into
the process. This systems architecture, where the local process control is performed
from an evaluation of the information content in the inputs, is no longer sufficient
today. It remains fundamental for dynamic process control, but it is accompanied by
process supervision performed by each sub-system.
It is considered today that, in the same way as sensors have their own operating
modes, algorithmic subdivisions also have different processing modes and that it is
possible either to switch locally, or under the control of a supervisor, from one mode
to another, or to locally make some of the processes compete with each other, then to
aggregate, combine or fuse all of the outputs in order to provide either an optimal out-
put or a summed-up output of the information acquired. We are no longer discussing
only fusion, but the more general field of artificial intelligence. In order to achieve
an operational status, all of the various subdivisions have to interact with each other
according to one or several plans, such as those suggested in [GAR 97] (Figure 2.6).
The most common simple plan works by using a supervision mechanism based on
competing processing modes, in a fusion architecture comprised of low-level signal
processing phases and of two types of streams: a continuous stream of data and a
stream for controlling processing modes on demand [CHE 97]. An accurate way to
describe reality is to take into account, on the processing mode level, the concept of
object behaviors by using different models expressing behavior variations in these
objects [REY 96].