Final Long
Program
Monday - Tuesday - Wednesday - Thursday
Thursday July 13
Session ThA1: Plenary
talk
Unifying Estimation and Decision
Fusion
X. Rong Li, University of New Orleans, LA, USA
Slides : 520 Ko pdf file
Session ThB1: Defense
Systems
Chair: Stefan Arnborg, Royal Institute of Technology,
Stockholm, Sweden
Co-Chair: John Kent, Hi-Q Systems Limited, Winchester,
UK
1 - A Platform for Interoperable Fusion
Models
Jean-François Challine, Thomson - TCC/SIC/SEA Colombes,
France
Véronique Royer, ONERA, Châtillon, France
Abstract: The paper presents a generic and extensible
prototype platform for Intelligence fusion. The fusion process is designed as
an heterogeneous “algebra” of fusion operations. Each fusion
operation works on given abstract data types. They are generic being
independent of the semantics of data. Data and knowledge semantics is given
explicitly by an input ontology together with descriptive models of the
ontology concepts. The models are automatically encoded into representations
suitable for the fusion operations. The approach allows to integrate external
fusion algorithms, as well as to change the ontology. Interoperability is
obtained through the common ontology : the consistent combination of fusion
operations is indeed made possible because of a common semantics of data and
common reference concepts, not because of common data and knowledge.
2 - Deploying Tactical Fusion Systems:
the Challenges
John Kent, Hi-Q Systems Limited, Winchester, UK
Abstract: Although fusion tools have successfully been
applied to a wide variety of problem domains, fusion research has failed to
produce systems that can support the needs of the commanders of land-based
military operations, especially those that involve operations other than war.
This paper explores the reasons for this failure to date, summarising the
lessons of recent operations and relevant NATO research, and shows that even
the apparently simple task of training fusion system users can impose
significant technical challenges. The paper concludes by suggesting an agenda
for further research based on the development of man-in-the-loop systems
consisting of appropriate decision support tools embedded in a wider
application framework.
3 - Intelligent Threat Assessment
Processor (ITAP) using Genetic Algorithms and Fuzzy Logic
Paul Gonsalves, Rachel Cunningham, Nick Ton, Dennis Okon,
Charles River Analytics, Inc., Cambridge, USA
Abstract: The explosive growth in the area of information
technology provides a tremendous opportunity for enhancing military warfighting
capabilities. The management and processing of military intelligence
information, the requisite assessment of enemy capabilities, intent, and
objectives, and the generation of appropriate response recommendations form a
critical element of battlespace operations. Here, we develop an Intelligent
Threat Assessment Processor (ITAP) for enhancing tactical threat assessment.
Our novel system integrates a genetic algorithm approach to predicting enemy
courses of action (eCOAs), a fuzzy logic-based analysis of predicted eCOAs to
infer enemy intent and objectives, and in conjunction with our on-going
development of an Intelligent Fusion and Asset Management Processor (IFAMP),
provides the necessary functionality to support multi-level data fusion. We see
considerable potential for this approach in enhancing existing tactical
decision-aiding systems and addressing future information dominated battlespace
requirements.
4 - Information Awareness in Command
and Control: Precision, Quality, Utility
Stefan Arnborg, Joel Brynielsson, Royal Institute of Technology,
Stockholm, Sweden
Henrik Artman, Swedish Defence College, Stockholm,
Sweden
Klas Wallenius, CelsiusTech Systems AB, Järfälla,
Sweden
Abstract: In current Command and Control system design,
the concept of information plays a central role. In order to find architectures
for situation and threat databases making full use of all dimensions of
information, the concept of information awareness must be understood. We
consider and define some information attributes: measures of precision, quality
and usability, and suggest some uses of these concepts. The analysis is
Bayesian. A critical point is where subjective Bayesian probabilities of
decision makers meet the objective sensor-related Bayesian assessments of the
system. This interface must be designed to avoid credibility problems.
Session ThB2: Target
Tracking - 6
Chair: Emil Semerdjiev, Bulgarian Academy of Sciences,
Bulgaria
Co-Chair: Kuo-Chu Chang, George Mason University, Fairfax, VA, USA
1 - A Comparison of Fixed Gain IMM
against two other Filters
Eric Derbez, Bruno Remillard, Université du
Québec, Trois-Rivières, Québec, Canada
Alexandre
Jouan, Lockheed Martin Canada, Montréal, Québec, Canada
Abstract: The purpose of this paper is to present an
alternative to the constant acceleration Kalman filter which requires half the
computational load and yet performs almost as well as the IMM filter. The
theoretical justication for this filter came from a study of the IMM filter by
two of the authors. The results of this study are recalled, and illustrative
simulations using these filters are carried out by transforming noisy radar
data into Cartesian coordinates and then applying a filter to each coordinate
separately. The proposed filter is analysed against a constant velocity,
constant acceleration IMM filter, and a constant acceleration regular Kalman
filter. The stability properties of each of these filters are also
addressed.
2 - Progressive Correction for
Regularized Particle Filters
Nadia Oudjane, Christian Musso, ONERA, Châtillon,
France
Abstract: Particle methods have been recently proposed to
deal with the nonlinear filtering problem. These are Monte Carlo methods that
can provide a nonparametric approximation to the signal conditional
distribution even in nonlinear and non Gaussian cases, without depending on the
state space dimension. In this article, we present a new version of regularized
particle filter using a progressive correction (PC) principle which improves
the approximation, in introducing a decreasing sequence of (fictitious)
variance matrices for the observation noise. This method is applied to the
multisensor tracking problem (radar and IR sensor) and compared to the
classical regularized particle filter and the EKF.
3 - Optimal Single Sensor MHT
Improvements
Huub W. de Waard, Hollandse Signaalapparaten B.V., The
Netherlands
Abstract: The primary contribution of this paper is the
definition of a marginal bearing density function for the measurements produced
by a certain track which forms the theoretical foundation for the different
proposed promising improvements. Stepwise implementation of the different
measures can provide MHT applications with the means to use the available
computer storage and computation time resources as eÆcient as possible.
Furthermore, delay-times that can occur before processing/pruning are
minimized.
Session ThB3: Security
and Surveillance - 1
Chair: Robert H. Bishop, University of Texas, Austin, TX, USA
Co-Chair: Jean-Pierre Le Cadre, IRISA/CNRS, Rennes, France
1 - Decision Fusion using Support
Vector Machines (SVM)
Bert Gutschoven, Patrick Verlinde, Royal Military Academy,
Brussels, Belgium
Abstract: The contribution of this paper is twofold: (1)
to formulate a decision fusion problem encountered in the design of a
multi-modal identity verification system as a particular classification
problem, (2) to propose to solve this problem by a Support Vector Machine
(SVM). The multi-modal identity verification system under consideration is
built of d modalities in parallel, each one delivering as output a scalar
number, called score, stating how well the claimed identity is verified. A
fusion module receiving as input the d scores has to take a binary decision:
accept or reject identity. This fusion problem has been solved using Support
Vector Machines. The performances of this fusion module have been evaluated and
compared with other proposed methods on a multi-modal database, containing both
vocal and visual modalities.
2 - A Temporal Fusion Algorithm for
Multi-Sensor Tracking in Wide Areas
Olivier Wallart, C. Motamed, M. Benjelloun, Université du
Littoral Côte d'Opale, Calais, France
Abstract: This article presents a distributed vision
system for tracking of mobile objects over wide areas. A temporal data fusion
is used in order to improve the decision making at the data association stage.
The temporal fusion is performed with a possibilistic M.H.T. (Multiple
Hypothesis Tracking). We have decided to use the Possibility Theory which
handle efficiently uncertainties. The originality of this M.H.T. is the control
of its development according to the online quality estimation of the data
association based on a Necessity measurement.
3 - Information Fusion for Intrusion
Detection
Nong Ye, Mingming Xu, Arizona State University, Tempe, AZ,
USA
Abstract: There are many intrusion detection techniques
existed. Each of them can produce the value of Indication and Warning (IW) to
account for how serious the intrusion is. However, each technique reveals
different aspect of the intrusion and has it’s own strength and weakness.
An information fusion technique can be a relatively useful reference for
detecting sophisticated and novel attacks. This paper presents three
information fusion techniques starting from Artificial Neural Network(ANN) and
linear regression , and finally to logistic regression. The performance of
these systems is compared by testing them with data provided by DARPA Intrusion
Detection Evaluation Program (1998).
4 - Integrated Vision and Sound
Localization
Parham Aarabi, Safwat Zaky, University of Toronto, Ontario,
Canada
Abstract: This paper illustrates the synergic advantages
of a multi-modal object localization system utilizing vision and sound
localization. Prototype vision and sound localization systems were developed
and integrated using spatial probability maps, which allow any number of
cameras or microphones with arbitrary orientation to be easily integrated. Test
results show a significant improvement in the system’s ability to
accurately localize objects in low signal to noise situations. Furthermore, the
performance of the integrated system was shown to surpass that of the
individual sub-systems.
Session ThB4:
Detection Fusion 2
Chair: Peter K. Willett, University of Connecticut, Storrs,
CT, USA
Co-Chair: Thiaglingam Kirubarajan, University of Connecticut,
Storrs, CT, USA
1 - Global Optimization for Distributed
and Quantized Bayesian Detection System
Ming Xiang, Chongzhao Han, Jiaotong University, Xi’an,
China
Abstract: Global optimization of distributed detection
system with multi-bit sensor output requires simultaneous solution of optimum
fusion rule and of optimum quantizer mappings for individual sensors. For fixed
sensor quantizer mappings, the optimal fusion rule can be easily shown to be a
likelihood ratio test. But for a fixed fusion rule, the optimal quantizer
mappings are very difficult to determine. In this paper, we consider the case
of conditionally independent sensors. The optimal quantizer mappings for fixed
fusion rule are derived, and optimal solution to the global optimization
problem is obtained through a numerical algorithm.
2 - Effective Simplified Decentralized
Target Detection in Multisensor Systems
Victor S. Chernyak, Scientific Research Institute of Radio
Device Engineering, Moscow, Russia.
Abstract: Decentralized target detection optimization in
multisensor systems, particularly in multisite radar systems, is usually
reduced to the optimum choice of detection thresholds at all sensors and of a
decision rule at the information fusion center. For each decision rule, optimum
peripheral thresholds are calculated by solving a system of m+1 complicated
nonlinear equations in m+1 unknowns where m is the number of sensors. Since
optimum thresholds depend on signal-to-noise ratios, those equations are to be
solved practically in real time. Such a procedure is very cumbersome and
requires large computational resources. In this paper, an alternative simple
and effective approach to the problem is described. The key idea is a "uniform
distribution" of any given output false alarm probability between all sensors.
Thus solving the above mentioned equations and "tuning" local false alarm
probabilities in real time are avoided. The performance analysis has shown that
energy losses are small (with respect to the optimum procedure) for both fixed
threshold and CFAR peripheral detection. Therefore, the suggested approach may
be recommended for practical use.
3 - Unified Fusion Rule in Multisensor
Network Decision Systems
Yunmin Zhu, Department of Mathematics, Sichuan University,
Chengdu, Sichuan, China
X. Rong Li, University of New Orleans, LA, USA
Zhisheng You, Sichuan University, Chengdu, Sichuan, China
Abstract: In this paper, we present a unified fusion rule
in distributed multi-hypothesis multisensor network decision systems, where
communication pattern between sensors is given and fusion center can also
observe data. For the above decision networks, we propose a specific fusion
rule, which is in fact of the most general form (i.e. other fusion rules are
all its special cases) and independent of the statistical characteristics of
observations and decision criteria. It is therefore called unified fusion rule
of the decision network systems. To reach the globally optimal decision
performance, people only need to optimize sensor rules under the unified fusion
rule for given conditional distributions of observations and decision
criterion. Computer simulations support the above results and show some
guidance on how to assign sensors to nodes in a signal detection network with
given communication pattern.
4 - Detection Fusion under
Dependence
Peter K. Willett, University of Connecticut, Storrs, CT, USA
Peter Swaszek, University of Rhode Island, RI, USA
Rick Blum, Lehigh
University, Bethlehem, PA, USA
Abstract: Most results about quantized detection rely
strongly on an assumpt on of ndependence among random variables.With this
assumption removed,little is known. Thus, in this paper, Bayes-optimal binary
quantization for the detection of a shift in mean in a piar of dependent
Gaussian random variables is studied. For certain problem parametrizations
(meaning: the signals and correlation coefficient) optimal quantization is
achievable via a single threshold applied to each observation – the same
as under independence.In other cases one observation is best ignored,or is
quantized with two thresholds; neither behavioris seen under independence.
Further,and again in distinction from the case of independence,it is seen that
in certain situations an XOR fusion rule is optimal, and in these cases the
implied dec sion rule is bizarre.
Session ThB5: Data
Fusion Systems Evaluation and Test-Beds - Invited
Session
Chair: Uri Degen, Advanced Technology Ltd., Tel-Aviv, Israel
Co-Chair: Victor Remez, Advanced Technology Ltd., Tel-Aviv, Israel
1 - On Measures of Performance to
Assess Sensor Fusion Effectiveness
A. Theil, L.J.H.M. Kester, TNO - Physics and Electronics Lab.,
The Hague, Netherlands
Éloi Bossé, DRE Valcartier, Val
Bélair, Québec, Canada
Abstract: In this paper, measures of performance
(MOP’s) with which the effectiveness of sensor fusion can be assessed are
discussed. A series of objective measures related to track management
statistics, to track quality and to reaction time is outlined. The
applicability of several MOP’s is demonstrated for an examples in which
radar and camera data are combined, showing the benefit of sensor fusion.
2 - Empirical Evaluation of Multi-Radar
Tracking Systems
Ari J. Joki, Jouko Saikkonen, Finnish Air Force Headquarters,
Tikkakoski, Finland
Abstract: The formal descriptions of data association and
target-tracking algorithms are forbiddingly formidable for non-expert eyes.
Evaluating performance of tracking systems commercially proposed, based on only
mathematics, requires the work of several full-time experts. It is still
doubtful if the supplier is really implementing what the formulae describe. A
subjective evaluation based on benchmark scenarios is seen to be meaningful for
non-technical decision-makers.
3 - A Test Bed for Sensor Suite and
Multi Sensor Tracking Algorithms Studies
Uri Degen, Victor Remez, Advanced Technology Ltd., Tel-Aviv,
Israel
Abstract: To specify and define a multi-sensor tracking
(MST) system, the tasks of sensor suite definition and data fusion concepts and
algorithms definition should be performed. Data Fusion Test Bed (DFTB) is
designed to model the MST system and to enable simulation of target behavior
(motion and emissions) and sensor detections with or without preprocessing, and
implementation of data fusion and sensor management algorithms and their
post-execution performance evaluation. DFTB outputs are scenarios, simulated
targets and sensor detections (including potential detections), illumination
statistics, intermediate data fusion results, final data fusion results and
analysis results. These results can be presented in form of data files, tables,
graphics and replay. DFTB can run multitude of scenarios automatically without
operator interaction according to operator defined task list.
4 - Refinement of Targets Situation
Picture Quality Evaluation Methodology
Leonid Shvartser, Victor Remez, Advanced Technology Ltd.,
Tel-Aviv, Israel
Abstract: Data Fusion (DF) performance evaluation is
based on the association of “true” and estimated (by DF) Target
Situation Pictures (TaSP), which is followed by a variety of Figures of Merit
(FOM) calculations. However, because of TaSP configuration complexity,
different approaches should be employed for evaluation of different target
activities. Therefore, the decomposition of “true” TaSP into target
activity patterns is a pre-condition for the meaningful DF performance
evaluation. Consequently, only the relevant FOM are calculated for each
particular activity.
Session ThC2: Target
Tracking and Identification
Chair: Subhash Challa, University of Melbourne, Australia
Co-Chair: Christian Musso, ONERA, Châtillon, France
1 - Target Tracking and Identification
Issues when Using Real Data
Elisa Shahbazian, Rodney Hallsworth, Daniel Turgeon, Lockheed
Martin Canada, Canada
Abstract: Since 1991, the Research and Development
(R&D) group at Lockheed Martin Canada (LM Canada) has been developing and
demonstrating the application of Multi-Source Data Fusion (MSDF) techniques for
target tracking and identification onboard the Naval Command and Control (C2)
for of the Canadian Patrol Frigates (CPF). The current C2 as well as the sensor
suite of the CPF were designed in the early 80s within a proprietary hardware
architecture. The sensor data is pre-processed and provided to the C2 without
much consideration of how the data should be combined within the C2. After a
sequence of simulation and modelling efforts for an MSDF capability within the
CPF C2, this project is now at a point when real data captured from a ship
trial on CPF is being injected into the MSDF. This is an on-going activity and
a number of iterations are foreseen, before the MSDF becomes part of CPF C2.
This paper provides a summary of lessons learned in this exercise at this
point.
2 - Target Recognition Using Fuzzy
Fusion Classifier
Pingkui Hou, Xuejun Wang, Xizhi Shi, Shanghai JiaoTong Univ.,
Shanghai, China
Liangji Lin, Mingzhi Zhang, Scientific Test and Control
Technology Institute, Dalian, China
Abstract: This paper makes available a feature fusion
technique in underwater target recognition and proposes the necessary
characteristics for fusion classifier design. Then, Fuzzy fusion classifier
(FFC) is designed based on the characteristics presented. FFC presented here
does not make any assumption about samples distribution and emphasizes mutual
restraints among different classes as well as synthesizes patterns like
combination operation in fuzzy logic. For the application of underwater target
recognition, FFC can efficiently improve classification performance of
recognition system by synthesizing features from multisensors.
3 - A Constructive Bayesian Approach
for Vehicle Monitoring
Y. Xiang, V. Lesser, University of Massachusetts, Amherst, MA,
USA
Abstract: A key componentof a vehicle monitoring system
is uncertainty management. Bayesian networks (BN) have emerged as a normative
and effective formalism for uncertain reasoning in many AI tasks. Since a
priori modeling of the domain into a BN is impractical due to the vast
interpretation space, the BN formalism has been considered inapplicable to this
type of task. We propose a framework in which the BN formalism can be applied
to vehicle monitoring. The framework explores domain decomposition, model
separation, model approximation, model compilation and re-analysis.
Experimental implementation demonstrated good performance at near-realtime.
4 - Target Tracking Incorporating
Flight Envelope Information
Subhash Challa, University of Melbourne, Parkville, Victoria,
Australia
Niklas Bergman, Linköping University, Sweden
Abstract: Target tracking in usually performed assuming
the target under consideration is adequately modelled asa point target with no
reference to constraints on its motion. However, in reality, target motion is
constrained by achievable maximum speeds and accelerations. This is
non-standard information and its incorporation into conventional tracking,
while promising significant benefits, poses a signifi cant challenge in
constructing the posterior probability density function (PDF) of target state.
In this paper, the optimal Bayes' recursion for this posterior PDF is derived
and an algorithm for implementing this recursion based on stochastic sampling
techniques is presented. An illustrative example demonstrating the benefits of
incorporating this information is also provided.
Session ThC3: Security
and Surveillance - 2
Chair: Isabelle Bloch, ENST Paris, France
1 - Dynamic Service Definition in the
Future Mixed Surveillance Environment
Christos M. Rekkas, Jean-Marc Duflot, Pieter van der Kraan,
Eurocontrol, Brussels, Belgium
Jean-Claude Rassou, Thomson-CSF/ISR, Massy,
France
Abstract: The future Surveillance environment is expected
to be heterogeneous and include new types of Data Sources, such as the Mode S,
ADS-Broadcast and ADS-Contract, in addition to the classical radar sensors
(i.e. PSR, SSR). The new types of sensors are capable of transmitting aircraft
derived data of high importance for the Air Traffic Management functions (such
as the conflict detection and resolution tools, flight data processing etc.)
which make use of Surveillance data. Various types of transmission
characteristics for the aircraft derived data are foreseen (periodic, event
driven etc.). This paper presents the work on the development of a prototype
function which dynamically defines the services to be requested from the
Surveillance Data Sources in the future Surveillance data fusion environment.
The Dynamic Service Definition (DSD) function has a set of inputs, including
sensor data, track data, geographical data etc., (reflecting the current
Traffic and Environment Situation), as well as the requests from the
Surveillance Users. On the basis of these inputs and defined operational
criteria, the function dynamically defines the most appropriate set of
contracts which will be requested from the Data Sources, in order to
“optimise” the Traffic Situation Picture and satisfy the requests of
the Users. The contracts will be requested in order to receive a user defined
set of data items for a corresponding set of tracks at specified transmission
characteristics.
2 - Sensor Modeling and Data Processing
for Airport Simulation
Antonello Pasquarelli, Alenia Marconi Systems, Roma, Italy
Andrea Corsini, Andrea Garzelli, University of Siena, Italy
Abstract: This paper presents the SEEDS simulation
environment for the evaluation of distributed traffic control systems. The
description starts with a general overview of the simulator, targeted for
airport surface traffic simulation, and then focuses on the sensor models
implemented in the prototype. The surveillance function foreseen in a real
Advanced Surface Movement Guidance and Control Systems (A-SMGCS) has been
studied and modeled; suitable set of sensors and signal processing algorithms
have been considered and their performances have been analyzed in order to be
compliant with the application performance requirements defined by
International Organizations. The paper shows the sensor module architecture,
how the sensors have been modeled and how the software module has been
implemented and integrated in the core simulator. The interactions with the
other modules of the simulator and the exchanged messages are also
described.
3 - Characterization of Mine Detection
Sensors in Terms of Belief Functions and their Fusion, First Results
Nada Milisavljevi¢, Royal Military Brussels, Belgium and
ENST Paris, France
Isabelle Bloch, ENST Paris, France
Marc Acheroy,
Royal Military Academy, Brussels, Belgium
Abstract: In this paper, characterization of mine
detection sensors in terms of belief functions and their fusion are presented.
Need for fusion of mine detection sensors is discussed, and reasons for
choosing Dempster-Shafer framework are pointed out, taking into account
speciÝcity and sensitivity of the problem. This work is done in the
scope of the HUDEM project, where three promising and complementary sensors are
under analysis. These sensors are presented, and detail analysis is performed
in case of fusing the data from them. A way for including in the model
inÛuence of various factors on sensors and their results is discussed as
well and will be further analyzed in the future. The application of the
approach proposed in this paper is illustrated on the frequent case of
detecting metallic objects, but the possibility for modifying it to some other
situations exists.
Session ThC4: Target
Detection
Chair: Chongzhao Han, Jiaotong University, Xi'an, China
Co-Chair: Alexander Tartakovsky, University of Southern California, Los
Angeles, CA, USA
1 - Entropy Based Optimization for
Binary detection Networks
Denis Pomorski, Université des Sciences et Technologies
Lille I, Villeneuve d'Ascq, France
Abstract: This contribution deals with the binary
detection networks optimization using an entropy based criterion. The
optimization of an elementary detection component consists in applying a
variable threshold on the likelihood ratio, which depends on a posteriori
probabilities. A gradient algorithm is proposed in order to find this
threshold. The optimization results of the elementary detection component using
entropy and Bayes' criteria are compared : the proposed approach has a very
interesting property of robustness with respect to rare events, or with respect
to events for which a priori probabilities are uncertain. In particular, the
obtained ROC curve does not recede from the ideal point.
2 - Implementation of Hough Transform
as Track Detector
Kiril Alexiev, Bulgarian Academy of Sciences, Sofia,
Bulgaria
Abstract: Hough Transform is a convenient tool for
features extraction from images. In this paper an implementation of Hough
Transform is onsidered for automatic track initiation in the surveillance radar
space. The influence of Hough parameter space granularity upon probability of
track detection is analyzed. Analytical expressions for probability of track
initiation using Hough Transform are derived in the presence of normal
distributed additive system noise, measurement noise and without any noises. A
new parameter space structure, matching with measurement errors is proposed.
The Monte Carlo simulation confirms received analytical result.
3 - Sequential Testing of Multiple
Hypotheses in Distributed Systems
Alexander Tartakovsky, University of Southern California, Los
Angeles, CA, USA
X. Rong Li, University of New Orleans, LA, USA
Abstract: It is supposed that there is a multisensor
system in which each sensor tests a finite number of hypotheses in a sequential
manner. Then the decisions are transmitted to a fusion center, which combines
them to improve the performance of the system. First, we propose a local
multihypothesis sequential test procedure which allows one to fix the
probabilities of errors at specified levels and is asymptotically optimal for
general statistical models in the sense of minimizing the expected sample size
when the probabilities of errors vanish. We then construct two fusion rules :
non-sequential and sequential. The first fusion rule waits until all the local
decisions in all sensors are made and then fuses them. It is optimal in the
sense of minimizing the average probability oferror (Bayes criterion) or the
maximal probability of error (minimax criterion). In contrast, the sequential
fusion rule fuses local decisions one by one in the order they are made, and at
each stage decides whether to continue fusion or to stop and make a final
decision. It has an advantage over the first rule in that it reduces the total
time to make a final decision for a given average probability of error. An
example of fusion of binary local decisions shows that the final decision can
be made substantially more reliable even for a small number of sensors (3-5).
4 - Global Optimization for Distributed
Detection System under the Constraint of Likelihood Ratio Quantizers
Ming Xiang, Chongzhao Han, Jiaotong University, Xi’an,
China
Abstract: An optimal solution to global optimization
problem usually can be obtained only for conditionally independent sensors. As
for dependent sensors, although the necessary conditions for global
optimization can be found, an optimal solution usually can not be obtained.
Thus, for distributed detection systems consisting of dependent sensors, some
suboptimal global optimization method need to be considered. In this paper, we
consider this suboptimal global optimization problem for distributed and
quantized Bayesian detection system. We constrain the sensor quantizers to be
likelihood ratio quantizers, and optimize the system performance under this
constraint.
Session ThC5:
Recognition
Chair: André Ayoun, Thomson-CSF/ISR, Massy, France
Co-Chair: Michel Prenat, Thomson-CSF Optronique, Guyancourt, France
1 - Model Generation and Cooperation in
On-Line Omni-Writer Handwritting Recognition
Lionel Prevost, Maurice Milgram, Université Paris VI,
Paris, France
Abstract: In this paper, we introduce a new method for
on-line character recognition based on the cooperation of two classifiers. The
first one is a k-nearest-neighbor classifier, the second one is an evolutionary
neural classifier. Several cooperation architectures (already tested in OCR but
seldom in on-line recognition) are presented, from the easier (weigthed sum of
both classifier outputs) to the most complicated (integrating neural network).
The recognition improvement varies between 30% and 50% according to the merging
strategy. We try to appreciate each method asset on recognition rate and speed.
Results are presented on 52 different character classes (upper and lower case
letters) and more than 50000 examples from UNIPEN database.
2 - Multiple Expert System Design by
Combined Feature Selection and Probability Level Fusion
Fuad M. Alkoot, J. Kittler, University of Surrey, Guildford, UK
Abstract: We propose a novel design philosophy for expert
fusion by taking the view that the design of individual experts and fusion
cannot be solved in isolation. Each expert is constructed as part of the global
design of a final multiple expert system. The design process involves jointly
adding new experts to the multiple expert architecture and adding new features
to each of the experts in the architecture. We evaluate the performance of
different fusion strategies ranging from linear untrainable strategies like Sum
and Modified Product to linear and nonlinear trainable strategies like logistic
regression, single layer perceptron and radial basis function classifier. We
investigate two distinct design strategies which we refer to as parallel and
serial. In both cases we show that the proposed integrated design approach
leads to improved performance.
3 - An Entropy Method For Multisource
Data Fusion
Bienvenu Fassinut-Mombot, Jean-Bernard Choquel,
Université du Littoral Côte d'Opale, Calais, France
Abstract: The present paper proposes a generic model of
the multisource data fusion in the framework of the theory of information, with
closer attention being given the different nature ofdata processed incommon
cases. This model is then used to elaborate processing methods able to face
specific problems that may arise when multisource systems are implemented to
achieve functions like classification and pattern recognition, matching of
ambiguous observations, estimation, detection or tracking. Crucial practical
problems to data fusion are more specifically dealt with, such as information
representation, appropriate combination processing and decision making. Some
clues are given on the practical use and implementation of such an approach,
for example, in the distributed estimation problem.
Monday - Tuesday - Wednesday - Thursday
Last Updated: October 6,
2000
Web site by: dezert@onera.fr (content),
gaultier@onera.fr (form)
copyright © ISIF 2000 |