Title: Sequential%20Adaptive%20Multi-Modality%20Target%20Detec-tion%20and%20Classification%20using%20Physics-Based%20Models
1Sequential Adaptive Multi-Modality Target
Detec-tion and Classification using Physics-Based
Models
- Professor Andrew E. Yagle (PI) (EECS)
- Mine detection, channel identification
- Professor Alfred O. Hero III (EECS)
- Sensor scheduling, nonparametric statistical
models - Professor Kamal Sarabandi (Director, RadLab)
- Vehicle and foliage physics-based modeling
2SEQUENTIAL ADAPTIVE MULTI-MODALITY TARGET
DETEC- TION AND CLASSIFICATION USING
PHYSICS-BASED MODELS
- APPROACH
- Develop realistic physics-based models
- Perform statistical simulations to obtain
distributions of measured scattered fields - Develop sensor scheduling and detection
algorithms using these statistical models - Evaluate algorithms using statistical measures
- Apply algorithms to real multi-modal data.
- ARMY COLLABORATIONS
- Army Night Vision Lab (GPR IR mine field data).
(PICTURE)
- ACCOMPLISHMENTS
- Phenomenological studies of radar clutter and
targetclutter using realistic physics-based
models - Developed non-parametric MRF models for these
- Developed myopic sequential adaptive sensor
management algorithm for tracking problems - Developed migration (time-reversal) algorithm
for imaging land mines and evaluated on real GPR
data. - TRANSITION TO ARMY/ INDUSTRY
- In progress.
- OBJECTIVES
- Develop algorithms for detection of landmines
and tanks under trees using radar and IR
sensors - Develop data-adaptive algorithms for sensor
scheduling and multi-modal sequential detection - Evaluate the algorithms using Monte Carlo type
simulations on realistic models, and on real
data. - ARMY RELEVANCE
- Detection of landmines and tanks under trees
has obvious Army relevance
3(No Transcript)
4SM
Reduced Models
Reduced Models
Scenarios/Sensor models
Scenarios/Sensor models
Actual Data
Simulated Data
Performance Matrix
Hybrid simulated/real data
TUT
UXO
Environment Modification
5(No Transcript)
6Research Project Objectives
- Develop overall algorithm for detection of
Tanks under trees landmines (not structures). - Initial focus TUT (can hit the ground running).
- Algorithm Features sequential detection, sensor
management selection, physics-based models - Evaluate the resulting procedure on realistic
models (statistical simulations) and real data.
7Issues Overall Algorithm
- How to select which sensing modalities to use?
- What is the value-added for combining other
modalities? Is it worth the additional cost? - How to implement data-adaptive configurations,
e.g., selection of sources/receivers, based on
scattering of targets and propagation in medium? - How to select decision thresholds for detection?
- What are the figures of merit for evaluation?
8Issues Overall Algorithm II
- ALL of these issues require that we develop
- statistical models for scattered fields from
vehicles under foliage and from mines. - Development of multimodal target detection
algorithms take time we need to perform Monte
Carlo simulations using realistic physics-based
models.
9Sequential Adaptive Multi-Modality Target
Detection and Classification using Physics-Based
Models
- Professor Andrew E. Yagle
- RAs Jay Marble, Siddharth Shah
- Professor Alfred O. Hero III
- RAs Doron Blatt, Chris Kreucher, Raghuram
Rangarajan - Postdoc Cyrille Hory
- Professor Kamal Sarabandi
- RAs Mojtaba Dehmollaian, Feinian Wang
- Research Scientists Leland Pierce, Il-Suek
Koh
10Sequential Adaptive Multi-Modality Target
Detection and Classification using Physics-Based
Models
- Mine detection Yagle, J. Marble
- Blind Channel Deconvolution Yagle, S. Shah
- Vehicle modeling Sarabandi, M. Dehmollaian
- Foliage modeling Sarabandi, I. Koh, F. Wang
- Sensor scheduling Hero, Kreucher, Blatt
- Nonparametric statistics Hero, Blatt, Ragarajan
11HERO 1st-Year Accomplishments I
- Developed non-parametric statistical modelling of
scattered fields using Markov random fields. - Significance Can model scattered fields from
a few - observations, extrapolating the rest. This
saves much time in Monte Carlo statistical model
development. - Developed target model reduction technique.
- Significance Can model vehicles using a
lower dimensional manifold, simplifying detection.
12HERO 1st-Year Accomplishments II
- Developed myopic distributed multi-sensor
multi-look detection and tracking sensor
management algorithms using Renyi divergence. - Significance Optimal sensor scheduling too
- hard when many targets and sensors present.
- Particle filtering Renyi-divergence-based
scheduling reduces complexity. Tracking of - dozens of actual targets was demonstrated.
13HERO Progress since August
- Developing non-myopic distributed multi-sensor
multi-look detection and tracking sensor
management algorithms using Renyi divergence and
Q-learning. - Significance Can develop useful optimal SM
approximations and quantify performance-vs-complex
ity tradeoffs. -
- Developing aggregation strategies for distributed
sensors and quantifying performance tradeoffs. - Significance Allows fast and reliable
computation of maxima of objective functions
these dictate strategies.
14Sarabandi 1st-Year Accomplishments I
- Performed phenomenological studies of
- (a) physics-based clutter models
- (b) physics-based target models
- Significance Basic understanding of
effects is vital for interpreting results. - These proved very useful in developing the
statistical models of scattered fields.
15Sarabandi 1st-Year Accomplishments II
- New results Target-clutter interaction Multiple
scattering from needle clusters - Closed-form solution for scattering from a
disk of arbitrary shape. - Developed time-reversal method for foliage
camouflaged target detection. - Significance This is one of the
physics-based models for detection (project
title).
16Sarabandi Progress since August
- Developing iterative frequency-correlation based
forest radar channel identification - New approach for attenuation estimation.
- Significance Procedure for deconvolving
effects of propagation through foliage. - Developing iterative physical optics approach to
account for foliage shadowing. - Significance Greatly reduces computation
17YAGLE 1st-Year Accomplishments
- Developed mine detection algorithm from SAR GPR
using range migration imaging (with Jay Marble). - Significance Physics-based algorithm for
imaging mines from ground-penetrating radar
(project title). - Developed 2D and 3D blind deconvolution
algorithms for radar channel identification (with
Siddharth Shah). - Significance Apply to blind deconvolution of
channel propagation effects for mines, and
perhaps for foliage.
18YAGLE Progress since August
- Developed hyperbola-flattening transform
algorithm for feature detection in GPR data. - Significance Preliminary detection stage
using less computation than range-migration
imaging - Working on material discrimination using decay
rates from magnetometer (metal detector) data. - Significance Multi-modal mine detection.
19Synergistic Activities Hero
- General Dynamics (formerly Veridian, ERIM)
- C. Kreucher sensor management scheduling
- K. Kastella sensor management
- J. Ackenhusen mine detection
- ARL NAS-SED review panel member
- ARL N. Patwari (student) summer internship
- ERIM Int C. Shih (student) summer internship
20Synergistic Activities Sarabandi
General Dynamics John Ackenhusen BAE Norm
Byer FCS COMMUNICATIONS Jim Freibersiser
(DARPA PM) Barry Perlman (CECOM) ARL Ed Burke
(mm wave), Brian Sadler, Bruce Wallace
21Synergistic Activities Yagle
- General Dynamics (formerly Veridian, ERIM)
- Jay Marble, student (ARO mine research)
- Brian Fischer, student (Low RCS material design)
- Chris Wackerman, former Ph.D. student
22Summary of Results Hero
Quick Overview
- Statistical distributions and realizations of
backscatter from a plate in a pine forest, from
Sarabandis physics-based models. - Aggregation (centralization) of sensors data for
detection and estimation - 3. Sequential adaptive sensor management using
non-myopic strategies.
23Research Loci(2003)
- Statistical modeling of forward- and back-
scatter fields - Polarimetric Field Modeling and Reconstruction
(Hory/Blatt) - Adaptive multicomponent Pearson model
- Markov random field (MRF) model for
extrapolation/reconstruction - Adaptive decentralized detection and
classification - Aggregation strategies for distributed sensors
(Blatt) - Optimal estimator aggregation method developed
- Quantified and exploited tradeoff between local
on-board processing and centralized aggregation - Sequential adaptive sensor management
- Non-myopic multi-modality sensor
scheduling(Blatt/Kreucher) - Information-driven non-linear target tracking
algorithms - Markov decision process (MDP) for detecting smart
targets
24Detection Target or Clutter Alone?
25Experiment Plate in Forest of Pine Trees
Randomized tree positions
Trees
Plate
meters
meters
- 15cm x 15cm x 1cm plate at 1m from ground
- Plate under forest canopy (10 pine trees)
26Multi-Static Radar Platform
60
Statistical analysis
60
Illumination/detection radar array
27Backscatter realizations
Forest Alone Target in Forest
28Forest Alone and Target plus Forest Histograms
SNR0dB
SNR6dB
Additive and Gaussian
Not Additive or Gaussian
- 2GHz Spotlight SAR illumination
- Aggregate of three look angles (azimuth35,45,55,
elev180) -
29Target and Clutter Return Spatial Distribution
- I.i.d. forest-alone return over array
- Target introduces local spatial dependency
- Spatial dependency decays exponentially fast
- Dependency model required to capture presence of
target
30Q-Q Plot for Gaussianity Testing
Return from forest alone Return from
plate in forest
KS goodness-of-fit P value0.0303
KS goodness-of-fit P value0.9985
31Target in Forest Marginal Backscatter Density
Non-parametric estimates via multi-component
Gaussian mixtures
- Results for four corner sub-arrays
- Gaussian mixture components estimated by ML-EM
algorithm - Number of components adaptively estimated via
(MML) penalty
32Mixture models Goodness of Fit
33Joint Finite Mixture Models For Spatial Dependency
Non-parametric joint density estimation using
Gaussian mixtures - neighboring detector cells
- Advantages of Adaptive Mixture Modeling
- Better fit to the empirical statistics than
previous models - Simple ML estimator
- Well suited for GLRT
- Power of parametric and flexibility of
non-parametric approaches
34Centralized Approach to Estimation and Detection
Processing Unit
35Decentralized Approach to Estimation and Detection
Processing Unit
36A slice of ambiguity/likelihood function
Estimator Realizations
x xx x xx xxxx x xx
37Optimal Aggregation Algorithm
Sample Covariance Analysis
Estimation of Gaussian Mixture Parameters (EM
)
Aggregation To Final Estimate
38Distributed Sensing Performance Comparisons
39Implications
- 50 of the local estimates were aberrant
- With gt 10 sensors centralized estimates attain
the CRB - Naïve aggregation is 10dB worse than CRB
- Smart (clairvoyant) aggregation comes within 3dB
of CRB - New method is within 3.5dB of CRB
40Sequential Adaptive Sensor Management
- Sequential only one sensor deployed at a time
- Adaptive next sensor selection based on present
and past measurements - Multi-modality sensor modes can be switched at
each time - Detection/Classification/Tracking task is to
minimize decision error - Centralized decision making sensor has access to
entire set of previous measurements - Smart targets may hide from active sensor
Single-target state vector
41Sequential Adaptive Sensor Management
- Progress made on two fronts
- Non-myopic information-gain strategies for target
tracking - Value function approximation using visibility
constraints - Renyi-Divergence approximation
- Established link between Renyi info and decision
error exponents - Mitigated computational bottleneck by adaptive PF
- Coupled vs independent particle partitions for
tracking multiple targets - Exploitation of permutation symmetry
- Real-time operation demonstrated for tracking gt
10 real target motions - Partially Observed Markov Decision Process
strategies - Developed Q-learning approach to sensor
management - Applied to detecting smart targets
42Sensor scheduling objective function
- Action a deploy a sensor, probe a cell at time t
- Value of taking action a at time t after
observing
Sensor agility
Prediction
Retrospective value of taking action a
Available measurements at time t-1
43In Retrospect Posterior Density
x
x
Best action is a2 since its posterior update is
most concentrated ? induces highest information
gain
44Information-based Value Function
- Incremental information gained from taking action
a at time t can be measured by divergence - Requires updating posterior distributions of
future target state given future Z and given
present Z, resp., - Main issues for evaluation of ED(a,t)Z
- Computation complexity
- Robustness to model mismatch
- Decision making relevance
45Information Value FunctionAlpha Divergence
- Properties of Renyi divergence
- Simpler and more stably implementable than KL
(a1) (KreucheretalTSP04, SPIE03) - Parameter alpha can be adapted to non-Gaussian
posteriors - More robust to mis-specified models than KL
(KreucheretalTSP04, SPIE03) - Related directly to decision error probability
via Sanov (HeroetalSPM02) - Information theoretic interpretation
46Myopic Target Tracking Application
- Possible actions point radar at cell c and take
measurement, c1, , L - We illustrate the benefit of info-gain SM with AP
implementation of JMPD tracking 10 actual moving
target positions (2001 NTC exercise). - GMTI radar simulated Rayleigh target/clutter
statistics - Contrast to a periodic (non-managed) scan same
statistics - Coverage of managed and non-managed50 dwells per
second
47Comparison with Other Myopic Managed Strategies
- Renyi-Divergence method of sensor management
outperforms others - Periodic scan sweeps through all cells and then
repeats - Methods A and B point the sensor where
targets are estimated to be - Method A chooses cells randomly from cells
predicted to have targets and cells surrounding
those predicted to have targets - Method B chooses cells probabilistically based
on their estimated target count
48Multimode Radar Mode and Dwell Point
Selection(Myopic Sensor Management)
- The information based SM algorithm applies to a
sensor with multiple modalities. - Sensors make a total of L sensing actions each
(here L16) - For each mode SM determines which action
generates max. expected gain in information. - The mode corresponding to the largest expected
information gain is chosen and used - To accommodate switching times (between modes)
and different time scales required for the modes
expected Information / time (information rate).
MTI 100mx100m cells (GMTI) measures 49x1
strips Pd.5, Pf 1e-4 Detects moving targets
only FTI 100mx100m cells (SAR) measures 7x7
blocks Pd.5, Pf 1e-4 Detects stopped targets
only ID 100mx100m cells (HRR) measures 3x3
blocks Confusion matrix
True1 True2 True3 Empty Meas 1 0.600
0.200 0.200 0.333 Meas 2 0.200
0.600 0.200 0.333 Meas 3 0.200
0.200 0.600 0.333
49Toward Real-Time Operation
- Incorporation of advanced adaptive multitarget
sampling schemes allows tracking with small
numbers (a few hundred to a thousand) of
particles for tens of targets. - Modules for particle proposal, weighting, and
divergence expectations written natively - Simulation
- Multitarget particle filter tracker using
a-divergence sensor management (a.5). - Real targets taken from battle simulations at
NTC. - Number of sensor dwells per time scaled up with
number of targets.
50There is a Performance Loss Associated with
Myopic SM
- Myopic SM computes only one-step ahead
- Does not incorporate any information past one
step ahead, even information that may be known
perfectly - Vulnerable to situations which require planning
ahead - Sensor-to-target visibility changing due to
platform or target motion - Detection characteristics of target changing over
time - Non-myopic SM looks ahead multiple steps
- Computationally difficult to implement exactly
approximate methods necessary - Even two step look-ahead can be of value
51Non-myopic sensor management Relevant Situations
Sensor Position
Sensor Position
Visible Target
Region of Interest
Region of Interest
Shadowed Target
Extra dwells useful at time 1 not made by myopic
strategy
Time 1
Time 3
Time 4
Time 5
Time 6
52Non-myopic scheme makes use of this information
Myopic scheme uses only this information
Left Target Measured
One Realization of p(X2Z1) when left target
measured
Right Target Measured
Posterior at t0, P(X0Z0)
Prediction at t1, P(X1Z0)
- Simple illustration with Non-myopic information
gain criterion - Two targets in two cells
- At even time instants only one cell is visible
One Realization of p(X2Z1) when right target
measured
532-step Lookahead Non-Myopic Search Tree
54Comparison of Greedy and Non-Myopic (2 step)
decision making
Myopic Target lost 22 of the time
Non-Myopic Target lost 11 of the time
55General Non-Myopic Strategies
- Reward at time t for action sequence
- is
information state - Optimal action sequence
- Optimal action sequence satisfies Bellmans
equation - Value function
56Optimal Action Determined by Partition of
information state space
1
1
0
Special case of 3 state target
57Application to Optimal Sensor Management
- For discrete measurements and finite horizon (T),
solution to value equation is linear program - Krishnamurthy (2002) exploited this property for
SM - Problems with Krishnamurtys approach
- Complexity of linear program is geometric in T
- when number of states is large computations
become intractable - when measurements are continuous value equation
is non-linear
58 ? time t-1 ?
time t ? time t1 ?
- Impose simple form on scheduling function
infinite horizon - Time invariant function of information state
- Value Function Approximation
-
Approaches
- Exploring depth with particle proposals
- Optimal allocation of N particles
59Exploration with Particle Proposals
Model Update
Realized Information Gain k1 to k2
Model Update
Expected Information Gain, k2
. . .
ltDagt
ltDagt
60Non-Myopic Value Approximation
The Bellman equation describes the value of an
action in terms of the immediate (myopic) benefit
and the long-term (non-myopic) benefit.
Bellman equation
Non-myopic correction under a
Myopic part of V under action a
Value of state
For computational tractability approximate
non-myopic term Where Na(s) is an easily
computed measure of the future benefit of action
a (i.e. an approximate long-term value term).
61Target Tracking Application Visibility
Constraints
- Define visibility of cell c at time k as
Visk(c) - Visk(c)0 implies cell not visible and Visk(c)1
implies cell is perfectly visible. - A non-myopic strategy will place extra priority
on measuring a visible cell that will soon become
obscured to the sensor. - A candidate non-myopic approximation is to
optimize
Nc( )
Myopic Scheduling (.31s) Brute-force
non-myopic (102.5s) Non-myopic approx.
(.32s)
62Target Tracking Application Information
Divergence
- Let denote the expected myopic gain
when taking action c at time k - denote the distribution of
myopic gain when taking action c at time k - Approximate long-term value of taking action c
- Optimization becomes
- Gaussian approximation to
63Model Problem using Value function approximation
- At initialization, target is localized to a 300m
x 500m region. - GMTI Sensor must search the region for the
target. - Sensor visibility region changes with time.
- Non-myopic strategy scans regions that will be
obscured in the future while defering regions
that will be visible in the future.
64Q-learning Approach to SM
- Our results extend Krishnamurthys work
- Handles continuous measurement space
- Computational complexity is linear in T
- Applicable to infinite horizon, e.g. quickest
detection - Smart Targets state transition matrix affected
by action a - Two principal ingredients
- Using Monte Carlo simulation to approximate
expectation integrals - Performing dimension reduction of information
state s via function approximation
65Q-learning Background
- Main idea (Watkins89)
- simulate actions and the induced information
states (measurements) - Find the optimal schedules by stochastic
averaging - Q-function defined as indexed value function
- Algorithm For n1,2,
- Using
simulate trajectory - Update Q functions according to recursion
- Repeat until variance of Q-function is below
tolerance
66Example SM for Smart Target Detection
- Three possible target states
- No target present (static)
- Target present and exposed at time t
- Target present and hidden at time t
- Four possible actions at time t
- Stop and declare target present or absent
(stopping time is tT) - Defer decision and deploy strong active sensor
- Defer decision and deploy weak active sensor
- Defer decision and deploy passive sensor
- If deploy active sensor target may go into hide
mode. - Goal Deploy sensors so as to minimize time to
correctly decide target present or absence.
67Estimated Q-functions
- Q(s,a) measures value of taking action a at
information state s. - Three sensors available
- A1 strong active
- A2 weak active
- Pa passive
- Qs learned from 1M simulated trajectories of
sensor deployments.
68The resulting policy defined on information space
P(Exposed targetY)
P(Target absentY)
69Gain over myopic strategy drop passive
Detection gain relative to myopic policy never
use passive
70Foci for 2004
- Backscatter models for adaptive detection and
classification refining sensor performance
metrics (Pf, Pd, Pid). - Adaptive non-myopic sensor scheduling and
management combining Q-learning and particle
filtering - Time reversal 3D imaging with uncalibrated sensor
arrays -
71Summary of Results Yagle
Quick Overview
- Mine detection using ground-penetrating radar
(GPR) and range-migration imaging. - Hyperbola-flattening transform for feature
detection from GPR mine field data. - Active magnetometer (metal detector) for
multimodal mine detection. - OMITTED 3-D blind deconvolution
- Basis-function-based inverse scattering
72Quick Overview
The Mine Hunter / Killer
Metal Detector Coils
GPR Antenna 19
GPS Antenna
IR Camera
GPR Antenna 0
73Quick Overview
Wavenumber Migration Applied to GPR Data
Applied SAR imaging algorithm to GPR data. Able
to estimate size and depth of landmine.
Depth 6 6.4
Height 6 8
Width 13 14
TM-62M Russian Landmine
Actual Estimated
Thresholded
Imaged Data
Original Data
6
14
8
74Material Discrimination Using Decay Rates
Aluminum and Iron objects can be separated by
their different decay rates.
Pulsed Metal Detector
Iron Sphere
Aluminum Plate
Double Click to Run Movie
DARPA Backgrounds Dataset (1995)
75Hyperbola Flattening Transform Algorithm
A novel feature for detecting hyperbolic
signatures in GPR data. The Hyperbola Flattening
Transform converts the entire signature into a
point.
Original Hyperbola
Remapping y -gt 1/y
45 Rotation
Radon Transform
Simulation
Simulation
Simulation
Final Location
45 Rotation
Original Hyperbola
Remapping y -gt 1/y
Radon Transform
Final Location
761st Year Accomplishments
Wavenumber Migration Applied to GPR Data A SAR
image formation algorithm was applied to GPR
data. The end result was a repeatable estimate
of the size and depth of the landmine. This
info is very useful in eliminating false-alarms
from GPR data based on the known size and
typical burial depths of landmines. Small
clutter objects near the surface should be
especially easy to eliminate given an estimate
of the size and depth. Decay Rates of Metal
Detector Exploited for Object Discrimination Usin
g a pulsed metal detector, swirling currents can
be induced in metal objects. Theses induced
currents will die away in an exponential
manner. By measuring the rate of decay, certain
metals can be identified. Specifically, iron
and aluminum objects can be easily separated as
currents in iron objects decay more rapidly than
aluminum. Typically, metal landmines are high in
aluminum and low in iron content. Novel
Feature Developed for Hyperbola Detection in GPR
Data It has been shown that the hyperbolic
nature of landmine signatures provides great
discrimination capable over false-alarms from
soil layers and surface returns. A feature can
be computed for discriminating hyperbolic
signatures from non- hyperbolic signatures using
a new transformation called the Hyperbolic
Flattening Transform. This technique transforms
the data from a hyperbola into a single point.
The energy contained in this point becomes the
feature that can be utilized in discrimination.
Winter 03
Fall 03
Summer 03
77Quick Overview
OBJECTIVE
ILLUSTRATION
Determine size and depth of landmines using GPR
as part of a multimodal detection algorithm
APPROACH
ACCOMPLISHMENTS
Range Migration and phase compensation Stoltz
interpolation
Successful detection of Russian mines buried in
field from NVESD MH/K
78RANGE MIGRATION ALGORITHM EXPERIMENT
Quick Overview
USSR TM-62 LAND MINE
Army NVESD MH/K
Point-spread response
Imaging a single point
79RANGE MIGRATION ALGORITHM RESULTS
Quick Overview
TM-62 measured (6 depth)
TM-62 binary reconstructed
80Battlefield Vehicle Prototype
- Army Night Vision Electronic Science
- GPR, metal detector, infrared camera
- Robot arm will mark mine locations with ceramic
disks (arm is not shown at right)
81Ground-Penetrating Radar (GPR)
- Mine Hunter/Killer Designed by BAE
- Army Night Vision Lab (Fort Belvoir VA)
- 20 transmit/receive antenna pairs in front
- 256 frequencies 500 Mhz to about 2 MHz stepped
by 5 MHz
82Significance of Hyperbola
- Avoids false alarms due to clutter and noise
- Stratified ground appears as straight line
- Hyperbola indicates real, localized target
- Hyperbola indicates its depth, as well
83Active Magnetometer Data
Quick Overview
- Work in progress at present time
- Mostly comparing GPR magnetometer
- Multi-modal data GPR magnetometer
- Using previously-developed (Jay Marble)
electromagnetic induction model (1995) - Idea Distinguish aluminum from iron using
induction decay rate (like MRI)
84Quick Overview
Direction Recap The Mine Hunter/ Killer
utilizes 2 up close sensors
(1) array of ground penetration radar
(GPR) antennae (2)
array of metal detector coils, which are also
called
elecromagnetic induction (EMI) coils.
85Quick Overview
Ft. AP Hill - TEST and CAL Lanes
86EMI (6 sensors)
2
3
1
GPR (20 sensors)
1
2
4
3
This is a sample of the data produced by the
system. This is a test lane in Virginia. The
vehicle is moving to the right and the sensor
outputs are vertical. The GPR depths have
been summed into a plan view. The EMI is
showing 3 definite metal objects. The GPR
detects these objects plus a fourth (likely a
shallow, low-metal land mine).
87Physical Model
By creating simple physical models, the hope
is to generate signatures for cross channel
fusion algorithm development, when actual data
from all sensors does not exist for the same
objects.
Expected GPR Signature
This model is the main progress this period.
This FORTRAN model has been compiled
and integrated into MATLAB using mex.
Predicted Vertical Magnetic Field
Depth inches
Target Info Radius 0.07m (3)
Depth 0.1m(6) Conductivity
1000 Coil Height 0.3m(12)
Along Track m
88Measured Data
GPR Measured Signature
Depth
Along Track inches
Along Track samples
These signatures were extracted from the
actual MH/K data. They correspond to Target 1
in the upper left corner of the mine lane.
However, a great deal is not known about the
sensors Gain EMI dipole moment How much
current is exciting the coils? How many loops
does each coil have?
89Raw GPR Signature
Imaged GPR Signature
Depth
Depth
We need a way to measure this value for
the imaging subsection.
Mask For Estimating Size
er 9 (A guess that worked.)
Height 6 Width 13 Depth 6 (to top) Metal
Case
Depth inches
We can get size info by using SAR imaging on
the GPR data.
Russian TM-62M Landmine-
Along Track inches
90TM-62M Russian Landmine
Firing Pen (Always Metal)
91Electromagnetic Induction (EMI)
92Electromagnetic Induction (EMI)
Upper Coil
Lower Coil
- Upper Coil Receive only
- Lower Coil Transmit receive
93EMI Modeling
1
2
- (1) Current driven through coil generates a
primary magnetic field. - (2) Primary field induces magnetic source in
metal object. Induced source can be decomposed
into horizontal and vertical components. - (3) Secondary field produced by horizontal and
vertical induced magnetic sources can be sensed
at surface.
3a
3b
94 EMI Spatial Simulation
953D Metal Detector Data Set DARPA
Backgrounds (1995) Operator Parsons
Engineering System by Geonics
Receivers
Z Coil
Y Coil
X Coil
96Coil Current
Sampled Decay Rate
97Location SB (Ft. Carson, CO) Transmit
Z Receive Y Target Registration Targets
5
Iron Sphere
Aluminum Plate
98- The decay rates of all iron and
- aluminum test objects are shown here.
- The blue objects are aluminum.
- The red objects are iron.
- Decay rates here are for the vertical
- transmit and vertical receive pair.
99- Iron objects decay much faster
- than Aluminum Objects.
100The Hyperbola Flattening Transform
Quick Overview
- Feature detection in GPR data
- Map hyperbolas into spots in feature space.
- Perform 45 degree coordinate rotation.
- Perform reciprocal coordinate transform Maps
rotated hyperbola to a straight line. - Use Radon transform to look for lines
101Hyperbola Flattening Transformation
102Hyperbola Flattening Transformation
Sampled Form
103Hyperbola Flattening Transformation
104Hyperbola Flattening Transformation
105(No Transcript)
106Try this on actual GPR data from mine field
107Try this on actual GPR data from mine field
108Present Work on Landmines
- Issue detection performance post-migration
(easier to look for parallel straight lines) vs.
detection performance w/pre-migration data
(harder to look for hyperbolae, but apply to raw
data before migration processing) - Issue develop statistical physics-based model
- Issue how to combine with other modalities
109Summary of Results Sarabandi
Quick Overview
- Iterative physical optics for shadowing.
- Attenuation estimation in forest canopies.
- Frequency correlation for estimating forest
canopy parameters (trunk thickness, etc.)
110Sequential Adaptive Multi-Modality Target
Detection and Classification Using Physics-Based
Models
K. Sarabandi, M. Dehmolaian, F. Wang, T. Benjamin
Radiation Laboratory The University of Michigan,
Ann Arbor, MI 48109-2122 saraband_at_eecs.umich.edu
111Phenomenological Study
Physics-Based Scattering and Propagation Modeling
of Forest and Embedded Targets
- Forest is a complex random medium composed of
lossy scatterers arranged a semi-deterministic - Foliage cause significant attenuation,
scattering, field fluctuation - Target is in the close proximity of many
scatterers (strong field fluctuations and phase
front distortion)
- Signal level, fluctuations, polarization state,
impulse response, spatial coherence etc. depend
on Tree density, type, height, and structure
- Military targets are usually large and
structurally complex - Significant multiple scattering and shadowing
112Electromagnetic Scattering Simulation of Hard
Targets Embedded in Foliage
- Objectives
- To develop an accurate EM model for forest stands
to allow performance assessment of radar sensors
and target detection algorithms. - Determination of foliage channel, RCS of clutter,
target signature in foliage - Examination of different modalities (f, p, q) on
target/foliage signature.
- Challenges
- Hard target and clutter constitute a
computationally very large problem. - Target and clutter are structurally complex
(features vary from small to very large objects).
113- Forest Model
- Arbitrary fractal tree structures
- Discrete coherent scattering model
- First-order uniform near-field/far-field
calculation inside and outside forest
- Target Models
- Full-wave (MoM, FDTD) computationally
inefficient, good for flt300 MHz - Approximate solution (GO ray tracing, PO)
- Target-Foliage Model
- Low frequencies (flt100 MHz) Full-wave methods,
Scattering from foliage can be ignored - Mid-frequency range (flt1 GHz) Hybrid FDTD and
the forest code - High frequency (fgt1GHZ) Hybrid PO and improved
forest code
114Tasks Under Phenomenological Studies
Forest Model
Target Model
Hybrid Forest/Target Model
Forest parameter estimation
- Enhance model accuracy
- Improve computational efficiency
- Improve range of validity of models
- Provide simulated data to SM team
- Work with reduced models to improve computation
time.
115Progress
- Forest Model
- Accurate estimation of attenuation rate for
near-grazing incidence (long distance
propagation) - Efficient method for inclusion of multiple
scattering - Forest Parameter Estimation
- Application of frequency correlation function
(FCF) - Hybrid Target/Foliage Model
- Direct computation of forest scattered magnetic
field. - Implementation of iterative PO to efficiently
account for target shadowing and double bounce
effects on the target.
116Accurate Estimation Long-Distance Signal
Attenuation in Foliage
Forest Model Improvement
- Issues related to direct wave propagation over
long distances in foliage - A novel model for accurate predication of signal
attenuation based on a renormalization approach - Estimation of forest block statistical parameters
using a numerical approach - Overall signal estimation using a network theory
117Estimation of Path-lossin dense random media
- Experimental data indicates signal attenuation
with distance shows a nonlinear behavior with
distance - Path loss is usually computed from Foldys
approximation (single scattering, far-field
approximation) - Overestimation of attenuation rate
- Significant error over long distances
- Signal attenuation
- a - absorption
- b - scattering loss
- c scattering gain (multiple scattering)
118Statistical WAve Propagation (SWAP) Model
A Hybrid Statistical and Wave Theory Approach 1-
Statistically homogeneous forest properties can
be used to localize the field computation.
- A forest environment can be divided into
statistically identical blocks along the
direction of wave propagation. - Each block of the forest can be considered as an
N-port network with similar statistical
properties. - Once the input-output relation is determined, it
can be used in a network approach to find the
forest channel path-loss.
119SWAP Model
2- Break received power into coherent and
incoherent components.
jth block
Rx
- Received field contains mean and fluctuation
components, received power contains coherent and
incoherent components. - Coherent power comes from the mean field which is
the incident wave attenuated by the effective
forest medium (Foldys approximation). - Incoherent power comes from the fluctuation
field, which contains the contribution from
scatterers within each block of forest (assuming
the blocks are statistically independent).
120SWAP Model
3- Determine the input-output relationship of a
typical block.
Input
Output
Elementary currents computed from fluctuating
fields
Field components computed from the coherent
forest scattering model for each pixel
A forest block made up of many statistical
fractal trees with random location
- Assuming spatially uncorrelated input for
fluctuating fields and using Monte Carlo
simulation find the output mean-field and
standard deviation (fluctuating field)
- Repeat the same procedure for a plane wave
illumination (mean-field incident)
121One Block Simulation
z
y
- Single scattering theory
- plus Foldys approximation
- Coherent mean field incident at each individual
scatterer generates scattered field at the
observation point which is then coherently added.
- Monte-Carlo simulation
- Randomly distribute the tree locations to
simulate the statistical properties of forest.
x
Note Considering the statistical homogeneity
along y-dimension, only a line of observation
points along z-dimension are selected. The
spacing is half-wavelength for accurate
estimation of statistical parameters.
122Desired Statistical Parameters for Estimation
- Variation of fluctuation field
- Spatial Correlation function
- Foldys attenuation coefficient
- Input-output relationship transmission matrix
Assumption statistical properties of forest
depend on the forest itself, not of the
excitation, therefore planewave incidence is
chosen for simplicity.
Note the estimation is conducted within one
representative block of forest and the results
are reused for any blocks.
123Spatial Correlation Function
- C1(?y), C2(?y) are the spatial correlation
functions along a horizontal line at two vertical
points. - C3(?z), C4(?z) are the spatial correlation
functions along a vertical line at two horizontal
points. - C1(?y) and C2(?y) are very similar due to the
statistical homogeneity of forest along
horizontal dimension. - C3(?z) and C4(?z) are much different since the
vertical structure of the forest is not
homogeneous.
Along vertical direction
Along horizontal direction
124Algorithm Flowchart
125Computation of Incoherent Power
- Radiation from the output surface of the jth
block is computed using the field equivalence
principle. Only the fluctuating component is
considered.
- Ground effect is taken into account by using
image theory.
- Surface fluctuation field beyond the forest
dimensions (i.e. the broadening effect) can be
neglected.
126Computation of Incoherent Power, ctd.
- Incoherent power radiated from the jth block of
forest to the receiver.
- Stationary phase technique can be applied for the
integration along y-direction due to the
statistical homogeneity along that direction.
where,
127SWAP Model Validation
Fractal pine trees generated Tree height 8m,
Trunk height 1.2m
- Three sets of simulations are performed
- Model validation (comparison between numerical
foliage model and SWAP model) - SWAP model simulation of signal attenuation at
different frequencies - SWAP model simulation of signal attenuation for
different tree densities at 500 MHz.
128Model Verification
- Comparison between numerical foliage model and
SWAP model - Frequency 0.5 GHz, Tree density 0.05/m2
- Observation point height 1.5m, distance from
forest edge 1m
- SWAP model is reasonably accurate compared to the
single scattering model. - Dual-slope phenomenon is clearly observed from
the SWAP model simulation result.
129Simulation Results
- SWAP model applied to same forest at different
frequencies - Tree density 0.05/m2, Forest range up to 500m
- Observation point height 1.5m, distance from
forest edge 10 m
- Dual-slope phenomena are observed at all
frequencies. - The knee point occurs at shorter distance as f
increases due to higher incoherent power. - Attenuation rate of the mean field is increasing
with f. - Scattering power is increasing with f. Incoherent
power tends to dominate the field after the knee
point.
130Simulation Results (III)
- Different Tree Densities
- Frequency 0.5GHz
- Observation point height 0.75m, distance from
forest edge 10 m
- Dual-slope phenomena are observed at all tree
densities. - The knee point occurs at shorter distances as
tree density increases. - Higher tree density causes more attenuation
effect on the coherent power but gains more
incoherent power which dominates after the
slope-turning point.
131Conclusions
- The SWAP model efficiently includes effects of
scattering in foliage attenuation. - The model for all single scattering effects.
- The model accurately predicts the change in
attenuation rate as a function of distance as
observed in measurements. - Future improvements
- Including multiple scattering among scatterers
within one block - Improve the calculation of mean field at the
output surface of each block by considering the
scattered field from scatterers within adjacent
blocks (both forward and backward)
132Forest Model Enhancement
- At High frequencies the effect of multiple
scattering among tree components become important - To account for all multiple scatterings the
simulation becomes computationally intractable,
however the interaction up to the second order
seems to be sufficient . - Efficient methods for inclusion of multiple
scattering - far-field method
- Near-field method
133Second Order Scattering
Objects are in the near field of each other
Apply Reaction Theorem
The incident field induces a current density
on the particle 1 in the absence of particle 2.
is the near-field scattered field from
particle 2 when it is excited by an
infinitesimal current source along at the
observation point.
is the first plus second order scattered
field from particle 1.
134Complete Second-order for two broad leaves
Using the VIPO approximation the far field
expression for the scattered field from a
circular disk is,
As the leaves get near to each other the exact
near field expression for the scattered field is
used,
135Back Scattered RCS versus tilt angle of the
second leaf
Validation using MoM for d2l (Far-field Method)
Vertical Polarization
Horizontal Polarization
136Validation using MoM for dl (two leaves are in
the near field zone of each other)
Vertical Polarization
Horizontal Polarization
137Phase of back scattered field versus tilt angle
of the second leaf
dl
Vertical Polarization
Horizontal Polarization
138Estimation of Forest Channel Parameters From The
Frequency Correlation Function of Radar
Backscatter
- Goal Need to remove the effects of foliage from
the target signature for target detection and
identification (Electromagnetic defoliation in a
statistical sense) - Require Parameters
- Tree height
- Foliage attenuation rate
- Volume scattering
- Ground reflectivity
139Theoretical Formulation
FCF of a Homogeneous Foliage Layer above a Ground
Plane
- Consider a uniform distribution of scatterers
above a dielectric ground causing attenuation
and volume scattering. - Effective propagation constant , scattering per
unit volume
Indirect term
Clutter-ground term
Ground-clutter term
Direct term
Radar
z
Volume Scattering
d
Ground
140Backscattering Decomposition Using Fourier
Transform of FCF
.
141Simulated Frequency Response of a Tree Stand
Magnitude
Phase
Trunk-ground
FT
Canopy
142Frequency Correlation Function
Frequency spacing 2 MHz Number of
realizations 50
Canopy
Ground-trunk
27m2H
143Simulation Data
Tree trunk
Simulation of 50 trees over 500 MHz
f (GHz)
Tree canopy
Realization
Realization
Simulated data contains tree trunk, tree
canopy, and noise floor
144Choose SAR Data Similar to Simulated Data
X-band SAR image (B500 MHz)
? Range
Tree canopy, ground, and noise floor (this
resembles the simulated data)
Azimuth ?
Ground only
Perform similar FCF analysis on these two
SAR patches
145Preparing SAR Image for FCF
Bandwidth to get high resolution Can we extract
FCF from high resultion SARS Sacrifice resolution
for achieving tree structures.
146Analyzing SAR FCF Using Small Correlation Windows
High attenuation at X-band does not allow
extraction of tree height and structure
Homogeneous Tree Area
147Overview of High-frequency Model
Hybrid Target/Foliage Model
- Calculate scattering from the target inside a
forest using PO approximation - Valid for targets large compared to l and in
specular directions - Forest scattering at high f is significant,
- hence the target is illuminated from all
directions. - Independent of observation point there will be
- many specular contributions.
- Process
- Calculation of field distribution on the
scatterer using the coherent forest model. - Based on these calculated fields derive PO
currents on the target. - Apply the reciprocity theorem to calculate
scattered field from the target that includes the
effects of trees.
148Hybrid Target/Foliage Model
Calculation of PO currents requires
Z
Y
Complex near to far-field expressions of forest
code provides
X
The code is modified to calculate directly
speed upgt2
Example Tree trunk near-field
Frequency 2 GHz X 50 l7.5m Observation
Height 1 m Trunk Height 5.64 m Trunk diameter
20 cm
Trunk Height
Observation Height
149Magnetic Near-Field using the old and new methods
V-Pol. Incidence
H-Pol. Incidence
H
H
Distance 4.2 m
Distance 4.2 m
Time to Run a simulation for lXl plate behind a
10 trees is approximately 3.125 times faster.
150Hybrid Target/Foliage Model
Dimension of Computational Domain
80lX100lX100l Number of scatterers excuding
needles gt 50,000
Sensitivity analysis
Frequency 2GHz Number of Trees 10
Simulation Scenario
Incident direction
Z
Y
q
3 l
3l
f
X
Height 1m
151Sensitivity of the electric current on the plate
to the forest realization
Realization 2
Realization 1
Realization 3
The electric current induced on the plate is
highly sensitive to the arrangement of
trees. Scattering from nearby trees is very
significant.
152Sensitivity analysis
- Enhanced SAR Target Detection Methods
- Multi-incidence angle data
- Spotlight SAR
- SAR tomography
153Sensitivity to elevation angle
For fixed f 177 the induced current on the
plate is plotted for 3 close q.
154Calculation of backscattering Using Reciprocity
Elementary source at the excitation point
Field computation inside forest
Induced current calculation on the scatterer
Scattered field at the excitation point
Apply reaction theorem
No need for computation of scattering from forest
155Backscattering sensitivity to Azimuthal and
elevation angles
Clutter
Plate
- Backscatter
- from plate
- fluctuates more
- along the elevation angle than the
azimuthal angle. - Backscattering is sensitive to
- forest realization, elevation and azimuthal
angles. -
Clutter
Plate
156Backscattering for different elevation
azimuthal angles for 2 different realizations
f
f
q
q
f
f
q
q
Note Fluctuation along the elevation angle is
more than that along azimuthal angle.
157Cross pol Comparison
Level of the xpol from the forest is about 10dB
more than that of the plate.
158Backscattering sensitivity to forest realization
Back scattering from forest and plate are highly
sensitive to the forest Realization.
159Another Example 3D Box
For 3-D objects the lit and shadow area from all
scatterers in forest must be identified
POGO Approach PO current estimation GO
shadowing
Direct Wave
Shadow for reflected
Lit for direct
Reflected Wave
Shadow for direct
Lit for reflected
Direct is shadowed if
Ground Plane
160Simulation Results
Z
Freq 2 GHz 10 Pine trees
Y
3l
Two view of the box
3l
3l
Height 1m
X
Ground Plane
Note Direct Incident field has strong effect on
the level of current.
161Backscattering plots versus elevation angle
- Freq 2 GHz
- Pine Trees
- Target Metallic Box
- f 0 Degrees
svv
Note Level of backscattering from the box Is
comparable to that of the forest.
shh
shv
162Complex Objects
- For complex objects GO-PO Solution becomes
intractable - Estimation of shadowing is difficult, the
algorithm is very complex and becomes the
bottleneck in the scattering computation - For each forest scattere and for each observation
point, shadowing should be estimated. -
Incident wave
Shadow
Lit
Very complicated algorithm for an arbitrary
object.
163Iterative PO Approach
Iterative near-field PO approach
Incident field
Plate 2