







Advanced Object Tracking and Data Fusion: Methods and Applications



Session organizer: Wolfgang Koch
Tutorial Abstract:
In many engineering applications, including surveillance, guidance, or
navigation, single standalone sensors or sensor networks are used for
collecting information on time varying quantities of interest, such as
kinematical characteristics of moving or stationary objects of interest
(e.g. manoeuvring air targets or ground moving vehicles). More strictly
speaking, the state vectors of stochastically moving objects are to be
estimated from time series of sensor data sets, also called scans or data
frames. The individual measurements are produced at discrete instants of
time, being referred to as scan or frame time, target revisit time, or data
innovation time. In case of moving pointsource objects or small extended
objects often relatively simple statistical models can be derived from basic
physical laws describing their temporal behaviour and thus defining the
underlying dynamical system. In addition, appropriate sensor models are
available or can be constructed, which characterize the statistical
properties of the produced sensor data sufficiently correct.
Illustrated by examples from real applications, methods for dealing with
selected aspects of sensor data fusion will be discussed: At first, `fusion
of data produced at different times', i.e. the tracking problem, will be
considered. The use of multiple sensors is illustrated for improving
detection and track accuracy (`fusion of data produced by multiple
sensors'). This is followed by a discussion of methods for integrating
nonsensor context information (`fusion of data with refined sensor models',
`fusion of data with road maps'. The overview is concluded by a sensor
management example. A written manuscript will be made available in
electronic form.



Fusion of Quantitative and Qualitative Information with NeuroFuzzy Systems and Bayesian Networks



Session organizer: Rudolf Kruse
Tutorial Abstract:
In many engineering applications, empirical data and expert knowledge are
used simultaneously. Often there is a need to fuse these different types of
information. In many cases a common framework for the representation of
quantitative and qualitative knowledge is needed, because methods for
learning, adaptation, updating, revision, etc. have to be integrated. In
this tutorial we consider two fusion frameworks in more details:
Probabilistic graphical models and neurofuzzy systems.
In the last decade graphical models have become one of the most popular
tools to structure uncertain knowledge about highdimensional domains in
order to make reasoning in such domains feasible. Their most prominent
representatives are Bayesian networks and Markov networks, but also
relational and possibilistic networks turned out to be useful in practical
applications. For all types of networks several clear, correct, and
efficient propagation methods have been developed, with join tree
propagation and bucket elimination being among the most widely known. In
practice, however, the need also arises to support a variety of additional
knowledgebased operations on graphical models. In this turorial the fusion
of networks with relational rule systems, and operations such as revision,
updating, network approximation, and learning from data samples are studied.
As a second framework for the fusion of qualitative and quantitative
information, neurofuzzy systems are very popular. Nowadays fuzzy systems
are frequently applied in data analysis problems, because they can provide a
simple, inexpensive and interpretable modelling of a priori knowledge in
terms of IFTHENrules. In order to take advantage of the knowledge hidden
in data sets, it must be possible to improve the priori knowledge in the
light of this additional data material. The so called learning and
adaptation in fuzzy systems is most often implemented by learning techniques
derived from neural networks. In this tutorial we study the methods used in
our neurofuzzy system NEFCLASS in some more detail.
The research to be reported about in this tutorial was mainly triggered by
consulting of the automobile manufacturer DaimlerChrysler and Volkswagen
Group, where graphical models are now established for several tasks, and by
projects with British Telecom, BMW, Siemens and Deutsche Giro und
Sparkassenverband, where fuzzy systems and neural networkshave been used in
the context of exploratory data analysis.



A New Class of Computation for Distributed Sensor Networks



Session organizer: S. S. Iyengar
Tutorial Abstract:
“The Next Generation Distributed Sensor Network Simulator.” This talk provides an overview of the development of theory and implementation of real
time sensor simulators for unstructured applications. The talk also discusses a comparison of this simulator with the other existing simulator
currently available in the literature.



Multisource Multisensor Information Fusion: Architectures, Algorithms, and Applications



Session organizer: Belur Dasarathy
Tutorial Abstract:
This tutorial will offer an overview of multisensor, and
multisource information fusion field from the author's perspectives along
three fronts: architectures, algorithms, and applications. This is
accomplished through the process of addressing the questions of what, why,
when, and how as it relates to information fusion. The presentation starts
with an introduction to the terminology, motivation for the study of
information fusion, followed by a delineation of the various taxonomies
discernible in the field. This will be followed by a brief discussion of
some basic architecture design options. A sampling of algorithms conceived
for fusion in different modes will be presented next. The tutorial will
emphasize fusion at feature and decision levels since this is less directly
driven by sensor and application specific considerations and hence amenable
to a discussion that is independent of any specific application. The
presentation will close with a panoramic view of the application domains.



An Introduction to Dirichlet Processes



Tutorial Organizer:
Volker Tresp
Tutorial Abstract:
Bayesian modeling is a principled approach to updating the degree of belief
in a hypothesis given prior knowledge and given available evidence. Both
prior knowledge and evidence are combined using Bayes' rule to obtain the a
posteriori hypothesis. In most cases of interest to machine learning, the
prior knowledge is formulated as a prior distribution over parameters and
the evidence corresponds to the observed data. By applying Bayes' formula we
can perform inference about new data. In hierarchical Bayesian modeling, one
is not primarily interested in learning the true parameters, but in learning
the true distribution of parameters. Consider the situation of learning a
model for predicting the outcome for patients with a particular disease
based on patient information. Due to differences in patient mix and hospital
characteristics such as staff experiences the models are different for
different hospitals but also will share some common effects. This can be
modeled by assuming that the model parameters originate from a particular
distribution of parameters that can be learned from data from a sufficiently
large number of hospitals. If applied to a new hospital, this learned
distribution assumes the role of a learned prior. Now a technical problem
arises: the parameterization of this learned prior distribution needs to be
sufficiently flexible to truthfully represent the actual distribution. Thus
instead of selecting a prior distribution out off a class of standard
distributions, it is more appropriate to assume a nonparametric
representation. The most common way of generating a nonparametric
distribution is by means of a Dirichlet process. The Dirichlet process can
be thought of as a generalization of a Dirichlet distribution if we let the
number of states go to infinity.
In my tutorial I will give an introduction to Dirichlet processes and
present various applications. I hope that this tutorial might inspire the
participants to think of novel applications of nonparametric Bayesian
modeling to multisensor fusion and to the integration for intelligent
systems.


Image Fusion



Session organizer: Michael Heizmann
Tutorial Abstract:
For many tasks of automated visual inspection and environmental perception, acquisition
and evaluation of a single image is not sufficient. The reason is that the
information of interest cannot be acquired from a single image at all or with the
necessary quality. This insufficiency can result from sensor characteristics such as
sensor noise. In addition, the information of interest may be unavailable by direct
observation. In such cases, the desired information can often be obtained by recording
multiple images (socalled image series), which are then combined in an
image fusion in order to achieve a new or improved representation of the image
content or the scene properties.
This tutorial gives an introduction to common fusion approaches for image
data and illustrates main aspects of image fusion. Image sensors are characterized
with respect to the fusability of the acquired data. Abstraction levels on which an
image fusion can be performed are explained. To describe processing stages for
image fusion in general, a universal fusion scheme is introduced. The mentioned
aspects of image fusion are illustrated by means of practical examples.



