robust localization, classification, and temporal evolution tracking in auroral data
DESCRIPTION
Robust Localization, Classification, and Temporal Evolution Tracking in Auroral Data. G. Germany * C.C. Hung † T. Newman * J. Spann ‡ * Univ. of Alabama in Huntsville † S. Polytechnic State Univ. ‡ MSFC. AISR Meeting, April 4-6, 2005. Visual Perception I. What do you see?. - PowerPoint PPT PresentationTRANSCRIPT
Robust Localization, Classification, and Temporal Evolution Tracking in
Auroral Data
G. Germany * C.C. Hung †T. Newman * J. Spann ‡
* Univ. of Alabama in Huntsville† S. Polytechnic State Univ.
‡ MSFC
AISR Meeting, April 4-6, 2005
Visual Perception I
• What do you see?
from Torrans 99 (GMU) from Schiffman’s Sensation and Perception, 2000, as presented by Loken et al., Macalester Coll.
Visual Perception I
• What do you see?
from Torrans 99 (GMU) from Schiffman’s Sensation and Perception, 2000, as presented by Loken et al., Macalester Coll.
Gestalt = Form, a unit of perceptionNote: Avoid strict Gestaltism!
Visual Perception II: Some Gestalt
• Continuity
• Closure
• Proximity
• Similarity
EEEEEEEEEEFEEEEEEEEEE
A Problem: Finding Auroral Oval
UVI Image Region-Based Seg.
A Problem: Finding Auroral Oval
UVI Image Region-Based Seg.
* Incomplete
Toward a Better Solution?
• Exploit:
– Continuity
– Closure
– Proximity
– Similarity
Toward a Better Solution?
• Exploit:
– Continuity
– Closure
– Proximity
– Similarity
General Goal: Support Study of Aurora
• UVI Polar Datasets– 9M+ images!
• Exploitation Challenge
– OST (Germany Talk @ AISR05)• Queries based on:
– Sensor Location– Some Image Features:
• Integrated Intensity• Boundary Position• Image Quality
• Popular:– 1000/100
Specific Goal: Effective Retrieval
• UVI OST: Mining Pre-Processing
– Removes Non-Auroral Features
– Extracts Auroral Oval
• Inefficient
• Often fails
• Support Auroral Feature Search
– Shape-Based Processing to Localize Oval
– Better-support Auroral Feature Retrieval
Shape Recovery
• Our prior work:– Spheres– Cylinders– Cones– Ellipsoids– Paraboloids– Hyperboloids– Compound Structures
Exploiting Shape I
• Exploiting Shape: Preliminaries– What Shape?– Shape Invariant Descriptors
2 2 1Ax By Cxy Dx Ey
/ 2 / 2
/ 2 / 2
/ 2 / 2
A C D
C B E
D E F
/ 2
/ 2
A CJ
C B I A B
0, 0, 0 : ellipseJI
Exploiting Shape II: Experiments
• Shape Testing
– Manual Tracing
– Linear Least Squares Fitting
– Shape Invariant Extraction
97/0
11/0
5342
297
/011
/094
246
FittingsManually traced boundaryColorized Image Invar.
∆ = 2E-10J = 1E-9I = -8E-5∆/I < 0
∆ = 9E-9J = 2E-8I = -3E-4∆/I < 0
Some Fitting Results
Initial Work: Localization
• Hough-Based Processing– Democratic:
• Pixels Vote
– Example• y = mx + b
PixelA
y=ax+cy=px+q
y=fx+g
y=dx+e
Approx. Auroral Oval Boundary
• Thresholding with density examination– Thresholding value: (mean of the pixel intensities)
• LoG edge detection
99/006/093006 Thresholded image Dense part Boundary
Shape-based Detection
• Mod. Randomized Hough Transf. (MRHT) for ellipse:– Randomly Select Eedgels– Fit General Quadratic– Determine Shape– Ellipse Parameter Recovery (next page)– Parameter Space Decomposition
• 2D accumulate array for center• 2D accumulate array for axes• 1D accumulate array for orientation
• MRHT for ellipse– Recover:
0 0 2 2
2 2( , ) ( , )
4 4
BD CE AE CDx y
C AB C AB
2 2 1Ax By Cxy Dx Ey
2
arctan( 1 )B A B A
C C
1 2
( , ) ( , )a bJr Jr
where
2 21
2 22
1 ( ( ) ),
21
( ( ) ).2
r A B B A C
r A B B A C
Shape-based Detection (cont.)
center
orientation
axes
Shape-based Detection (cont.)
• Separation of inner and outer boundary– Based on distance to center of detected ellipse
from boundary pixels
Green = detected ellipse from boundary pixels. Red = inner boundary. Blue = outer boundary
99/006/093006
• Shape detection of inner and outer boundary. Shape-based Detection (cont.)
99/006/093006
99/006/070027 Green = RHT ellipses. Red = Inner. Blue= outer boundaries of the aurora arc.
More Experimental Results
96/344/233133 97/011/094246
99/006/21312397/010/150934
Experimental Configuration
• Machine:– CPU: 2.3 GHz– Memory: 512 MB– OS: Windows XP
• Run time (wall clock)– 3 MRHT runs each with 100,000 fittings
• One run for center estimation, two runs for inner and outer boundary detection
45 seconds for each image
Conclusion
• Shape-Based Fitting
– Goal: Better-Localizing Auroral Oval
– Upcoming:
• Couple w/ k-means (Hung and Germany, 2003)
• Longer Term:
– Auroral Classification
– Temporal Evolution Tracking