TOPIC 8: LEVEL 1 IDENTIFICATION

David L. Hall

TOPIC OBJECTIVES

Continue introduction of Level-1 processing with focus on attribute fusion (e.g., for target identification)

Introduce common pattern recognition algorithms

Understand issues and limitations

LEVEL-1 (IDENTITY DECLARATION)

JDL LEVEL ONE PROCESSINGOBJECT REFINEMENT

JDL Level One Processing Object Refinement

Data Alignment

•Spatial Reference Adjustment

•Temporal Reference Adjustment

•Units Adjustment

Data/Object Correlation

Object

Positional Estimation

• System Models• Optimization Criteria• Optimization Approach• Processing Approach

Object Identity

Estimation

• Physical Models• Feature-based Inference Techniques

• Cognitive-based Models

CA

TEG

OR

YFU

NC

TIO

NP

RO

CES

S

• Gating• Association Measures• Assignment Strategies

Sources HumanComputerInteraction

DATA FUSION DOMAIN

Level OSignal

Refinement

Level OneObject

Refinement

Level TwoSituation

Refinement

Level ThreeThreat

Refinement

Level FourProcess

Refinement

Database Management System

SupportDatabase

FusionDatabase

CONCEPTUAL PROCESSING FLOW FOR LEVEL 1 FUSION

BulkGating

DataAssociation

Position/Kinematic/Attribute

Estimation

IdentityEstimation

• Observation File• Track File• Sensor Information

Sensor#1

PreprocessingData

Alignment

Sensor#2

PreprocessingData

Alignment

SensorN

PreprocessingData

Alignment

•

•

•

•

•

•

•

•

•

METHODS FOR ATTRIBUTE FUSION

Identity or classof entity, object or activity •

• •

Feature-basedClassificationMethods• Neural nets• Cluster algorithms• Parametric templates

Decision-based methods• Voting• Decision trees• Logical templates• Bayesian Belief nets• Dempster-Shafer method• Fuzzy logic• Rule-based

Systems

Fea

ture

ex

tra

ctio

n

S1

S2

SN

Model-based methods(high-fidelity physical models)

Raw Data FeaturesDeclarationof Identity

Declaration of Identity

EXAMPLE OF SINGLE-SENSOR FEATURE BASED OBJECT IDENTITY

DECLARATION

Pro

pag

atio

n M

edia

Sensors

Signal Space

Fea

ture

Ext

ract

ion

y

Target Models of a priori data

Classifier

• Cluster Methods• Neutral Networks• Templating• etc.

Feature Space

Target Class A

Target Class B

DecisionSpace

EnergySensorReaction

Signal,Image

Feature Vector

Declaration Of Identity

CLUSTER ANALYSIS

Cluster Analysis:Cluster Analysis: Basic use is for classification analysis based on multi-parameter

similarity. Provides estimation of pair-wise/cluster-wise similarities. Supports parametric model development. Helpful when studying new entities/new parameters. Can be computationally demanding.

CONCEPT OF CLUSTER ANALYSIS

Sensor A

Sensor B

Sensor C

• • •

TaggedDataSet:

ObservationsAssociated

withSpecificObjects

Selectionand

Calculationof

ResemblanceCoefficients

Selectionand

Calculationof

ClusteringMethod

ClusteringThresholdSelection

ClusterDefinition

Object

Ob

jec

t

ResemblanceCoefficients

Observation/Featurei

Observation/Featurej

Cluster i

Cluster j y

EXAMPLE: HELICOPTER TRANSMISSION FAULT

CLASSIFICATION

Purpose: Classify faults using data collected from the

aft transmission of a Westland CH-46E helicopter 8 accelerometers 7 faults and a no-fault Faulty and seeded fault components

used

Results: Achieved robust classification using feature

reduction Separates faults with similar signatures but

with significantly different criticalities failure progressions

Sp

iral

Be

vel

Inp

ut

Pin

ion

Sp

alli

ng

Input PinionBearing Corrosion

Quill ShaftCrack

CollectorGearCrack Helical

Input PinionChipping

HelicalIdler Gear

Crack

NoDefect

A. K. Garga et al, ”Fault Classification in Helicopter Signals,“ Proc. Amer. Helicopter Soc. 53rd Annual Forum, 1997.

CLUSTER ALGORITHMIC APPROACHES

Hierarchical agglomerative methods Iterative partitioning methods Hierarchical divisive methods Density search methods Factor analytic methods (based on correlation

matrix processing) Clumping methods (allows membership in more

than one class) Graph theoretic methods

ASSESSMENT OF CLUSTER METHODS

The good news Requires no a priori “knowledge” of data or

target physical characteristics Allows exploration of features and classes Simple to use Extensive COTS software available, see the

review at site http://www.pitt.edu/~csna/software.html

The bad news Requires extensive training data Results dependent upon

Association measure Scaling of feature components Order of data processed Clustering scheme Specific training data etc

OVERVIEW OF ADAPTIVE NEURAL SYSTEMS

InputVector

OutputVector

w0

w2

w1

wn

x0

x2

x1

xn

f(y)Output

Adaptive linear combiner

y

THE ACTIVATION FUNCTION

*V.R. Hush and B.G. Hornefs(y) = (1 - ey)-1

1.0

0.0

-10.0 0.0 10.0

f(y)

y(a)

1.0

0.0

-10.0 0.0 10.0

f(y)

y(b)

=5.0

=1.0

=0.2

EXAMPLE: HANDWRITING RECOGNITION

F

eatu

re V

ecto

r

1 = a

0 = b

0 = c• • • • • •

0 = z

NEURAL NETWORK ISSUES AND LIMITATIONS

• Choosing the Network SizeChoosing the Network Size- Tradeoff between too large and too small- Emerging systematic techniques for size selection

• Complexity of LearningComplexity of Learning- Back-propagation (BP) Methods notoriously slow- Weighting search problem is NP-complete

• GeneralizationGeneralization- How much training data required for general results?

(rule of thumb is 10 x wij)

- Generalization error (error on training data vs actual problem)

• Network InterconnectivityNetwork Interconnectivity- Optimal brain damage- Complexity regularization- Weight sharing

PARAMETRIC TEMPLATING AND DECISION TREES

If we know a priori the parametric “boundaries” related to decision or identification classes then we may represent these via parametric templates, decision-trees, or rule-based systems; e.g.

Mechanical system fault if Engine temperature exceeds Tcritical

Possible bearing failure if vibration exceeds X, Etc.

COMBINED SYNTACTIC/CONTEXTUAL TARGET MODELING

• Found at specified altitudes• Minimum speed equals 150mph• Travels in specified groups

Model consists of target signature and contextual informationModel consists of target signature and contextual information

Stored Models

Weather (MET)

Time/Season

Range

Sensor Phenomenology

TargetSignature

ContextualModel

Design Target Model

PATTERN RECOGNITION

SYNTACTICAL COMPOSITION

CONTEXTUAL INTERPRETATION

ALTERNATE ARCHITECTURES FOR MULTISENSOR IDENTITY FUSION

ASSOCIATION

DecisionLevel

Fusion

IdentityDeclaration

JointIdentity

Declaration

IdentityDeclaration

IdentityDeclaration

IdentityDeclaration

I/DA

I/DB

I/DN

• • •

FEATURE

EXTRACTION

SensorA

SensorB

SensorN

• • •

A. Decision-Level FusionA. Decision-Level Fusion

ALTERNATE ARCHITECTURES FOR MULTISENSOR IDENTITY FUSION

B. Feature-Level FusionB. Feature-Level Fusion

ASSOCIATION

FeatureLevel

Fusion

IdentityDeclaration

JointIdentity

Declaration

FEATURE

EXTRACTION

SensorA

SensorB

SensorN

• • •

ALTERNATE ARCHITECTURES FOR MULTISENSOR IDENTITY FUSION

C. Data-Level FusionC. Data-Level Fusion

IdentityDeclaration

JointIdentity

Declaration

FEATURE

EXTRACTION

ASSOCIATION

SensorA

SensorB

SensorN

• • •

DataLevel

Fusion

FUZZY MATHEMATICS

The world of human cognition is not binaryThe world of human cognition is not binary Many concepts are not defined with math precision:Many concepts are not defined with math precision:

Examples of fuzzy notions about two • somewhat heavy ugly • handsome tall • borderline

Interpretation is context dependent Fuzzy set theory argues that imprecision is an intrinsic property of various Fuzzy set theory argues that imprecision is an intrinsic property of various

notionsnotionsNot an approximation of truthNot a failure to comprehendAn admission that some notions may forever be imprecise

Do not try to quantify the unquantifiable; but formalize a way to deal with itDo not try to quantify the unquantifiable; but formalize a way to deal with it

FUZZY SETS: MATHEMATICS

Introduce the membership function:Introduce the membership function:

A(X) [0, 1] V x E

Fuzzy sets are sets of ordered pairs:Fuzzy sets are sets of ordered pairs:

(x, (x)) Note:Note:

BOOLEAN FUZZY SETs

(x) [x, (x)]

T or F partial truth/uncertainty feasible

Membership functions are not unique:Membership functions are not unique:

Varying solutions

Sensitivity analysis to choice of membership function

FUZZY SETS:ELEMENTARY OPERATIONS

INCLUSION:INCLUSION: A B A (X) ≤ B (X) EQUALITY:EQUALITY: A = B A (X) = B (X) COMPLEMENTATION:COMPLEMENTATION: A (X) = 1 - A (X) UNION:UNION: (X)AB = MAX [A (X), B (X)] INTERSECTION:INTERSECTION: (X)AB = MIN [A (X), B (X)] DIFFERENCE: DIFFERENCE: (X)AB = MAX [A (X), B (X)]

FOR EXAMPLE, LETFOR EXAMPLE, LETA = (X10.2, X2 0.7, X3 1, X4 0.1)

ANDANDB = (X10.5, X2 0.3, X3 1, X4 0.0)

THENTHEN AB = (X10.5, X2 0.7, X3 1, X4 0.1)

ANDAND AB = (X10.2, X2 0.3, X3 1, X4 0.0)

INTERPRETING SOME FUZZY SET OPERATIONS: INTERSECTION

By definition, an element in a fuzzy set can reside partly in one set By definition, an element in a fuzzy set can reside partly in one set and partly in another (including the complementary set)and partly in another (including the complementary set) An element cannot be more true in the intersection than it is in either set An element cannot be in the intersection to a degree more than it is in one of

the subsets; this argues for min () Intersection creates a middle level type set

EXAMPLE: the intersection of TALL and NOT TALL setsEXAMPLE: the intersection of TALL and NOT TALL sets

DATA FUSION WITH FUZZY LOGIC

SENSOR1

y1

SENSORN

yN

FU

ZZ

IFIC

AT

ION

A

B

FUZZY LOGIC

• If A and B C• If A and B D

C

D

D

EF

UZ

ZIF

ICA

TIO

N

QuantifiedInferences

FuzzyMembershipFunctionXforms

Fuzzy RulesFuzzy Calculus

InverseFuzzyMembershipXforms

• Ad Hoc• Neural Nets• Templates• Etc.

APPLICABILITY OF TECHNIQUES FOR LEVEL 1 FUSION

Spatial Adj.

Temp. Adj.

Units Adj.

Screen-ing

Correla-tion

Assign-ment

Obs. Predict

State Update

Uncer-tainty Mgmt.

Object Mgmt.

ID Mea-sures

Com-pari-son

Declare

ID

Uncer-tainty Mgmt.

Algorithms and Techniques

Data

Alignment Data/Object Correlation

Object positional/ Kinematic/Attribute Estimation

Object Identity Estimation

Coordinate Transforms X X X X X Sensor Models X Physical Models X X X X X Association Measures X X X X X Assignment Logic X X X Equations of Motion X Optimization Methods X X

Kalman Filters X X Covariance Error X X Bayesian Inference X X X X Dempster-Shafer X X X X Voting X X Pattern Recognition X X X X Templating X X X X X Expert Systems X X X X Fuzzy Sets X X X

TOPIC 8 ASSIGNMENTS

Preview the on-line topic 8 materials Read chapter 5 of Hall and McMullen (2004) Writing assignment 7: Develop a one-page discussion of how

level-1 identification and pattern recognition applies to your selected application.

Discussion 4: Discuss the concept of identification; how have automated identification processes and sensors (e.g., tags on objects, cell phones, smart cards, etc) become integrated into common activities? What are issues of failure in automated identification techniques?

DATA FUSION TIP OF THE WEEK

Here is an ancient Chinese classification of animals:"Animals are divided into (a) those that belong to the Emperor, (b) embalmed ones, (c) those that are trained, (d) suckling pigs, (e) mermaids, (f) fabulous ones, (g) stray dogs, (h) those that are included in this classification, (i) those that tremble as if they were mad, (j) innumerable ones, (k) those drawn with a very fine camel's hair brush, (l) others, (m) those that have just broken a flower vase, and (n) those that resemble flies from a distance." from Other Inquisitions: 1937-1952 by Jorge Luis Borges Downloaded from http://www.alaska.net/~royce/Funny/classify.html July 30, 2008

It is easy to forget that identification and classification are inherently a labeling process (attaching labels to physical objects, activities and events);

Such classifications may not actually be observable or possible – the link between observable features and classes may not be feasible with any technique