ACDC CLUSTERING DECLUTTERING SUMMARY Outline  Automated Change Detection and Classification (ACDC) System  Computer-Aided Detection (CAD), Classification (CAC), Search (CAS), and Change Detection.  Clustering  NRL 6.2 FY05 New.

Download Report

Transcript ACDC CLUSTERING DECLUTTERING SUMMARY Outline  Automated Change Detection and Classification (ACDC) System  Computer-Aided Detection (CAD), Classification (CAC), Search (CAS), and Change Detection.  Clustering  NRL 6.2 FY05 New.

ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Outline
 Automated Change Detection and
Classification (ACDC) System
 Computer-Aided Detection (CAD),
Classification (CAC), Search (CAS),
and Change Detection.
 Clustering
 NRL 6.2 FY05 New Start
 Automated declutter mechanism for
electronic displays
 Summary
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
ACDC
 Ability to automatically detect /
classify / identify objects in imagery
and perform change detection
 Current project: Side-scan imagery
(SSI) for mine counter-measures
(MCM)
 Future/potential applications




Real-time Imagery in Cockpit (RTIC)
ECDIS
Weather / meteorological
Common Operational Picture (COP)
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Change Detection (using SSI)
1. Detect seafloor features
(shadows, bright spots)
2. Classify detections
(mines, rocks, sand waves, etc.)
3. Search historical database (position error)
4. Match new feature (to ideal features)
5. Perform area matching (uses clustering)
6. Identify features that don’t match:
change detection
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Computer Aided Detection
MILECs found by:
 Bright spots
 Sizes
 Shapes
 Type of Mines
Mine-Like Echoes (MILECs)
Automatically detected in SSI
 Shadows
 Length (look angle)
 Correct side
 Proximity to bright
spot
ACDC
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Real-Time CAD
SSI stored in UNISIPS format as separate records (Lat/Lon, Altitude, Heading).
Thresholded using a hard limiting transfer function to
find Bright Spots:
a = func(n) | a = 1 if a > min_bright_spot_value
000000000000000000000000000000
000000000000000000000000000000
000000000000000000000000000000
000000000000000000000000000000
000000000000000000000000000000
000000000000000000000000000000
Thresholded using a hard limiting transfer function to
find Shadows:
Bright Spot Bitmap
00000000000000000000000
00000010000000000000000
00011100000000000000000
00001100000000000000000
00000000000000000000000
Shadow Bitmap
00000000000000000000000
00000000000000000000000
00000001111111111100000
00000001111111111100000
00000000000000000000000
a = func(n) | a = 1 if a < max_shadow_value
Shadows & Bright Spots Marked
ACDC
ACDC
CLUSTERING
DECLUTTERING
Computer Aided Classification
CAD SNIPPET
ADAPTIVE FILTERING
AUTO-COMPLETE
SUMMARY
ACDC
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Historical SSI Database
(historical survey area)
 Imagery
 Features
 Snippets
 Classification
 Attributes
ACDC
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Vector Searchable Database
(Handles position error!)
 CAD / CAC new features (N) in areas
where historical features (H) exist
 Populate search database with H’s
 Query search database for each N
“ANDing”
position error
ellipses
H2
N1 

Spatial Query
H
10
 H3
New: N = N1, N2, …, Nn
Historical: H = H1, H2, …, Hn
Results: N1 = H3 | H10
ACDC
ACDC
CLUSTERING
SUMMARY
DECLUTTERING
Wavelet Networks for
Feature Matching
rectangle
rectangle
rectangle
rectangle
rectangle
rectangle
rectangle
rectangle
rectangle
triangle
triangle
triangle
triangle
triangle
triangle
ask = wavelet coefficients
triangle
circle
circle
circle
circle
wsk = basis functions
circle
circle
circle
unknown
unknown
Training Set
EXAMPLE: N1
y(x,y) = Σ ask wsk(x,y)
Neural Network
H10
ACDC
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Wavelet Networks for Area
Matching
N1
N4
N2
N1
NEW
N3
N2
H2
N4
H5
Cluster
Region
H10
N3
New Survey Data
Historical Database
ACDC
CLUSTERING
SUMMARY
DECLUTTERING
Clustering for MCM
 NRL has transitioned new algorithm to NAVO
to cluster MILECs detected in SSI, in support
of MCM efforts.
I. Detect / classify / search for / identify mine-like objects
(MILECs) in SSI.
II. Cluster objects
into regions.
III. Smooth clusters,
calculate density.
3
MILECs
Clutter
/ km2
category
x<4
1
4 < x < 12
2
x > 12
3
1
2
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Clustering algorithm
 NRL algorithm clusters mine-like objects
detected in SSI
 NRL 7440.1 invention disclosed June 2003.
 Uses geospatial bitmapping technique
patented by Code 7440.1 in 2001
(U.S. Patent 6218965).
 Unique method of clustering objects in
2D / 3D space: computationally efficient,
single-pass, repeatable, operates
on user-defined space, autonomous.
Latitude (Y)
Longitude (X)
Collection of points in geographic space (here, 2D)
Represent points as a geospatial bitmap. Each bit containing a point
is “set;” all other bits are cleared. Bit size depends on scale.
“Grow” each set bit (representing each point) by setting the
surrounding bits to form a predefined expansion shape.
 Expansion shape dictated by data
characteristics and user requirements.
 Can use any shape that can be
mathematically defined.
 Size of expansion shape determines density
of resultant clusters.
Points that are geographically close to each other will grow
or cluster together. Result: new geospatial bitmaps.
Traverse each bitmap (in a consistent direction)
to get vertices.
Smooth the bitmaps by dropping vertices if the resulting
polygon still contains all the original points and has an
area equal to or less than the unsmoothed bitmap.
Result of final iteration.
Density of region = # original points / area of smoothed polygon.
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
FY05 6.2 New Start
 Internally funded by NRL 6.2 program
(FY05-07)
 After FY07, will need transition
sponsors to implement in fleet systems
(NGA / VVOD, NGA / DNC2, NAVAIR /
TAMMAC, others)
 Leveraging ongoing work from ACDC
and other projects
 Collaborating with Dr. Greg Trafton
(Engineering Research Psychologist)
Clutter
a confused multitude of objects
Clutter as it
pertains to this
project: when
additional
information
would result in
performance
degradation.
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
The Clutter Problem
 Our ability to collect data & “create” information
is outpacing our ability to use and visualize the
results. Clutter in all types of electronic displays
is a massive and rapidly escalating problem.
 Many researchers have documented link
between increased clutter and degraded
performance.
 E.g., in cockpit displays, visual clutter can
disrupt a pilot's visual attention, resulting in
greater uncertainty concerning target locations.1
1Aretz
(1988), Wickens (1993), Wickens & Carswell (1995)
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Problem (cont.)
 Pilots want ability to declutter displays (e.g., driven by
vector-based GIS-like databases) but such “flexibility
means integration complexity and added pilot workload.
Pilots should be flying, not building a map!" 1
 "If the map display is too cluttered, I just turn it off!" 1
 Automated decluttering requires good clutter metrics.
Reasonably good clutter metrics exist for text displays,
but not for graphical displays.
 We propose to apply NRL detection / clustering
algorithms and human factors principles to quantify
clutter in electronic displays.
1 Quotes of F/A-18 pilots, from Lohrenz, et al. (1999)
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Relevance to COP
charts
imagery
obstructions
terrain
NGA
data
recon.
routes
troops
targets
C4I
data
Over the past decade, the COP has grown in functional complexity … (DISA, 1998)
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Current declutter method
Sample display downloaded from Nobeltec company website.
 Warfighter must manually remove an entire layer at once (brute-force filtering).
 Need a more “intelligent” way to declutter electronic displays.
 Warfighters and decision-makers should see all that is needed – but only what
is needed – without extraneous data obscuring critical information.
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Measuring Clutter
Considerations:
 Display type: aviation, meteorological, ECDIS, etc.
 User expertise: novice vs. expert
 Task: read-off, integrate, infer, working-memory
Incorporate established cognitive theory into
new / enhanced clutter metrics:





“Global” vs. “local” information density
“Salience” (e.g., M. Zuschlag, 2004)
#Colors, color contrasts among adjacent objects
#Features of each “type” (points, lines, areas, text)
#Clusters of features, cluster density - NRL algorithm
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Clustering Algorithm
 Expand technique to cluster objects in N x 2d
(2d geospatial location + other attributes)
 Primary challenges:
 Mathematically define meaningful expansion shapes
for feature layers and attributes.
 Determine how clusters of various display feature types
(points, lines, areas, text) interact with each other.
 Apply established theories of human visual attention
and search strategies to our methodology.
 Bound the problem: focus on clustering features
in NGA Vector Product Format (VPF) databases
 Standard database for many DoD applications
 Can be tailored with mission-specific data sets
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
Validate Clutter Metrics
 Can our metrics predict good display design
(subjective / user preference)?
 Interview technical experts
 Compare clutter metrics with subjective evaluations
 Can our metrics predict user performance and
workload? Requires experimentation …
 Independent variables
•
•
•
•
Display type (aeronautical; nautical; meteorological; etc.)
Clutter metric (uncluttered  very cluttered)
User expertise (novice, expert)
Task performed (read-off, integrate, infer, working memory, etc.)
 Dependent variables
• Performance: time, accuracy, method/logic (e.g., w/ eye-tracker)
• Workload: subjective evaluation, secondary task, pupil dilation
ACDC
CLUSTERING
DECLUTTERING
SUMMARY
SUMMARY
Summary
 NRL developing ACDC system with algorithms to
autonomously detect, classify, and cluster mine-like
objects in SSI, and perform change detection via
historical contact databases.
 ACDC concepts/functions applicable to other types
of imagery and objects  COP.
 Detection and clustering algorithms will be exploited
for new NRL project (FY05-07) to develop clutter
metrics for electronic displays.
 Will attempt to validate clutter metrics by comparing
with measures of user performance and workload,
and subjective evaluations of display design.
 Results should be significant for future geospatial
databases, db upgrades and display designs  COP.