real world data fusion - enseeihtsc.enseeiht.fr/doc/seminar_daum_2012_1.pdf · le 19/06/2012 à...
TRANSCRIPT
real world
data fusion
Fred Daum
15 June 2012
data fusion
Copyright © 2012 Raytheon Company. All rights reserved.
Customer Success Is Our Mission is a trademark of Raytheon Company.
1
PATRIOT
Firefinder
F/A-18
Global HawkFirefinderPAVE PAWS
Global Hawk
2
theoretical optimal multi-sensor data fusion
fusion of measurements
performance
fusion of
tracks
interesting parameter3
real world multi-sensor data fusion
fusion of tracks
performance
fusion of measurements
interesting parameter4
real world issues in multi-sensor data fusion
• limited resolution of sensor data
• residual bias & drift errors of sensor data
• physics of real world bias & drift errors
• data association errors
• not all objects detected & resolved & reported by sensor • not all objects detected & resolved & reported by sensor
A are detected & resolved & reported by sensor B
• error covariance matrix inconsistency
• nonlinear & non-Gaussian filtering errors
• ill-conditioning
• limited bandwidth & latency & poor connectivity & poor
content of data communication between multiple sensors
• other5
Séminaire de Statistique. Le 19/06/2012 à 11h00,
Salle de séminaire du 1er étage, bât.1R3
Frederic Daum : Nonlinear filters with particle flow
We have invented a new particle filter, which improves accuracy by several orders of magnitude
compared with the extended Kalman filter for difficult nonlinear problems. Our filter runs many
orders of magnitude faster than standard particle filters for problems with dimension higher than
6
orders of magnitude faster than standard particle filters for problems with dimension higher than
four. We do not resample particles, and we do not use any proposal density, which is a radical
departure from other particle filters. We show very interesting movies of particle flow and many
numerical results. The key idea is to compute Bayes’ rule using a flow of particles rather than as a
point wise multiplication; this solves the well known problem of “particle degeneracy”. Our
derivation is based on freshman calculus and physics. This talk is for normal engineers who do not
have log-homotopy for breakfast.
multi-sensor data fusion literature
publication sensor bias sensor resolution
physics of real world bias
Fusion Conferences 1% 0% 0%
Handbook of Multisensor Data Fusion
0% 0% 0%
Fusion
Sam Blackman’s books
1% 1% 0%
Yaakov Bar-Shalom’sbooks
2% 3% 0%
Oliver Drummond’s SPIE conferences
1% 1% 0%
IEEE Aerospace & Systems Transactions
1% 1% 0%
7
bias can ruin multi-sensor fusion
300 total targets: 30 missiles, 10 targets per missile
Position error σ = 100m, Separation of targets in missile complex = 500m, 1500m
60
70
80
90
100
% C
orr
ec
t a
ss
ign
me
nts
GNPL
0
10
20
30
40
50
60
1 10 100 1000 10000
Magnitude of bias (m)
% C
orr
ec
t a
ss
ign
me
nts
JVC
GNPL jointly
optimizes the
bias estimates &
data association
8
GNPL vs. JVC with Bias
• 7 remote tracks and 29 local tracks (2 of the remote tracks have no
local track) with residual radar bias
10
GNPLJVC
2 4 6 8 10 12
2
4
6
8
2 4 6 8 10 12
Perfect
fusion
All fusion is
incorrect
9
GNPL explicitly models bias*
• Score function jointly optimizes data association and bias
estimation:
( )[ ] [ ]∑
−−
=
≠+−−−=
+=
−−=
miii
T
iMT
a
iaii
iaii
iag
iaSxSxRxRxJ
QPS
xBAx
1
1
)(
)(
0)(
0)(ln2ln
δδπ
δestimated bias
• The optimal estimate of bias is:
[ ] ∑=
=i
aiag1 0)(
[ ] [ ] [ ]
∑
=
≠−+
∑ ++=
=
−−
=
−−m
1i
ai1
ai
1m
1i
1ai
1
0)i(a0
0)i(aBAQPQPRx ii
i
10*Mark Levedahl, proceedings of SPIE conference 2002.
should be computed
adaptively
comparison of algorithms
algorithm bias estimation? performance
1. GNPL + adaptive gating yes (jointly with
association)
2. iterative bias estimation & JVC yes
3. histogram of bias over association hypotheses
yes
4. covariance inflation & JVC no
5. association of objects (JVC) no
11
GNPL with adaptive gating is much better than Iterative JVC
(d = 6 & N = 10 objects)
0 10 200
0.2
0.4
0.6
0.8
1
Perc
ent C
orr
ect M
atc
h
100% Track Overlap
0 10 200
0.2
0.4
0.6
0.8
190% Track Overlap
0 10 200
0.2
0.4
0.6
0.8
180% Track Overlap
0 10 200
0.2
0.4
0.6
0.8
170% Track Overlap
0 10 200
0.2
0.4
0.6
0.8
160% Track Overlap
Pro
babili
ty o
f P
CA
0 10 200
GNP L
IJV C
0 10 200
0 10 200
NNp / 1-sigma
0 10 200
0 10 200
0 10 200
0.2
0.4
0.6
0.8
1
Pe
rce
nt C
orr
ect M
atc
h
50% Track Overlap
0 10 200
0.2
0.4
0.6
0.8
140% Track Overlap
0 10 200
0.2
0.4
0.6
0.8
1
NNp / 1-sigma
30% Track Overlap
0 10 200
0.2
0.4
0.6
0.8
120% Track Overlap
0 10 200
0.2
0.4
0.6
0.8
110% Track Overlap
GNP L
IJV C
Pro
babili
ty o
f P
CA
12
GNPL with adaptive gating is much better than iterative JVC
(d = 6 & N = 10 objects)
• Delta (GNPL – IJVC) 10 Object Association Performance
– 1σ Bias Error = 300m , 1σ Track Error = 100m, TOL = track overlap, Monte Carlo runs = 200
0.25
0.3
0.35
0.4
Pe
rce
nt C
orr
ect A
sso
cia
tion
Delta Percent Correct Association (GNPL minus IJVC)
100% TOL
90% TOL
80% TOL
70% TOL
60% TOL
Pro
babili
ty o
f P
CA
0 5 10 15 20 25-0.05
0
0.05
0.1
0.15
0.2
0.25
NNp / 1-sigma
Pe
rce
nt C
orr
ect A
sso
cia
tion
50% TOL
40% TOL
30% TOL
20% TOL
10% TOL
Pro
babili
ty o
f P
CA
13
physics of real world “bias”
physical source of error bias error varies with
1. misalignment of electrical boresite & mechanical IMU axes
azimuth & elevation and/or array sine space (u & v)
2. IMU drift time
3. monopulse slope error angle from beam center
4. scan dependent monopulse bias scan angles from array boresite, temperature of array face & refractivity near array face
5. tropospheric refraction azimuth & elevation & time
6. radome refraction & reflections azimuth & elevation and moisture on radome
14
antennanear ty refractivi and
antenna of re temperatuoffunction
:bias monopulsedependent scan
=
+=
+=
k
kvvv
kuuu
m
m
15
what is “resolution”?
Consider a sensor with resolution volume V, and two
objects A and B. We say that the sensor “resolves” A
and B if the resolution volumes are disjoint.
• A
• B
• A
• B
Resolved Unresolved
16
Probabilities of Resolution and Data Association for
One-Dimensional Measurements
17
simple back-of-the-envelope formulas
PR = probability of resolution
PR ≈ exp (-λV)
λ = density of objects
V = volume of sensor resolution cell
P = probability of correct data associationPDA = probability of correct data association
PDA ≈ exp (-cN)
N = average number of objects in the one-sigma prediction error volume
c = constant (on the order of unity) for a given dimension
18
fundamental theorem of multiple target tracking*
PR < PDA
P = probability of resolutionPR = probability of resolution
PDA = probability of correct data association
*assumes that prediction errors are dominated by noise rather than
target maneuvers. see Daum & Fitzgerald, “The importance of
resolution,” Proceedings of SPIE Conference on signal & data
processing, vol. 2235, pages 329-338, April 1994.19
example of data association problemR
an
ge
t t t t t t t t0 1 2 3 4 5 6 7
Time
Ran
ge
20
solution to data association problemR
an
ge
0 1 2 3 4 5 6 7
Time
Ran
ge
t t t t t t t t
21
example of resolution problem
Ran
ge
t t t t t t t t0 1 2 3 4 5 6 7
Time
Ran
ge
22
0.0
111.6
208.9
273.0
600
540
480
420
RE
FE
RE
NC
E T
RA
CK
AL
TIT
UD
E (
km
)
TIM
E (
s T
AL
O)
(43
80
0.7
7s
UT
C)
BRV
CSO CLOUD
PROPELLANT
DEBRIS
GCS
CSO CANISTERS
CSO PANCAKE
0.0
111.6
208.9
273.0
600
540
480
420
RE
FE
RE
NC
E T
RA
CK
AL
TIT
UD
E (
km
)
TIM
E (
s T
AL
O)
(43
80
0.7
7s
UT
C)
BRV
CSO CLOUD
PROPELLANT
DEBRIS
GCS
CSO CANISTERS
CSO PANCAKE
304.8
304.5
271.9
207.0
-1.6 0.0 1.6 3.2 4.8
360
300
240
180
120
RE
FE
RE
NC
E T
RA
CK
AL
TIT
UD
E (
km
)
RANGE RELATIVE TO REFERENCE TRACK (km)
TIM
E (
s T
AL
O)
(43
80
0.7
7s
UT
C)
BRVPROPELLANT
DEBRIS
M57
BRV / GCS
UNITARY
304.8
304.5
271.9
207.0
-1.6 0.0 1.6 3.2 4.8
360
300
240
180
120
RE
FE
RE
NC
E T
RA
CK
AL
TIT
UD
E (
km
)
RANGE RELATIVE TO REFERENCE TRACK (km)
TIM
E (
s T
AL
O)
(43
80
0.7
7s
UT
C)
BRVPROPELLANT
DEBRIS
M57
BRV / GCS
UNITARY
23
normal debris for ICBM
24
performance improvements due to explicitly
modeling unresolved data
without explicit model of unresolved measurements
with explicit model of unresolved measurements
Koch & van Keuk (1997) unstable tracks and bad estimation accuracy
stable tracks with excellent estimation accuracyaccuracy
Blom & Bloem (2005) 50 % probability of track coalescence
2 % probability of track coalescence
Daum (1986) 50 % probability of maintaining track and poor velocity estimation accuracy
99 % probability of maintaining track and excellent velocity estimation accuracy
25
question:
Suppose that there are 1,000 objects in a single radar
beam, randomly and uniformly distributed in range
over 100 km. Suppose that the radar range
resolution is 10 m.
What is the probability that a given object is
resolved from all other objects?resolved from all other objects?
(a) 0.98
(b) 0.82
(c) 0.26
(d) 0.0
26
answer*
λλλλ = 1,000 objects/100 km
λλλλ = 1 object/100 m
V = 2∆∆∆∆R
V = 20 mV = 20 m
PR ≈≈≈≈ exp (-λλλλV)
PR ≈≈≈≈ exp (-20 m/100 m)
PR ≈≈≈≈ exp (-0.2)
PR ≈≈≈≈ 0.82
*pessimistically assuming no Doppler resolution 27
resolved measurements are good & unresolved data
are essentially useless
Data
Measurement Accuracy of Radar
Data
Value for
Discrimination and Tracking
Resolved
Unbiased and limited by noise approximately as per standard
Good (assuming adequate signal-approximately as per standard
handbook formulas
adequate signal-to-noise ratio)
Unresolved
Strongly biased with large
variances, with errors roughly an order of magnitude larger than
resolved measurements
Essentially worthless
28
resolved measurements are good
Data Measurement Accuracy of Radar Data
Value for Discrimination and Tracking
Resolved
Unbiased and limited by noise approximately as per Cramér-Rao
bound:
NS2nT4
12
NS2B
2cRR
ππππ
λλλλ≈≈≈≈σσσσ≈≈≈≈σσσσ
Good (assuming adequate signal-
to-noise ratio) •
NS2
1
NS2
ANS2kNS2k
NS2nT4NS2B
A
m
3EL
m
3AZ
≈≈≈≈σσσσ≈≈≈≈σσσσ
θθθθ≈≈≈≈σσσσ
θθθθ≈≈≈≈σσσσ
ππππ
φφφφ
Unresolved
Strongly biased with large variances:
12
2badvery
1212B
2c
A
3ELAZR
ππππ≈≈≈≈σσσσ≈≈≈≈σσσσ
θθθθ≈≈≈≈σσσσ≈≈≈≈σσσσ≈≈≈≈σσσσ
φφφφ
Essentially worthless
29
definitions of symbols
λ = radar wavelength
c = speed of light
B = chirp bandwidth of waveform
S/N = signal-to-noise ratio
T = coherent integration time
L = width of radar antennaL = width of radar antenna
D = height of radar antenna
R = slant range
R = range rate
A = target amplitude
φ = target phase
n = number of pulses coherently integrated
DorL
2k
3
m
λλλλλλλλ====θθθθ
ππππ====
•
30
radar resolution formulas*
dimension
formula for Rayleigh
resolution
examples
azimuth LAZ λλλλ====∆∆∆∆ λλλλ = 3 cm and L = 10 m ⇒⇒⇒⇒ ∆∆∆∆AZ = 3 mrad
elevation DEL λλλλ====∆∆∆∆ λλλλ = 3 cm and D = 5 m ⇒⇒⇒⇒ ∆∆∆∆EL = 6 mrad elevation DEL λλλλ====∆∆∆∆ λλλλ = 3 cm and D = 5 m ⇒⇒⇒⇒ ∆∆∆∆EL = 6 mrad
range B2cR ====∆∆∆∆ B = 10 MHz ⇒⇒⇒⇒ ∆∆∆∆R = 15 m
B = 100 MHz ⇒⇒⇒⇒ ∆∆∆∆R = 1.5 m
B = 1000 MHz ⇒⇒⇒⇒ ∆∆∆∆R = 15 cm
Doppler T2R λλλλ====∆∆∆∆ secm3.0Rsecm100Tandcm3
secm3Rsecm10Tandcm3====∆∆∆∆⇒⇒⇒⇒========λλλλ
====∆∆∆∆⇒⇒⇒⇒========λλλλ
• •
•
*Assuming no weighting and no super resolution
31
industrial strength data fusion
explicit model
of resolution
explicit model
of bias in data
association
use sensor A data
to diagnoseunresolved
data forsensor B
robust
fusionalgorithm
monopulsequadrature
for unresolved
data
use chi-square
residuals for unresolved
data
retrodictionof
unresolved data
predictionof
unresolved data
32
• Wolfgang Koch & Günter van Keuk, “Multiple hypothesis track maintenance
with possibly unresolved measurements,” IEEE Trans. AES, July 1997.
• Shozo Mori & Chee Chong, “Multiple Target Tracking with Possibly Merged
Measurements,” SPIE Conference, San Diego, August 2005.
• Henk Blom & Edwin Bloem, “Descriptor system approach towards multi-
target tracking under limited sensor resolution,” March 2005.
• Mark Levedahl, “An explicit pattern matching assignment algorithm,”
Proceedings of SPIE Conference on Signal and Data Processing, Volume
4728, 2002.
• Martin Dana, “Registration: A prerequisite for data fusion,” Chapter 5 in
Multi-target Multi-sensor Tracking, Volume I, edited by Yaakov Bar-Shalom,
Artech House Inc., 1990.
• Fred Daum, “A System Approach to Multi-target Tracking,” Chapter 6 in
Multi-target Multi-sensor Tracking, Volume II, edited by Yaakov Bar-Shalom,
Artech House, Inc., 1992.
• Dimitri Papageogiou & John Sergi, “Simultaneous track-to-track association
and bias removal,” Proceedings of IEEE Aerospace Conference, Big Ski
Montana, March 2008.33
Séminaire de Statistique. Le 19/06/2012 à 11h00,
Salle de séminaire du 1er étage, bât.1R3
Frederic Daum : Nonlinear filters with particle flow
We have invented a new particle filter, which improves accuracy by several orders of magnitude
compared with the extended Kalman filter for difficult nonlinear problems. Our filter runs many
orders of magnitude faster than standard particle filters for problems with dimension higher than
34
orders of magnitude faster than standard particle filters for problems with dimension higher than
four. We do not resample particles, and we do not use any proposal density, which is a radical
departure from other particle filters. We show very interesting movies of particle flow and many
numerical results. The key idea is to compute Bayes’ rule using a flow of particles rather than as a
point wise multiplication; this solves the well known problem of “particle degeneracy”. Our
derivation is based on freshman calculus and physics. This talk is for normal engineers who do not
have log-homotopy for breakfast.
likelihood of
measurementprior
densityoptimal
accuracyg h
root cause of
curse of dimensionality:
curse of dimensionality:
particles to represent the prior
pdf pdf
particles particles
flow of density
flow of particles
sample from
density
sample from
density
λ=0 λ=1
prior posterior
)(log)(log),(log xhxgxp λλ +=
),( λλ
xfd
dx=
f.for PDE above thesolving
by flow particle design the We
loglog)(
+−=
λd
Kdhppfdiv
35
method to solve PDE how to pick unique solution comments
1. generalized inverse of linear differential operator minimum norm* very difficult to design robust stable & fast algorithm
2. Poisson’s equation gradient of potential*(assume irrotational flow)
very difficult to design robust stable & fast algorithm
3. generalized inverse of gradient of log-homotopy assume incompressible flow & pick minimum L² norm solution
workhorse for multimodal densities
4. generalization of method #3 most robustly stable filter or random pick, etc.
workhorse for multimodal densities
5. separation of variables (Gaussian) pick solution of specific form (polynomial) extremely fast & hard to beat in accuracy for many problems
6. separation of variables
(exponential family)
pick solution of specific form (finite basis functions)
needs theoretical work & numerical experiments
7. variational formulation (Gauss & Hertz) convex function minimization needs work
8. optimal control formulation convex functional minimization (e.g., least action like Monge-Kantorovich)
very high computational complexityaction like Monge-Kantorovich)
9. direct integration (of first order linear PDE in divergence form)
choice of d-1 arbitrary functions should work with enforcement of neutral charge density & importance sampling
10. generalized method of characteristics more conditions (e.g., small curvature or specify curl, or use Lorentz invariance)
needs theoretical work & numerical experiments
11. another homotopy (inspired by Gromov’s
h-principle) like Feynman’s QED perturbation
initial condition of ODE &
uniqueness of sol. to ODE
needs theoretical work & numerical experiments
12. finite dimensional parametric flow
(e.g., f = Ax+b with A & b parameters)
non-singular matrix to invert needs numerical experiments
13. Fourier transform of PDE (divergence form of linear PDE has constant coefficients!)
minimum norm* or most stable flow very difficult to design robust stable & fast algorithm
14. small curvature flow & assumed prior density solve d x d system of linear equations new in 2012. Beats other methods for difficult nonlinear problems.
15. small curvature flow & homotopy for inverse of A + B (sum of two linear operators)
numerically integrate ODE new in 2012. extremely cool theory.
16. small curvature flow & homotopy for generalized inverse of A + B (sum of two linear operators)
numerically integrate ODE new in 2012
36
exact flow filter is many orders of magnitude faster per
particle than standard particle filters
- - - - -
bootstrap
EKF proposal
incomp flow
exact flow
3
104
105
106
107
Med
ian
Co
mp
uta
tio
n T
ime
fo
r 30 U
pd
ate
s (
sec)
d = 30
d = 20
d = 10
d = 5 bootstrap
particle filter
EKF proposal
* Intel Corel 2 CPU, 1.86GHz, 0.98GB of RAM, PC-MATLAB version 7.7
25 Monte Carlo trials10
210
310
410
510
-1
100
101
102
103
Number of Particles
Med
ian
Co
mp
uta
tio
n T
ime
fo
r 30 U
pd
ate
s (
sec)
37
exact flow
EKF proposal
incompressible
flow
particle flow filter is many orders of magnitude faster
real time computation (for the same or better
estimation accuracy)
3 or 4 orders of
3 or 4 orders of magnitude faster
per particle
avoids bottleneck in
many orders of
magnitude faster
3 or 4 orders of magnitude
fewer particles
bottleneck in parallel
processing due to resampling
38
10-1
100
101
102
EKF
PF Incompressible
new filter improves accuracy by
two orders of magnitude
median error
N = 500 particles
extended Kalman filter
standard particle filter
0 2 4 6 8 1010
-3
10-2
10
Time (sec)
PF Incompressible
PF Ax+BN = 500 particles
new filter
key idea: particle flow inspired by fluid
dynamics to make solution of PDE
div(pf) = η much faster
Euler’s equations:
3121233
2313122
1232311
)(
)(
)(
MIII
MIII
MIII
=−+
=−+
=−+
ωωω
ωωω
ωωω
&
&
&
39