Upload
dalton
View
73
Download
1
Tags:
Embed Size (px)
DESCRIPTION
AUTOMATIC TARGET RECOGNITION AND DATA FUSION. March 9 th , 2004 Bala Lakshminarayanan. Outline. Introduction Distributed processing DSN topologies Data fusion SFTB project Classification results. Introduction. Automatic Target Recognition Classify civilian targets with high accuracy - PowerPoint PPT Presentation
Citation preview
AUTOMATIC TARGET RECOGNITION AND DATA FUSION
March 9th, 2004
Bala Lakshminarayanan
Bala Lakshminarayanan, 9th March 2
Outline
Introduction Distributed processing DSN topologies Data fusion SFTB project Classification results
Bala Lakshminarayanan, 9th March 3
Introduction
Automatic Target Recognition Classify civilian targets with high accuracy
7 targets 3 sensors (IR, Grayscale, Acoustic) 3 nodes 3 scenarios
Nodes are placed along the road
Bala Lakshminarayanan, 9th March 4
Processing paradigms…(1)
Centralized processing – fusion center High communication bandwidth Higher network cost Non-optimal processing, esp. when sensor
coverage does not overlap Central node dependence
Bala Lakshminarayanan, 9th March 5
Processing paradigms…(2)
Distributed processing Redundancy – accurate classification Lesser network cost Reduced bandwidth requirement though
increased communication between nodes Better response to rapid changes Needs proper architecture Energy efficiency
Bala Lakshminarayanan, 9th March 6
Serial topology
Event (H)
Sensor 1 Sensor 2 Sensor n
y1y2 yn
u1 u2 un
yi : Local observation
ui : Local decision variable
n : Number of sensors
u0 : Global decision variable
Bala Lakshminarayanan, 9th March 7
Parallel topology
Event (H)
Sensor 1 Sensor 2 Sensor n
y1y2 yn
u1 u2 un
• Classifier Selection model
• Each classifier is an “expert”
• For feature x, classifier in its vicinity is given highest credit
Bala Lakshminarayanan, 9th March 8
Parallel topology
Event (H)
Sensor 1 Sensor 2 Sensor n
y1y2 yn
u1 u2 un
Fusion Centeru0
• Classifier Fusion model
• All classifiers trained over entire feature space
• Competitive model
Bala Lakshminarayanan, 9th March 9
Data fusion…(1)
Disparate sensors used for data collection Need to integrate results – fuse sensor data
to give user ability to decide better Objective of fusion is to give one reliable,
robust decision rather than many uncertain decisions
Fusion levels Temporal Multi-modal Multi-sensor (from different nodes)
Bala Lakshminarayanan, 9th March 10
Data fusion…(2)
Temporal fusion Independent frames Majority voting
Multi-modality fusion Different sensing modalities, all exposed to whole
feature space Competitive rather than complementary BKS algorithm
Multi-sensor fusion Handles faulty sensors MRI algorithm
Bala Lakshminarayanan, 9th March 11
Data fusion…(3)
Multi sensor fusion(From different
nodes)
Multi-modality fusion(From different
sensing modalities)
Multi-modality fusion(From different
sensing modalities)
Temporal fusion(From different
frames)
Temporal fusion(From different
frames)
Temporal fusion(From different
frames)
Temporal fusion(From different
frames)
GRAYSCALE NODE(Frame1, frame2,
Frame3….frameN)
IR NODE(Frame1, frame2,
Frame3….frameN)
Bala Lakshminarayanan, 9th March 12
BKS Classification…(1)
Behaviour Knowledge Space Aggregates results obtained from individual
classifiers Statistically, gives the optimal result OCR on 46,451 numerals shows BKS
outperforms voting, Bayesian and Dempster-Shafer
These approaches require the independence assumption – not so in real applications
Bala Lakshminarayanan, 9th March 13
BKS Classification…(2)
Independence assumption All classifiers are assumed to be equal Information for fusion is taken from confusion
matrix of single classifier BKS avoids the independence assumption by
concurrently recording decisions of all classifiers
Behaviour of all classifiers recorded on a knowledge space - BKS
Bala Lakshminarayanan, 9th March 14
BKS Classification…(3)
x R n
D R n c: [ , ] { } 0 1 0
M xD ( )
Feature Vector
Classifier
Class Label
Crisp Classifier, Fuzzy classifier, Possibilistic classifier
Decision can he hardened using the maximum membership rule
D x k M x M x iDK
Di( ) ( ) m ax ( ( ))
Bala Lakshminarayanan, 9th March 15
BKS Classification…(4)
Majority voting Class labels are crisp or hardened Crisp label most represented is assigned to x Ties are broken randomly
Bala Lakshminarayanan, 9th March 16
BKS Classification…(5)
BKS s1, s2, …, sL are crisp labels assigned to x by
classifiers D1, D2, …, DL respectively Every combination of labels is an index to an
LUT
Bala Lakshminarayanan, 9th March 17
BKS Classification…(6) Example of BKS c = 3, L = 2, N = 100
s1, s2 Number from each class
Label
1,1 10/3/3 1
1,2 3/0/6 3
1,3 5/4/5 1,3
2,1 0/0/0 0
2,2 1/16/6 2
2,3 4/4/4 1,2,3
3,1 7/2/4 1
3,2 0/2/5 3
3,3 0/0/6 3
Bala Lakshminarayanan, 9th March 18
SFTB Framework Start
Grab frames from datasetContinuous, non continuous
Segment using bgSubtract()Background image
Extract features invMoment()Normalize, write database files
ClassifyreadData(), knn()
End
Inputs-nodeID, scenario…
Bala Lakshminarayanan, 9th March 19
Results…(1)
30 frames from each node – 5 testing, 25 training
7 targets 3 scenarios 2 nodes (IR, Grayscale) 6 databases with 210 feature vectors 2 versions – database1, database2
Bala Lakshminarayanan, 9th March 20
Results…(2)
Node px12, Scenario 1, k=7, Classification accuracy 62.86%
Targets 1 2 3 4 5 6 7
1 4 - - - - - 1
2 - 5 - - - - -
3 - - 4 - - 1 -
4 - - - 5 - - -
5 - - - - 1 - 4
6 - 2 2 - - 1 -
7 3 - - - - - 2
Targets 1 2 3 4 5 6 7
1 3 - 1 - 1 - -
2 - 5 - - - - -
3 1 - 4 - - - -
4 - - 2 2 1 - -
5 - - - - 4 - 1
6 1 1 - - - 3 -
7 - - - - - 2 3
Node in23, Scenario 1, k=7, Classification accuracy 62.86%
Bala Lakshminarayanan, 9th March 21
Results…(3)
Scenario k=7 k=9 k=11
1 62.86 68.57 68.57
6 68.57 68.57 71.43
25 88.57 88.57 80.00
Scenario k=7 k=9 k=11
1 62.86 45.71 51.43
6 57.14 54.28 57.14
25 48.57 48.57 34.28
Node PX12
Node IN23
Bala Lakshminarayanan, 9th March 22
Future work
Perform fusion between IR and grayscale image data
Perform fusion between images from different scenarios
Bala Lakshminarayanan, 9th March 23
References
X. Wang, H. Qi, S. S. Iyengar, “Collaborative multi-modality target classification in distributed sensor networks”, International Conference on Information Fusion (ICIF), July 2002
R. Viswanathan, P.K. Varshney, “Distributed Detection with multiple sensors: Part 1-Fundamentals”, Proceedings of the IEEE, Vol 85, Jan 1997
L.I. Kuncheva, J.C. Bezdek, Robert P.W. Duin, “Decision templates for multiple classifier fusion: an experimental comparison”, Pattern Recognition, Vol 34, 2001
Y.S. Huang, C.Y. Suen, “A method for combining multiple experts for the recognition of unconstrained handwritten numerals”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol 17, Jan 1995