Title: Multisensor Data Fusion
1 - Multisensor Data Fusion
- Andrés Navarro
2Contents
- Introduction
- Definition
- Sensors and levels
- Fusion models
- Fusion techniques
- Distributed Data Fusion
- Scenarios
- Conclusions
3Introduction
- Motivation
- Master Thesis
- First approach to sensor fusion
- Overview of state-of-the-art
- Guide for people who do not know about sensor
fusion to get introduced to the issue
4Introduction
- Why is sensor fusion needed?
- Research questions
- What is multisensor data fusion? Is there an
unanimous definition? - What models for multisensor data fusion exist in
the literature? Do they have common descriptions?
Do they contradict each other? - What techniques or methods can be used in
multisensor data fusion? - Centralized or decentralized data fusion?
- What scenarios can multisensor data fusion be
applied?
5Definition
- JDL Data Fusion Lexicom
- A. Steinberg and C. Bowman.
- Wald
A process dealing with the association,
correlation, and combination of data and
information from single and multiple sources to
achieve refined position and identity estimates,
and complete and timely assessments of situations
and threats, and their significance. The process
is characterized by continuous refinements of its
estimates and assessments, and the evaluation of
the need for additional sources, or modification
of the process itself, to achieve improved
results.
Data fusion is the process of combining data or
information to estimate or predict entity states.
Data fusion is a formal framework in which are
expressed means and tools for the alliance of
data originating from different sources. It aims
at obtaining information of greater quality the
exact definition of greater quality will depend
upon the application.
6Sensors and levels
Commensurate multisensor data
Non commensurate multisensor data
Feature-level Fusion
Information extraction
Direct Data Fusion
High-level Fusion
7Fusion models
- JDL data fusion model
- Dasarathy's functional model
- Waterfall fusion process model
- Boyd Loop
- Thomopoulos' Fusion Model
- Durrant-Whyte architecture
- The Omnibus process model
- Endsley's Situation Awareness
- General Data Fusion Architecture
8Fusion models
- JDL data fusion model
- Sources
- Sources preprocessing
- Level 1
- Level 2
- Level 3
- Level 4
- Database management system
- HCI
9Fusion models
- JDL model revisions
- Drawbacks
- Different revisions
- New definitions
10Fusion models
- Dasarathy's functional model
- Levels of abstraction
- Data
- Feature
- Decision
- Categorization of data fusion functions in terms
of the type of data level at input/output.
11Fusion models
- Waterfall fusion process model
- Fusion processin stages
- Omission offeedback dataflow is themajor
limitation.
12Fusion models
13Fusion models
- Thomopoulos' Fusion Model
14Fusion models
- Durrant-Whyte achitecture
15Fusion models
- The Omnibus process model
16Fusion models
- Endsley's Situation Awareness
17Fusion models
- General Data FusionArchitecture
- Network based
- Levels described as classes with attributes and
functions.
18Fusion models
- Models classification
- Elmenreich
- Abstract Generic Rigid
- Durrant-Whyte and Henderson
- Architecture Meta, Algorithmic, Conceptual,
Logical and Execution - Centralized - Decentralized
- Local Global
- Modular Monolithic
- Heterarchical - Hierarchical
19Techniques
- Classification
- Overview
- Kalman Filter
- Probabilistic Inference
- Artificial Neural Networks
- Fuzzy Logic
- Support Algorithms
- Selection
20Techniques
21Techniques
- Classified by
- JDL levels
- Type of method
- Fusion problems
22Techniques
- JDL levels
- Type of method
23Techniques
- Classified by
- JDL levels
- Type of method
- Fusion problems
24Techniques
- JDL levels
- Type of method
- Fusion problems
- Data Association
- Estimation
- Identity declaration
- Decision-level identity fusion
25Techniques
Extended Kalman Filter
Discrete Kalman Filter
Assumed noise
26Techniques
- Kalman Filter with INS
- Inertia System
- Good high frequency information
- Drift at a slow rate
- Other Position System
- Good data at low frequency, on the average
- High frequency noise
The Kalman Filter approach is instead to use the
statistical characteristics of the errors in both
the external information and the inertial
components to determine this optimal combination
of information. Actually, the filter
statistically minimizes the errors in the
estimates of the navigation parameters on an
ensemble average basis, no other means of
combining the data will outperform it, assuming
the internal model in the filter is adequate.
27Techniques
- Filter fails ? System Fails
- High sample rate ? CPU load
- Erros in the inertial must remain of small
magnitude
28Techniques
- Bayesian Inference
- It can be used to discriminate between
conflicting hypotheses - Initial beliefs are needed before any evidence is
ever collected - Sensorfusion
29Techniques
- Bayesian Networks
- Probabilistic graphical model that represents a
set of variables and their probabilistic
interdependencies - Algorithms to perform inference and learning
- Dynamic Bayesian Networks
- Extension of Bayesian networks that allows the
representation of temporal information ? Signals
- Hidden Markov models
- Model for Markov process a stochastic process in
which the probability distribution of the current
state is conditionally independent of the path of
past states
30Techniques
- Probabilistic Inference
- Dempster-Shafer Theory
- Generalization of Bayesian Theory Instead of
requiring probabilities for each question, belief
functions are used. - Two ideas
- Obtain degrees of belief for one question from
subjective probabilities fora related question. - Use Dempster's rule for combining these degrees
of belief
Sensor1
Sensor2
m1(u0)?
m2(u0)?
m(u0)m1(u0)m2(u0)?
31Techniques
- Probabilistic Inference
- Dempster-Shafer Theory
- Generalization of Bayesian Theory Instead of
requiring probabilities for each question, belief
functions are used. - Two ideas
- Obtain degrees of belief for one question from
subjective probabilities fora related question. - Use Dempster's rule for combining these degrees
of belief
Sensor1
Sensor2
m1(u0)?
m2(u2)?
m(u0)m1(u0)m2(u2)?
32Techniques
- Probabilistic Inference
- Dempster-Shafer Theory
- Generalization of Bayesian Theory Instead of
requiring probabilities for each question, belief
functions are used. - Two ideas
- Obtain degrees of belief for one question from
subjective probabilities fora related question. - Use Dempster's rule for combining these degrees
of belief
Sensor1
Sensor2
m1(u0)?
m2(u1)?
Dempster's rule
33Techniques
- Generalized evidence processing theory
- Unifies the Bayesian theory with the D-S,
combining their advantages and avoiding their
disadvantages - Each sensor collects evidence and assigns the
evidence via probability masses unlike D-S, GEP
assigns and combines probability masses based on
the a priori conditional probability of the
hypotheses.
34Techniques
- Artificial Neural Networks
- Computational model of biological neural
networks - Densely interconnected set of artificial neurons
simple units as perceptron - Feed-forward / recurrent
- Non linear statistical data modelling
- They can learn a complex relationship between
inputs and outputs, normally established by the
unit weights - Learning
- Units' weights ? Backpropagation algorithm
- Network structure
35Techniques
- Artificial Neural Networks
- Neurons can be trained to represent sensory
information and, through associative recall,
complex combinations of the neurons can be
activated in response to different sensory
stimuli. - The main advantage of neural networks for
multisensor fusion is that there is no need of a
model for the sensors or the uncertainties
36Techniques
- Fuzzy Logic
- Fuzzy set theory
- Elements have degrees of membership to the
different sets, differing from classical set
theory, where elements belong or do not belong to
a certain set. - Rules IF antecedent THEN consequence
- Operators OR, AND and NOT
- Steps
- Fuzzyfcation
- Rule evaluation
- Aggregation
- Defuzzyfication
37Techniques
- Fuzzy Logic
- Uncertainty in multisensor fusion can be directly
represented in the inference process by allowing
each proposition to be assigned a degree of
truth. - Fuzzy Fusion Network
- Input data
- Feature Extraction
- Feature Level Fusion
- Decision Level Fusion
38Techniques
- Support Algorithms
- Required functions for the fusion system
- Library of basic numerical methods
- Database management
- Man-Machine Interaction
- Sensor Management
39Techniques
- Techniques Selection
- Goals
- Maximum effectiveness Algorithms should make
inferences with maximum specificity in the
presence of uncertain and missing data, dealing
with minimal or no available a priori
information. - Operational constraints and time constraints must
be considered. - Resource efficiency in CPU and communications
load. - Operational flexibility to account for
operational needs with changing a priori data. - Functional growth.
40Techniques
- Techniques Selection
- Steps
- Identifying categories of data-processing
techniques or algorithms. - Surveying existing prototype and fielded data
fusion systems. - Analyzing system requirements.
- Analyzing and defining operational concepts for
manual and automatic processes. - Identifying preliminary algorithms
- Performing trade-off analyses of algorithm
effectiveness versus required system resources. - Preparing detailed designs and prototypes of
selected algorithms. - Refining and tuning the algorithms.
41Distributed Data Fusion
- A collection of processing nodes, connected by
communication links, in which none of the nodes
has knowledge about the overall network
topology. - Requirements
- Dynamic Topology Management
- Information- Sharing Strategies
- Algorithms
42Scenarios
- Image Data Fusion
- Advantages
- Reduction of overall uncertainty and increase of
accuracy. - New features in a scene can be perceived
- More timely information is available
43Scenarios
44Scenarios
- Ambient Intelligence
- Electronic environments that are sensitive and
responsive to the presence of people. - Characteristics
- Embedded
- Context-aware
- Personalized
- Adaptive
- Anticipatory
45Scenarios
- Augmented Reality
- Integration of virtual content in a real
environment in real time.
- Alignment between virtual content and real world
- Fusion of tracking systems
- INS
- GPS
- Ultrasound
- Vision-based
Accurate position
46Scenarios
- Augmented Reality
- Interaction with other perceptual systems
- Orienting, Auditory,Haptic, Taste, Smell
- Feature extraction ? Human Behaviour
- Sensors Position, orientation, body gestures,
speech, vital signs, eyetracking... - Human behaviour and experience models and
simulations
47Conclusions
- What is multisensor data fusion? Is there an
unanimous definition? - Most of definitions are restrictive to a certain
terminology and applications - A broader definition is needed to cover such a
wide diversity of sensor fusion applications - Wald's definition is chosen
- Discussion will continue
48Conclusions
- What models for multisensor data fusion exist in
the literature? Do they have common descriptions?
Do they contradict each other? - A common point in most of them is the need of
divide the fusion process in levels of data
abstraction. - More disagreement is found in the idea of a
cycling processing. - Relationship between specification and usability
- Combine the underlying ideas for the final design.
49Conclusions
- What techniques or methods can be used in
multisensor data fusion? - Data fusion at different levels of abstraction
implies the use of multiple techniques. - A layout or scheme for the implementation of any
kind of sensor fusion application is not
feasible. - The design of the fusion algorithms is a lengthy
task where multiple fusion techniques can be
combined.
50Conclusions
- Centralized or decentralized data fusion?
- The decentralized fashion has some advantages and
some disadvantages comparing to the centralized
one. - The implementation of Distributed Data Fusion
requires - The use of certain fusion models that allow
decentralization. - Specific algorithms
51Conclusions
Multisensor Data Fusion is a broad issue due to
the wide range of scenarios that it can be
applied to. Therefore, to find a definition, a
model or an algorithm scheme that is explicit,
meaning that it can be followed to implement a
real system, and at the same time usable for any
kind of application, is an unfeasible task.
Hence, a view of the different approaches,
theories and implementations in the issue of
sensor fusion can be presented, intending to be
useful as a collection of different ideas that
should be combined in the implementation of a
real fusion system.
52 - Multisensor Data Fusion
- Andrés Navarro