Radar Signal Processing: Opportunities for SSPerseecs.umich.edu/ssp2012/moses.pdf · Radar Signal...

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Radar Signal Processing: Opportunities for SSPers

Randy Moses

Dept. of Electrical and Computer Engineering The Ohio State University

This research was supported by AFOSR, AFRL, and DARPA

TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: AAAAA

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Context

n  Advances in digital processing are enabling revolutionary opportunities for radar signal processing

n  Opportunities for Statistical Signal Processing – Persistent sensing over space and time – More sophisticated radar image/volume reconstructions – Multi-function radars that can simultaneously perform imaging,

detection, moving object tracking and recognition, … – Uncertainty analysis and estimation bounds

n  Challenges – Traditional models for radar backscattering may not apply over

wide angles – Large data, processing, and communications tasks

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Persistent, Wide-Angle Radar

Persistent Sensing enables: • High resolution, volumetric imaging of stationary objects and

scenes • Continuous tracking of moving objects

UAV UAV

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Outline

n  Radar 101 n  Revisit modeling assumptions for wide angle radar n  How can sparsity play a role?

– Parametric Modeling – Sparse Reconstruction

n  Feature-Based Classification n  Transmit adaptation

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Outline

n  Radar 101 n  Revisit modeling assumptions for wide angle radar n  How can sparsity play a role?

– Parametric Modeling – Sparse Reconstruction

n  Feature-Based Classification n  Transmit adaptation

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

SAR Data Collection

Frequency Space!

θ

φ

• • • • •

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

SAR Image Formation

n  Traditional approach: tomography

n  Tomographic image I(x,y) is a matched filter for an isotropic point scatterer at location (x,y). [Rossi+Willsky]

I(x,y)

2D IFFT

E(f,φ)→E(fx,fy)

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

3D Reconstruction

n  Large data size and processing requirements

n  Filled aperture is difficult to collect

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

9

X-Band Radar 3° aperture 1ft x 1ft res

Example: Ohio Stadium

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

10

SAR Image Detail

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Outline

n  Radar 101 n  Revisit modeling assumptions for wide angle radar n  How can sparsity play a role?

– Parametric Modeling – Sparse Reconstruction

n  Feature-Based Classification n  Transmit adaptation

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Persistent, Wide-Angle SAR

Persistent Sensing enables: • High resolution, volumetric imaging of stationary objects • Continuous tracking of moving objects

UAV UAV

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

AFRL Gotcha Radar

5 Km

15.25 Km

Data Storage: 90 G samples/circle

Image formation: 45 Tflops/sec

Communications: 190 M samples/sec

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Wide Angle Scattering Behavior

n  At high frequencies, radar backscatter is well-modeled as a sum of responses from canonical scattering terms.

n  EM scattering theory provides a rich characterization of backscatter behavior as a function of object shape – Azimuth, elevation, frequency dependence – Polarization dependence – Phase response - range

n  Most backscatter does NOT behave like a point scatterer over wide angles – Standard imaging is not statistically (close to) optimal

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Scattering Model

Jackson & RLM: 2009

Frequency!Dependence"

Location!Dependence!

"

Aspect!Dependence"

Polarization!Dependence!

f 2 [9.7, 10.3] GHz

�m = 3�

⇡ c2�fx

⇡ c2�fy

S(f,�, ✓) =

AHH AHV

AVH AV V

� ⇣jffc

⌘�

M(�, ✓) ej4⇡ffc

�R(x,y,z;a)

⇥ =

2

6666664

p0p0p0

p1 � p0...

pn�1 � p0

3

7777775

F⇥ =

2

6666664

F0 Fp0

p0p0

p1 � p0...

pn�1 � p0

3

7777775

~pk p0, p0

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Wide-Angle Data Collections

When the radar measurement extent is ≤ scattering persistence, the isotropic assumption is ~satisfied, and tomographic imaging is ~a matched filter.

azimuth

Sc. Ctr Responses

φc φc

α α

Radar measurements

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Wide-Angle Data Collections

For wide-angle measurements the isotropic scattering assumption breaks down.

– Tomography is no longer a matched filter

azimuth

Sc. Ctr Responses

φc φc

α α

Radar measurements

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

φc=−40°

φc =−20°

φc =0°

φc=20°

φc=40°

Frequency Support Image

φc=−40°

φc =−20°

φc =0°

φc=20°

φc=40°

Frequency Support Image

Scattering Aspect Dependence

Most scattering centers have limited response persistence 20° or less [Dudgeon et al, 1994]

Image response is no longer characterized by a single impulse response shape.

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Coherent wide-angle SAR Images

Coherent wide-angle image is not well-matched to limited persistence scattering behavior!

500 MHz Bandwidth 110 degrees az

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

GLRT Image Formation GLRT Image: Image I(x,y) is GLRT output at (x,y) to a limited-

persistence scattering center with center φc and width α."

αφαφ

αφαφ

width, center withimage std cc

c

yxR

yxRyxIc

=

=

),,,(

),,,(maxarg),(,

GLRT Image: !

Approximation: fix width α; quantize φc. !!Then the GLRT image is approximately

max over sub-aperture images.!

azimuth φc

α

RLM, Potter, Cetin: 2004

azimuth

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

GRLT Imaging

max

Frequency Data

GLRT Image

Generalizes Rossi+Willsky matched filter result to wide-angle imaging with limited-persistence scattering

RLM, Potter, Cetin: 2004

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Coherent and GLRT Image

110° Coherent Image 110° GLRT Image

4 GHz bandwidth

RLM, Potter, Cetin: 2004

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Outline

n  Radar 101 n  Revisit modeling assumptions for wide angle radar n  How can sparsity play a role?

– Parametric Modeling – Sparse Reconstruction

n  Feature-Based Classification n  Transmit adaptation

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Sparse 3D Radar Reconstruction

n  3D radar reconstruction necessarily will use (very) sparse measurements

n  Is the radar reconstruction sufficiently sparse to overcome measurement sparsity?

k-space

AFRL Backhoe Data Dome, with sparse “squiggle path” shown

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Polarization

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Squiggle Path 3D Tomographic Reconstruction

Largest

Smallest

Voxe

l mag

nitu

de Squiggle

PSF

Top 25 dB voxels shown

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Parametric: Canonical Scattering Model

Frequency!Dependence"

Location!Dependence!

"

Aspect!Dependence"

Polarization!Dependence!

Jackson & RLM: 2009

f 2 [9.7, 10.3] GHz

�m = 3�

⇡ c2�f

x

⇡ c2�f

y

S(f,�, ✓) =

AHH AHV

AVH AV V

� ⇣jffc

⌘�

M(�, ✓) ej4⇡ff

c

�R(x,y,z;a)

⇥ =

2

66666664

p0p0p0

p1 � p0...

pn�1 � p0

3

77777775

F⇥ =

2

66666664

F0 Fp0

p0p0

p1 � p0...

pn�1 � p0

3

77777775

~pk p0, p0

S(f,�, ✓) =

KX

k=1

AHH AHV

AVH AV V

�

k

✓jf

fc

◆�k

Mk(�, ✓) ej 4⇡f

f

c

�R(xk

,yk

,zk

;ak

)

Θ

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Sparsity – Measurements y (M £ 1) : sparse sampling of full (f,az,el) radar

measurement space – Reconstruction: x (N £ 1): sparse set of (x,y,z) locations with significant

radar scattering energy

Sparse reconstruction:

Nonparametric: lp Regularized Least-Squares

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Algorithmic Challenges

n  For large scale problems, the algorithm can become very memory and computationally expensive. – E.g. for backhoe squiggle problem:

l  M ≈ 105

l  N ≈ 107

n  A is structured and may not satisfy RIP for reconstruction samplings of interest. – Recent advances [e.g Austin, Fannjiang] incorporate structure in x to

allow high coherence in A.

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Squiggle Path Collection: lp Regularized LS Reconstruction

Top 30 dB voxels shown; p=1

Austin, Ertin, RLM, 2011

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Backhoe Squiggle Image

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Gotcha Vehicle Data

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Gotcha lp Reconstructions: Camry

Austin, Ertin, RLM, 2011

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Outline

n  Radar 101 n  Revisit modeling assumptions for wide angle radar n  How can sparsity play a role?

– Parametric Modeling – Sparse Reconstruction

n  Feature-Based Classification n  Transmit adaptation

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Vehicle Classification

Camry

Maxima Prism Taurus

Malibu Sentra

•  Six sedans from a 10-class problem •  Spatially-varying signatures across

large scene

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Vehicle Classification; Attributed Point Sets

0

5

10

0

5

100

100

200

300

x(m)

Camry q=0.3, θv=0, pose=0°

y(m)0

5

10

0

5

100

100

200

300

x(m)

Camry q=0.0, θv=0, pose=0°

y(m) 0

5

10

0

5

100

100

200

300

x(m)

Camry q=0.4, θv=0, pose=216°

y(m)

>95% correct classification

PHD,  dij  

Squared Euclidian Distance Matrix

MDS

x

y

φ

Dungan and Potter, 2011

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Outline

n  Radar 101 n  Revisit modeling assumptions for wide angle radar n  How can sparsity play a role?

– Parametric Modeling – Sparse Reconstruction

n  Feature-Based Classification n  Transmit adaptation

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Joint Communication and Radar Sensing

Can adaptive transmit waveforms be used to simultaneously sense a scene and communicate sensed data to a receiver?

AFRL Gotcha Radar Communications:

190 M samples/sec

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

OSU Software Defined Radar

n  Compact n  Wide Bandwidth (125 MHz) n  Real-time (DSP + FPGA)

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Joint Sensing-Comm Experiment

n  Self-adaptive joint radar/communication system – PN transmit signal waveform

n  Measured and communicated range-Doppler maps – nth range-Doppler map used to adapt (n+1)st waveform set.

Doppler Range

Measured Communicated Rossler, Ertin, RLM: 2011

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Take-Home Points

n Advances in sampling and digital processing are moving radar systems more firmly in the digital realm.

n Persistence and wide-angle sensing motivate rethinking the models and algorithms for radar processing.

n Effective 3D reconstruction from sparse apertures is possible – Surprising fidelity – Huge issues in computation, communication remain

n Huge opportunities for SSPers – Modeling; tractable algorithms; adaptation; persistent tracking;

classification; performance estimation

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Data Resources

n  AFRL Sensor Data Management System – https://www.sdms.afrl.af.mil – Backhoe Volumetric Data (synthetic) – Civilian Vehicle Data Domes (synthetic) – Gotcha data (measured) – SAR-GMTI Challenge Problem

Presented at: the 2012 Statistical Signal Processing Workshop, Ann Arbor

Thank you!