a. hero, afosr muri review 11/08 adaptive radar sensing strategies afosr muri integrated fusion,...
TRANSCRIPT
A. Hero, AFOSR MURI Review 11/08
Adaptive radar sensing strategies
AFOSR MURI
Integrated fusion, performance prediction, and sensor management for ATE
(PI: R. Moses)
A.Hero
Univ. of Michigan – Ann Arbor
2nd Year Review, AFRL,11/08
A. Hero, AFOSR MURI Review 11/08
Outline• I. Broad aims of our research
• II. Progress in sensor management
– New effort: multiple platform radar provisioning with guaranteed uncertainty
– Continuing effort: sparsity constrained spatio-temporal search using convex resource allocation criteria
• III. Progress in front-end processing and fusion
– New: graphical models for distributed decomposable PCA
– New: graphical models for hyperspectral image unmixing
• Information items
– Synergistic Activities
– Personnel
– Publications
A. Hero, AFOSR MURI Review 11/08
I. Broad aims of our research• Integration of modeling, inference, planning
– Integration of multi-platform data
– Performance prediction
– Information-directed sensor management
• Constraints
– Limited time/energy/resources
– Brute force optimal approaches are intractible
• Components of our research approach
– Sequential resource allocation
• Multiresolution wide area search
• Multiple platform provisioning
– High level fusion with hierarchical graphical models
• Decomposable PCA
• Hyperspectral unmixing
– Performance prediction
• Guaranteeduncertainty management
• Bayesian posterior analysis
This talk
Agile Multi-Static Radar system illustration
A. Hero, AFOSR MURI Review 11/08
Part II: Sensor Management
• II.A Performance prediction: multitarget
multiplatform multifunction radar systems
• II.B Adaptive wide area search: sparsity-
constrained multiresolution radar search
A. Hero, AFOSR MURI Review 11/08
II.A Performance prediction: multitarget multiplatform multifunction radar systems
High confidence target regions
time=t time=t+targets targets
A. Hero, AFOSR MURI Review 11/08
Target track update
time=t time=t+
High confidence target regions
Trac
k up
date
A. Hero, AFOSR MURI Review 11/08
Wide area search
time=t time=t+
High confidence target regions
Wide area search
A. Hero, AFOSR MURI Review 11/08
Objective: performance prediction
• Radar constraints: – multipulse radar can be allocated to multiple
tasks: target tracking, wide area search,...– number of radar pulses affect MSTE/ROC and
time spent on a given task
• Objective: predict overall system capabilities – maximum number of targets that can be
reliably tracked with a given number of radars?
– system loading and load margin available for other tasks (discrimination, kill assessment, search)?
A. Hero, AFOSR MURI Review 11/08
Our approach • A guaranteed uncertainty management (GUM)
framework– Radar system performance prediction– Guarantee specified level of track/detection
accuracy (std error of 2%, 5% FA and 1% M)– Specify stable regime of system operation
• An combination of information theoretic uncertainty management and prioritized longest queue (PLQ) resource allocation– related to optimal multiprocessor policy of
Wasserman&etal:2006 for multi-queueing systems.
A. Hero, AFOSR MURI Review 11/08
Uncertainty management and PLQ
ServiceLoad
TargetUncertainty
Policy is analogous to optimal processor allocation in heterogeneous multiple queueingsystems (Wasserman&etal:2006)
A. Hero, AFOSR MURI Review 11/08
PLQ Stability Analysis• Radar load for nth target after secs ellapsed
• As radar load grows superlinearly in time system stability is the central issue
• Cumulative service time to revisit all N targets
A. Hero, AFOSR MURI Review 11/08
Track-only stability condition• For stable operation of radar system
where (balance equation)
• Track-only system capacity: = maximum number of targets for
which solution exists
A. Hero, AFOSR MURI Review 11/08
Multi-tasking stability: load margin
• Assuming radar operates below capacity headroom exists for other tasks.– Search load:– Discrimination load:
• Condition for stability with additional load
• Excess capacity and occupancy
A. Hero, AFOSR MURI Review 11/08
Illustration: 24 Swerling II targets
• C-band radar (4Mhz)• PRI=1ms (150km)• Range res=150m• # pulses=10
– (Pf,Pd) = (0.000001, 0.9999)
• Target speed=300m/s• Speed std error=30m/s• Direction std error=18deg
Load curve lies above diagonalMax number of trackable targets is 23
• System is underprovisioned• Stable track maintenance impossible
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80
0.2
0.4
0.6
0.8
1
1.2
1.4
Single target service time (secs (PRI
Np)-1)
Tar
ge
t re
visi
t pe
riod
(se
cs)
N=24, R=1, RevisitR=0, =0%, Nmax
/R=23
Tcell
Radar load curveStability lineOptimal load curve
A. Hero, AFOSR MURI Review 11/08
Illustration: 12 Swerling II targets
• C-band radar (4Mhz)• PRI=1ms (150km)• Range res=150m• # pulses=10
– (Pf,Pd) = (0.000001, 0.9999)
• Target speed=300m/s• Speed std error=30m/s• Direction std error=18deg
•Track-only load curve below diagonal•Can handle up to 23 targets •With12 targets extra 0.2 secs to spare • System has excess capacity
• Load margin is 0.176 and occupancy is 70%
A. Hero, AFOSR MURI Review 11/08
Discussion• Take home message: GUM performance prediction framework specifies
capacity and stability of radar systems with information theoretic performance measures
• Theory can be used to evaluate radar systems for given scenario• The system capacity and stability depend on the presribed maximum track
and detect uncertainty • Priority longest queue (PLQ) allocation policy is a natural but not the only
radar resource allocation policy that can be studied in this framework.
A. Hero, AFOSR MURI Review 11/08
II.B. Adaptive wide area searchSt
age
1W
ide
area
sea
rch
Stag
e 3
Ref
ined
sea
rch
Stag
e 2
Ref
ined
sea
rch
Stag
e 4
Ref
ined
sea
rch
A. Hero, AFOSR MURI Review 11/08
Problem setup (same slide as last year)
• Set of all cells • ROI• ROI indicator• Spatio-temporal energy allocation policy
• Observations • Uniform spatial allocation:• Ideal spatial allocation:• Optimal N-step allocation: multistage stochastic control problem
• Simpler objective: find two-step optimal allocation that minimizes
A. Hero, AFOSR MURI Review 11/08
Recall optimal strategy
A. Hero, AFOSR MURI Review 11/08
Recall comparisons
Wide area SARacquisition
Optimal two step SARacquisition
Overall energy allocated is identical in both cases
A. Hero, AFOSR MURI Review 11/08
Year 2 progress on ARAP: M-ARAP
• Extend ARAP to account for
– time constraints (number of chips acquired)
– radar beam shape (footprint)
– extended targets
– multi-resolution search implementation
• Modified measurement model incorporates
spatial point spread function H(t)
A. Hero, AFOSR MURI Review 11/08
Simulation of M-ARAP for MTI • Uniformly attenuating beampattern
• FOV is 66 x 66 km with pixel dimensions of 20 × 20 m
• Radar resolution cell is 100 × 100 × 150 m.
• Sparsity level p = 0.0007 was selected Q = 4082
• Identical targets with target reflection distribution modeling an
aircraft similar to an Airbus A-320.
A. Hero, AFOSR MURI Review 11/08
Simulation of M-ARAP for MTI • Target velocities are isotropically normally distributed
• Swerling II noise model
• Clutter (rain) intensity was random between 0-6 [mm/hr] and
spatial correlation on the order of 1x1 km.
• Maximal clutter velocity was 30 m/sec
• Standard single pass MTI filter is compared to a two pass
multiresolution ARAP search
• M-ARAP search has lower MSE localization error, fewer false
alarms and higher detection rate than MTI for equivalent time
and energy.
A. Hero, AFOSR MURI Review 11/08
M-ARAP for MTI tracking radar
A. Hero, AFOSR MURI Review 11/08
Correct detection probability vs
false discovery rate
A. Hero, AFOSR MURI Review 11/08
Optimal energy allocation
A. Hero, AFOSR MURI Review 11/08
Discussion• Take home message: can attain 7dB MSE reduction at SNR of 5 dB using
only N=Q/P samples• M-ARAP searches for P sparsely distributed but clustered targets over Q
search cells with minimum time and energy constriants• Objective function J is related to the KL information divergence and the
Fisher information under a Gaussian measurement model. • J only depends on the cumulative energy allocated to each voxel in the
image volume (deferred reward)• Features of two-step M-ARAP search algorithm
– motivated by pooled statistical sampling (syphylis studies of Dorfman:AnnMathStat1943)
– assigns energy to regions with high posterior probability of containing targets– is an index policy with threshold k0– is a multi-resolution extension of the two-stage ARAP search algorithm presented
at last review.– Is low computational complexity - O(Q)
A. Hero, AFOSR MURI Review 11/08
Part III: High level fusion
• III.A Distributed decomposable PCA
• III.B Hyperspectral imaging and unmixing
• Common theme: application of hierarchical
graphical models
A. Hero, AFOSR MURI Review 11/08
III.A Decomposable PCA• Principle components analysis (PCA) is a model-free
dimensionality reduction technique used for high level
data fusion (variable importance, regression, variable
selection)
• Deficiencies:
– PCA does not naturally incorporate priors on
• Dependency structure (graphical model)
• Matrix patterning (decomposability)
• Scalability problem: complexity is O(N^3)
– Unreliable/unimplementable for high dimensional data
– Ill-suited for distributed implementation, e.g., in sensor
networks
A. Hero, AFOSR MURI Review 11/08
Networked PCA• Network model: measure sensor outputs Xa, Xb, Xc
– Two cliques {a,c} and {b,c}
– Separator {c}
• Decomposable model: covariance matrix R unknown but
conditional independence structure is known.
• PCA of covariance matrix R finds linear combinations
y=UTX that have maximum or minimum variance
A. Hero, AFOSR MURI Review 11/08
DPCA formulation• Precision matrix K=R-1
• For decomposable model K has structure
• General representation
A. Hero, AFOSR MURI Review 11/08
1 dimensional DPCA• PCA for minimum eigenvector/eigenvalue solves
• Key observation:
• This constraint is equivalent to
• where
A. Hero, AFOSR MURI Review 11/08
Extension to k-dimensional DPCA• k-dimensional PCA solves sequence of eigenvalue
problems
• Dual optimization
A. Hero, AFOSR MURI Review 11/08
k-dimensional DPCA (ctd)
• Dual maximization splits into local minimization with
message passing
• Message passing
A. Hero, AFOSR MURI Review 11/08
Tracking illustration of DPCA • Scenario: Network with 305 nodes representing three
fully connected networks with only 5 coupling nodes
• C1 = {1, · · · , 100, 301, · · · , 305}, C2 = {101, · · · ,
200, 301, · · · , 305}, and C3 = {201, · · · , 300, 301, ·
· · , 305}.
• Local MLEs computed over sliding time windows of
length n = 500 with 400 samples overlap.
• Centralized PCA computation: EVD O(305)^3 flops
• DPCA computation: EVD O(105)^3 flops + message
passing of a 5x5 matrix M
A. Hero, AFOSR MURI Review 11/08
DPCA min-eigenvalue tracker
Iteration 1
Iteration 2
Iteration 3
A. Hero, AFOSR MURI Review 11/08
DPCA network anomaly detection
SNVA
STTL
LOSA KSCY
HSTN
DNVR
CHIN
IPLS
ATLA
WASH
NYCM
Multiple measurement sites (Abilene)
A. Hero, AFOSR MURI Review 11/08
DPCA anomaly detection
PCA (centralized)
DPCA (E-W decomp)
DPCA (E-W-S decomp)
DPCA (Random decomp)
A. Hero, AFOSR MURI Review 11/08
Discussion• Take home message: Combination of model-free dimensionality reduction
and model-based graphical model can significantly reduce computational complexity of PCA-based high-level fusion
• Complexity scales polynomially in clique size not in overall size of problem. Example: 100,000 variables with 500 cliques each of size 200
– Centralized PCA: complexity is of order 1015
– DPCA: complexity is of order 106
• If can impose similar decomposability constraints on graph Laplacian matrix, be extended to non-linear dimensionality reduction: ISOMAP, Laplacian eigenmaps, dwMDS.
A. Hero, AFOSR MURI Review 11/08
III.b Hyperspectral unmixing• Hyperspectral imaging model
Y = MA + N
• Y: L x P matrix over L spectral bands and P pixels
• M: L x R matrix of R endmember spectra
• A: R x P matrix of endmember mixture coefficients
• N: L x P noise residual matrix
• Hyperspectral unmixing problem is to estimate A given M
and Y. Usually broken into two steps (ENDFINDR, VCA)
– Endmember extraction algorithm (EEA)
– Inversion step to extract mixing coefficents
A. Hero, AFOSR MURI Review 11/08
Sample endmember spectra
• Concrete Redbrick
A. Hero, AFOSR MURI Review 11/08
Key observation• Mixing coefficients supported on R-1 dimensional simplex
• This suggests a natural dimensionality reduction approach
to estimation of mixing coefficient matrix A
A. Hero, AFOSR MURI Review 11/08
Hierarchical Bayesian model• Graphical model structure induces posterior
• T: projected endmember spectra
• C: projected mixing coefficents
• e,s: mean and variance of projected endmember
spectrum
A. Hero, AFOSR MURI Review 11/08
Moffet field hyperspectral image
AVIRIS image
A. Hero, AFOSR MURI Review 11/08
Unmixing results
A. Hero, AFOSR MURI Review 11/08
Unmixing results
A. Hero, AFOSR MURI Review 11/08
Discussion• Take home message: by using a unified graphical
model approach to hyperspectral unmixing can
significantly improve performance wrt state-of-the-art
(N-FINDR, VCR)
• Other Bayesian prirors can lead to sparsity
preserving solutions
A. Hero, AFOSR MURI Review 11/08
Synergistic Activities• Related funded activities
– NSF (Cozzens) Transductive anomaly detection– ARO (Harmon) Sparsity penalized 3D inverse scattering– ARO (Prater) Sparsity penalized 3D molecular imaging MRFM– ONR (Martinez) Network tomography and discovery
• DoD panel participant:
– National Research Council Workshop on Disrupting IED Terror Campaigns and Predicting IED Activities (Mar. 2008)
– Army Research Office CISD Strategic Planning Meeting (Aug 2008)
– NSF/IARPA/NSA workshop on the science of security (Nov 2008)
• Industry interactions
– Techfinity (MDA funded) Guaranteed uncertainty management for missile defence
– SIG (ATR Center dunded) information-driven dimensionality reduction
A. Hero, AFOSR MURI Review 11/08
Personnel
• Supported by MURI grant:– 2008- : G. Newstadt (2nd year MS) – 2007-2008: C. Kim (2nd year MS)– 2007-2008: E. Bashan (Graduated July 2008)
• Other (unsupported by UM)– Ami Wiesel (Post-doc, Umichigan)– N. Dobigeon (Univ of Toulouse)– S. Damelin (Prof at Georgia Southern)– Venkat Chandrasekeran (MIT Grad Student)
A. Hero, AFOSR MURI Review 11/08
Publications (2007-2008)• Appeared
– E. Bashan, R. Raich, R.; A.O. Hero, “Optimal two-stage search for sparse targets using convex criteria,” . IEEE Trans. on
Signal Processing Vol. ?, Issue 10, Oct. 2008 Page(s):?. – E. Bashan, “Efficient resource allocation schemes for search ,”
PhD Thesis, University of Michigan, May 2008.– H. Bagci, R. Raich, A. E. Hero, and E. Michielssen, "Sparsity-
Regularized Born Iterations for Electromagnetic Inverse Scattering," Proc. of IEEE Antennas and Propagation Symposium, June, 2008.
– A. Hero, Guaranteed uncertainty management (GUM) for sensor provisioning in missile defense, mid-term research report to the US Missile Defense Agency and Techfinity, Inc, Mar 2008.
• Submitted– N. Dobigeon, J.-Y. Tourneret, S. Massaoui, M. Coulon and A.O.
Hero, 'Joint Bayesian endmember extraction and linear unmixing for hyperspectral imagery,' IEEE Trans. on Signal Pocessing, submitted Sept 2008.
– A. Wiesel and A.O. Hero, 'Decomposable Principal Components Analysis,' IEEE Trans. on Signal Processing. submitted Aug 2008.