Computer vision for non-visual

unclesamnorweiganΤεχνίτη Νοημοσύνη και Ρομποτική

18 Οκτ 2013 (πριν από 3 χρόνια και 9 μήνες)

93 εμφανίσεις

Computer vision for non
-
visual
spectral regimes and non
-
traditional applications

Rama Chellappa

UMD

Opening remarks


non
-
visual sensors


Sensors in visual regimes are only a small % of
the possible sensors


Infrared, hyperspectral, LADAR, SAR, FOPEN
SAR, mmWave, polarimetric


Provide thermal signatures, material properties,
3D, all weather/all day coverage, looking behind
trees, walls, …


Sensor physics is important.


Low signal to clutter ratio.


Characterization of image statistics driven by
sensor physics

Problems addressed using
hyperspectral images


Sensor design


Material classification/segmentation


Anomaly detection


More than two decades of work


ARO MURI (2002
-
2007) on the science of land
-
based target
signatures.


DARPA DDB program, NGA


Glenn Healey: pioneer in HSI and computer vision


Larry Wolf, Equinox Corporation


Unmixing of pixels (Remote sensing)


Compressive sensing possibilities (SPC)


Numerous books, papers, conferences…






Estimating object reflectance

4



Radio transfer function






irradiance to the camera



reflectance at location (x,y)



main light source (e.g. sun light)



ambient light source coming from all directions,
assuming constant intensity for all directions



fraction of unblocked sky from (x,y) view






geometry factor








Nguyen and Chellappa,

CVPR 2010 Workshop on

Beyond visible spectrum

5

Tracking radiance

Reflectance tracking



Robust against illumination, abrupt motion


Capable of recovering after losing track


6

Detection of land mines


Statistical models of clutter (non
-
Gaussian)


Optimal detection methods


Sup
-
pixel detection methods


Detection of disturbed earth


Detection of mass graves in the Balkans

Broadwater and Chellappa, PAMI 2007, SP, 2010

Radar images


Synthetic aperture radar (SAR), foliage
-
penetrating SAR (FOPEN SAR), …


Speckle noise
(random fluctuations in the return signal from an object that is no
bigger than a single image
-
processing element)


Shape from shading


radar clinometry (USGS, Frankot, Chellappa, AI
Journal, 1990)


3D from interferometric SAR (Zebker and Goldstein), stereo SAR


Object detection, indexing recognition (DARPA programs MSTAR, SAIP,
DDB)


MSTAR program (feature extraction, indexing, prediction, recognition)


Typical
object recognition framework


Features can correspond to scatterers (supported by physics)


FOPEN SAR


Low signal to clutter ratio (tree trunks produce stronger returns than objects)


Symmetric alpha
-
stable noise


Global hawk, TESAR sensors

R.T. Frankot and R. Chellappa, Estimation of Surface Topography from SAR Imagery Using Shape

from Shading Techniques, in Physics
-
Based Vision: Shape Recovery, (eds.), L.B. Wolff, S.A. Shafer and

G.E. Healey, Jones and Bartlett Publishers, Boston, MA, pp. 62
-
101, 1992.


LADAR images


Took off in the mid eighties


From machine vision to outdoor conditions


Feature (step, crease edges, surfaces,) extraction,
matching and recognition (Hypothesis
prediction/verification)


Fundamental forms I and II (Besl and Jain)


Pulsed
-
Doppler LADAR (2km) for ATR


Better resolution at longer ranges


Kinect

R. Chellappa, S. Der and E.J.M. Rignot, Statistical Characterization of FLIR, LADAR

and SAR Imagery, in Statistics and Images, K.V Mardia, (ed.), Carfax publishers,

Oxfordshire, U.K., pp. 273
-
312, 1994
.

Opening remarks


non
-
traditional
applications


Road following, lane tracking and other
automotive applications


Dickmanns, Pomereleu, DOT, FHWA


Computer vision for the blind


Navigation, face/expression recognition


Analysis of Schlieren images


Detection of oblique structures such as shock waves
and shear layers


Understanding bee dances


Industrial inspection

Vision for Schlieren data reduction


Schlieren Imaging


Aerodynamic visualization technique in
wind tunnels
-
> long established


Capture density gradients in supersonic
flow


Shock waves, shear layers and turbulent
structures


High speed imaging
-
> data deluge


Images complete the picture


Offer what surface measurements may
not


Are we taking advantage of the data
collected?


Can vision extend analysis
capabilities?


Additional insight to flow unsteadiness


Desire for automation


Removal of human subjectivity


Recast as a segmentation/feature
extraction problem


Extraction of oblique structures


Oblique flow structures ubiquitous


Physically meaningful
segmentation


Bilateral filter
-
> isoperimetric cut


Shock wave and shear
-
layer
inclination


Canny & Hough transform


Classification


Length & quantifiable bounds


Location enforced from labeling


Scale implementation for
robustness?


Hough transform binning


Convergence from two Hough grids


Success of automation 92
-
94%


Outer shock motion history


Recommended for Publication
in AIAA Journal


Vision: viable analysis tool



Waggle dance

Orientation of waggle axis


Direction
of Food source.(with respect to sun).

Intensity of waggle dance


Sweetness of food source.

Frequency of waggle


Distance of
food source.

Parameters of interest in the waggle
dance

Waggle Axis : Average orientation
of Thorax during Waggle.

Duration of Waggle : Number of
frames of waggle in each segment
of the dance.

Anatomy/behavior modeling
-

prior


Three major body parts; each body part
modeled as an ellipse.


Anatomical modeling ensures


Physical limits of body parts are
consistent; accounts for structural
limitations and correlation among
orientation of body parts


Insects move in the direction of their head.

Veeraraghavan, Chellappa and Srinivasn, IEEE TPAMI, March 2008

Insects display very specific
behaviors
-

priors

Modeling behavior explicitly
improve

Tracking performance

Behavioral understanding

Position tracking and
behavioral analysis in a
unified framework.

Result

Detect Frames of Waggle Dance by
looking at

Rate of change of Abdomen
Orientation

Average absolute motion of center of
abdomen in the direction
perpendicular to the axis of the bee.


Parameters of Interest in the Waggle
Dance

Waggle Axis : Average orientation of
Thorax during Waggle.

Duration of Waggle : Number of
frames of waggle in each segment of
the dance.

Looking for a screw amid screws
(MERL)

My advice to the young ones


Look for collaborations outside EE, CS


Helps with multidisciplinary credentials


Will help with winning MURIs: best source of
support.


There are top Transactions and journals that
accept these papers!


Computer vision presents immense
opportunities outside traditional areas.