Upload
doantu
View
216
Download
2
Embed Size (px)
Citation preview
Insight
National Academies Board on Global Science and Technology
Committee on Integrating Humans, Machines and Networks:
A Global Review of Data-to-Decision Technologies
Washington, D.C.
February 26, 2013
Ben Cutler
Program Manager
Distribution Statement A Approved for Public Release, Distribution Unlimited
Distribution Statement A Approved for Public Release, Distribution Unlimited 2
• Combine data across multiple sources
• Manage efficient use of sensors and platforms
• Identify threats using behavioral discovery and prediction algorithms
• Collaborate
Insight: one unified global ISR picture
An adaptable, integrated human-machine ISR exploitation system
Doppler
Distribution Statement A Approved for Public Release, Distribution Unlimited
3
• Data from diverse sources
• Ambiguity/uncertainty of information
• Who? What? Where? When? Why?
• Pedigree, provenance
• Report of an observation (delivery of electrons) may be received long after the event
• Sequence neutrality – obtain the same results regardless of the order in which data is integrated
• Tactical relevance requires real-time insights
• Real-time analysis
• Real-time data sharing
• Data analytics at massive scale
• High computation requirement per unit data
• Fleeting existence of a critical piece of information buried in a sea of data
• “Bad guys hiding in the noise floor”
Challenges
4
• Threat networks include the following elements
• Movement
• Command and Control
• Intelligence (includes collection and deception)
• Logistics
• Fires (IEDs, artillery, air defenses, aircraft)
• Protection (physical defenses, air defenses, electronic systems)
• Identification by
• Signatures
• Physical characteristics indicate this aircraft is a fighter jet
• Activity Based Intelligence
• Pattern of life (e.g., activities, associations) indicate a group of people appears to be an insurgent cell
Finding threat networks
Distribution Statement A Approved for Public Release, Distribution Unlimited
Vast sea of data
Find the threat network (all names fictional)
5 Distribution Statement A Approved for Public Release, Distribution Unlimited
6
• Correlate unusual activity with historical information and intelligence estimates
• Determine past and current locations of people of interest, vehicles, and other artifacts, and estimate patterns of life
• Detect deviations from normalcy that warrant further analysis
• Set up tripwires or watch boxes that alert when an event partially or fully matches a pattern of interest based on some combination of space, time, and network relationships
• Alert the analyst to any meetings in the south end of town
• Task assets to fulfill information requirements
• And alert the analyst when an information requirement (information that is needed for understanding or for making timely decisions) is fulfilled
• Situational analysis: what has happened, what is happening, what is likely to happen?
Some system requirements for finding insurgent networks
Distribution Statement A Approved for Public Release, Distribution Unlimited
Distribution Statement A Approved for Public Release, Distribution Unlimited 7
Data to information: VIRAT interpretation of video
Activity types Digging Carrying Walking
Video data collected from Creech AFB, March 2009
8
Data to information: wide area motion imagery example
• 25 cm GSD
• Six cameras with ortho-rectified (stitched and geo-
registered) imagery
• NITF file format
with encoded sensor metadata
Distribution Statement A Approved for Public Release, Distribution Unlimited
AFRL RYA LAIR Dataset, 21 October 2009 WAMI Data
9
Tracking result
Wide area
motion imagery example
Tracking result
Distribution Statement A Approved for Public Release, Distribution Unlimited
10
Relevant information content (ground truth)
~6,500 tracks in 7 minutes
Distribution Statement A Approved for Public Release, Distribution Unlimited
11
• Raw sensor data
• ~200 megapixels/second
• Pixel format and interpretation is sensor-specific
• Generally, we just want information about the entities in the video
• Where are the cars and people?
• Features (e.g., large, red sedan)
• Activities – what are they doing? (e.g., car making a u-turn)
• Information: “tracks”, one per active entity
• In this example: 200 megapixels/second => 15 tracks/second
• Tracks distill data into computable information that Insight can process
• Track format accepted by Insight is standard across motion imagery sensors
• Model-based presentation allows downstream algorithms to be sensor-agnostic
• Track information may include feature or activity information
• Very large wide area motion imagery sensors
• 10s of gigapixels/second
• 10+ kilotracks/second
Data to information: wide area motion imagery
Distribution Statement A Approved for Public Release, Distribution Unlimited
• Real-time information integration and automated exploitation
• Multi-echelon, integrated, theater through tactical unit level support
• Sensor ISR: space, air, and ground
• Other information sources: soldier information, reports, CI/HUMINT, SIGACTs, IPB,
demographics, friendly locations (BFT), OSINT, law enforcement, commercial sources, etc.
Diverse information sources
12
• Automated resource
management for
dynamic tasking and
cross-cueing based on
intelligence
requirements
• Direct support
systems
• Organic systems
• General support
systems
Distribution Statement A Approved for Public Release, Distribution Unlimited
Insight World Model
14
Object types provided by Intel Object types created by reasoning components
Distribution Statement A Approved for Public Release, Distribution Unlimited
15
Data conditioning
• Physical sensors have unique
reference frames – space/time bias
• Observer location, perspective
• Observed entity locations
• Incomplete data
• Obscuration
• Omission (e.g., HUMINT uncertainty)
• Duplicate data
• Multiple tracks or detections for a single entity
• Conflicting elements of a textual report
• Conditioning
• Provides a common reference frame
• Estimates incomplete or conflicting data
4
Performance example - Blue Devil HDBlue Devil video frame overlay onto PeARL foundational image map
Using raw sensor metadata
Example - Blue Devil HDBlue Devil video frame overlay onto PeARL foundational reference map
Unregistered - using raw sensor metadata
PeARLReference
Blue Devil Unregistered
5
Blue Devil Registered
PeARLReference
Example - Blue Devil HDBlue Devil video frame overlay onto PeARL foundational reference map
Registered - using CMREG corrected metadata
Blue Devil HD video
frame overlay on
PeARL foundational
reference map
(using raw sensor metatdata)
(using corrected metadata)
unregistered error
registered
Simulated platform
trajectories (with biases)
-250 -200 -150 -100 -50 0 50 100 150 200 250-200
-150
-100
-50
0
50
100
150
200
X (km)
Y (
km
)
lon
git
ud
e
latitude -250 -200 -150 -100 -50 0 50 100 150 200 250
-200
-150
-100
-50
0
50
100
150
200
X (km)
Y (
km
)
latitude
lon
git
ud
e
Simulated platform
trajectories (with biases mitigated)
simulated platform Distribution Statement A Approved for Public Release, Distribution Unlimited
16
An example with biases 1 vehicle, 3 sensor tracks with biases
Tracks of the same vehicle
produced by 3 geographically dispersed ground-based GMTI radars
1. Beacon 2. Harry 3. Quarry
Distribution Statement A Approved for Public Release, Distribution Unlimited
An example with biases mitigated Track fusion after data conditioning
4 radars track different segments of a road network
Fusion provides a clear picture of vehicle activity on the road network
17
Tracks from 4 geographically dispersed ground-based GMTI radars
1. Beacon
2. Harry 3. Quarry 4. Brigade
Fused tracks
Distribution Statement A Approved for Public Release, Distribution Unlimited
18
Dynamically-generated traffic patterns provide context for comparing against real-time intelligence
• Traffic normalcy modeling
• Track starts, stops, and velocity distributions
• Vehicle/people track counts and
activity
• Compare to real-time fused tracks to detect anomalies
• Unusually fast vehicles or isolated track start/stops
• Unusual vehicle/people activity
• Correlate with other information sources (e.g., historical
information/SIGACTs, other sensor observations, etc.)
Traffic Pattern Analysis
All fused tracks Dynamically-generated traffic patterns
Heavy traffic Moderate traffic Light traffic True detections of isolated track starts/stops
False alarms of isolated track starts/stops
Detections of an isolated track start/stop
True detections of track starts/stops
Distribution Statement A Approved for Public Release, Distribution Unlimited
• EDL is the entry point for source information (e.g., HUMINT, SIGACTs, network information)
• EDL produces enriched graph fragments
• Brings in context information to enable correlation of source data
• Discovers roles and links
• Processes data (tracks, locations,
associations) and metadata (probabilities, pedigree)
• Enriched graph fragments are used to build up the dynamic graph
• Insight and 3rd party applications can process enriched data
• Analysts are able to navigate
source evidence in a broader context
EDL enriches source information to build up linked evidence
19
Element Discovery & Labeling (EDL)
Meetings Known associations
Tracks Communications
Enrich
ed a
bst
ract
ion levels
Distribution Statement A Approved for Public Release, Distribution Unlimited
Reasoning engines enable top-down exploration and bottom-up alerts to activities of interest
20
Analyst-vetted cases
New anomaly
Analyst-vetted cases
Resolved anomaly
Bottom-up inference management
Top-down hypothesis management
Patterns Pattern matches Hypotheses
• Increases analyst ability to
manage hypotheses derived
from uncertain data
• Continuously monitors evidence
streams to detect analyst-
defined patterns of interest
• Detects emerging threats
buried in multiple, complex,
high-volume data streams
• Explains patterns discovered in
the data
• Prioritizes and explains
anomalies using analyst-vetted
cases
• Focuses analyst attention on
anomalies of
importance/relevance
Evidence stream
Distribution Statement A Approved for Public Release, Distribution Unlimited
Analyst workstation at FT-1
21 Distribution Statement A Approved for Public Release, Distribution Unlimited
Distribution Statement A Approved for Public Release, Distribution Unlimited 22
FT-1: geospatial, network, and evidence manager views
Efficient Overview, Alerting, and Navigation
23
• Alert overview provides a summary of open information requests, their status, and related
information products • Incoming alerts allow rapid drill down to more detailed information views
Distribution Statement A Approved for Public Release, Distribution Unlimited
Lightweight evidence management tools
24
• The analyst can rapidly organize, mark up, and revise hypotheses and connections
between entities and activities
• Touch-point for seeding and reviewing results of automated reasoning capabilities
Distribution Statement A Approved for Public Release, Distribution Unlimited
25
Insight field test data volumes (GB)
132.7 TB database available at https://www.vdl.afrl.af.mil/ as the Global ISR project site Distribution Statement A
Approved for Public Release, Distribution Unlimited
26
Community use of Insight field test data
Unique, high-fidelity truthed data sets that have been made available to ISR researchers across
the Defense and Intelligence communities – 475 researchers as of January 2013
Industry 339
FFRDCs & National Labs (MIT/LL, LLNL, SDL)
18
IC (NGA) 3
JIEDDO 1
DARPA 13
Air Force (AFIT, AFLCMC, AFRL, ASC, ESC)
86 Navy (ONR, SPAWAR, NSWC) 8
Army (ARL, CERDEC I2WD, CERDEC NVESD)
7
Breakout of Global ISR Project Site Users
Distribution Statement A Approved for Public Release, Distribution Unlimited
27
• Data from diverse sources
• Ambiguity/uncertainty of information
• Who? What? Where? When? Why?
• Pedigree, provenance
• Report of an observation (delivery of electrons) may be received long after the event
• Sequence neutrality – obtain the same results regardless of the order in which data is integrated
• Tactical relevance requires real-time insights
• Real-time analysis
• Real-time data sharing
• Data analytics at massive scale
• High computation requirement per unit data
• Fleeting existence of a critical piece of information buried in a sea of data
• “Bad guys hiding in the noise floor”
Summary
Distribution Statement A Approved for Public Release, Distribution Unlimited
www.darpa.mil
Distribution Statement A Approved for Public Release, Distribution Unlimited 28