Conclusion - Information Fusion in Signal and Image Processing - page 248

belief functions and fuzziness. We thus know now which application frameworks are
right for these theories, their advantages and their limits in these application fields;
we also know how to represent and model information and numerical, symbolic or
structural data in each of the formalisms, and how to achieve their combination. There
have been many new developments, particularly for the traditional multi-source clas-
sification, structure or object recognition in images, tracking, localizing and planning
applications.
12.2. A few prospects
Despite this progress, some aspects remain poorly understood or require further
development.
Taking into account the origin of the data and knowledge, as well as the relations
between sources, is still often performed under supervision and therefore requires lit-
tle experience. One of the important questions involves the independence between
sources and conditioning (particularly in the case of sequential fusion, in dynamic
updating processes). The probabilistic framework offers methods to test statistical
independence, and those are usually the only tools at our disposal. But this type of
independence is often considered too restricting, and in other contexts, such as belief
function theory, the preferred concept is cognitive independence, which is related
to how the knowledge and data are acquired rather than to their nature [SHA 76,
SME 90]. When choosing the operators of fuzzy and possibilistic theories, indepen-
dence results in the operators being idempotent, whereas dependence requires rein-
forcement.
A very difficult question is conflict management. Insofar as it is possible, the
sources of the conflict have to be identified and made explicit, in order to avoid incon-
sistencies when making the decision. In particular, it is not always easy to tell the
difference between conflict and the complementarity of sources, or to know whether it
should be resolved or not. Conflict, which can be referred to as “apparent”, is actually
a form of complementarity. For example, if a source systematically includes class B
in class A whereas another source is good at distinguishing them, these two sources
seem to be in conflict. Recognizing complementarity often requires the use of prior
(or learned) knowledge regarding the low possibilities for the first source to tell the
two classes apart. Resolving such conflicts is easy once they have been properly iden-
tified. A second form of conflict, which is real this time, is due to inconsistencies
between sources, which are caused by their limited reliabilities, by changes that occur
in the scene between acquisitions, or also by the fact that they are not dealing with
the same thing. This type of conflict is more difficult to identify and to resolve. It is
sometimes even preferable to simply indicate it and not to try to solve it because it
often corresponds to a fundamental inadequacy of our knowledge of the problem.