Sheila Vaidya 
Deputy Program Director, Defense Programs, 
Office of Strategic Outcomes 
Lawrence Livermore National Laboratory 
This work performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.
Small 
country 
LA basin 
Large city 
1012 
1011 
1010 
109 
Lawrence Livermore National Laboratory LLNL-PRES-559252 
2 
City block 
1GB/min 1TB/min 1PB/min 
108 
107 
106 
66Mpix 2005 => Army, 
AF 
DOE 
2007 
DOE 
2004 
DOE 4Mpix, 
2Hz, 2003 
Small city 
MIT LL 2Hz, 
2009 
DARPA 1.8Gpix, 10Hz, 
2010 
HDTV 
m2 = # pixels 
LLNL-led wide area surveillance camera development trajectory
Lawrence Livermore National Laboratory LLNL-PRES-559252 
Constant Hawk ~ 
200 Mpix/sec 
~25 sq. km at ~ 0.5m GSD 
~100 Mbps 
 Objective: 
Video analysis at high resolution over entire field of 
regard, for anomaly detection and profiling 
 Challenges: 
 Limited flight/ground storage 
 Poor signal strength 
 Minimal platform computing resources 
 Low communication bandwidth to ground 
 << eyes-on-ground to distill information 
State-of-the-art: 
Still compression JPEG-2000 5-10 X lossless 
Video compression H.264 20-50 X lossy 
Analysis algorithms embryonic 
In Theater 
~12 Mpixel 
Can we exploit spatial and temporal coherency in video to assist data 
compression and automated analysis? 
3
Lawrence Livermore National Laboratory LLNL-PRES-559252 
4 
Innovative video conditioning and stabilization for orders of magnitude data reduction, coupled to 
activity detection and tracking 
 Compression: 100 –10,000X data reduction over raw collections 
 Demonstrated 1000X (TBytes GBytes) 
 Scalability: Adaptively assigned computational resources to regions of interest 
 Hardware : Ground station cluster embedded on-board processing, optimized for size, weight 
and power 
 Resolution and Accuracy: Geo-located, super-resolution motion 
imagery on “stationary” background 
 Output: Areas classified by content, with vehicles/dismount tracks and 
track behavior analysis 
 Query: Fast real-time interactive search, query, and 
playback (TiVo-like)
Lawrence Livermore National Laboratory LLNL-PRES-559252 
5 
Conventional camera stitch Subpixel merge 
Static corrections Dynamic, variable shape 
corrections 
Large residual-from-background with 
static-only corrections
Lawrence Livermore National Laboratory LLNL-PRES-559252 
6
10x20x10 = 2000x background compression 
Lawrence Livermore National Laboratory LLNL-PRES-559252 
7 
θ 
θ 
Persistent flight 
loops:>10x re-use 
of background 
80 background 
frames per quadrant 
compress to 4 cubic 
control images: 20x 
reduction 
Standard encoders 
compress 10x spatially 
<0.1% mover pixels 2000x compression; 
Net ~ 1000X data shrinkage: data dependent
ROI Selection Vehicles ROI’s: Automatically Generated 
Lawrence Livermore National Laboratory LLNL-PRES-559252 
8 
Analyst Prioritized/ Vehicles – high 
Roads – medium 
Buildings/Field – low 
- Scene segmented
Stabilized Imagery Tracks Activity/Events 
NW 
Threats* 10 
Events* 103 
Tracks* 106 
Geo-regist. Pixels 1013 
Lawrence Livermore National Laboratory LLNL-PRES-559252 
9 
1 
100 200 300 400 500 600 
50 
100 
150 
200 
250 
300 
350 
400 
450 
Analyst Cueing 
Threat Assessment 
• Activity classification and scenario discrimination 
• Visual template for merging of multi-sensor data 
2 
100 200 300 400 500 600 
50 
100 
150 
200 
250 
300 
350 
400 
450 
20 
100 200 300 400 500 600 
50 
100 
150 
200 
250 
300 
350 
400 
450 
(state | state ) t 1 t P  
Raw Pixels/ Day 1015 
Forensics 
Predictive
Single Mission View: Handoff Detection 
Lawrence Livermore National Laboratory LLNL-PRES-559252 
10 
Multiple Mission View: Complex Queries 
With filters for 
• Detecting activity 
• Removing spurious tracks 
• Height mapping 
• Time stamping events 
• …
Lawrence Livermore National Laboratory LLNL-PRES-559252 
11 
Pattern-of-life 
using traffic 
heat-maps
Lawrence Livermore National Laboratory LLNL-PRES-559252 
12 
Wide area video data processing for forensic backtracking and real time anomaly detection 
 Orders of magnitude data reduction 
 Spatially accurate , high resolution video exploitation and activity analysis 
– Mapped onto 3D terrain 
 On-board video analysis and streaming 
– Merged multi-sensor data product playback at the desktop 
 Rapid search architecture with network extraction & pattern of behavior from 
– Varying spatial and temporal time domains 
 Potential application areas: 
− Coastal/Border Surveillance, Law Enforcement, Disaster Imagery, Wideland Firefighting, 
Search and Rescue, Large Crowd Events … 
Single image Multiple images

Persistics: Wide Area Surveillance & Analysis by Sheila Vaidya

  • 1.
    Sheila Vaidya DeputyProgram Director, Defense Programs, Office of Strategic Outcomes Lawrence Livermore National Laboratory This work performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.
  • 2.
    Small country LAbasin Large city 1012 1011 1010 109 Lawrence Livermore National Laboratory LLNL-PRES-559252 2 City block 1GB/min 1TB/min 1PB/min 108 107 106 66Mpix 2005 => Army, AF DOE 2007 DOE 2004 DOE 4Mpix, 2Hz, 2003 Small city MIT LL 2Hz, 2009 DARPA 1.8Gpix, 10Hz, 2010 HDTV m2 = # pixels LLNL-led wide area surveillance camera development trajectory
  • 3.
    Lawrence Livermore NationalLaboratory LLNL-PRES-559252 Constant Hawk ~ 200 Mpix/sec ~25 sq. km at ~ 0.5m GSD ~100 Mbps  Objective: Video analysis at high resolution over entire field of regard, for anomaly detection and profiling  Challenges:  Limited flight/ground storage  Poor signal strength  Minimal platform computing resources  Low communication bandwidth to ground  << eyes-on-ground to distill information State-of-the-art: Still compression JPEG-2000 5-10 X lossless Video compression H.264 20-50 X lossy Analysis algorithms embryonic In Theater ~12 Mpixel Can we exploit spatial and temporal coherency in video to assist data compression and automated analysis? 3
  • 4.
    Lawrence Livermore NationalLaboratory LLNL-PRES-559252 4 Innovative video conditioning and stabilization for orders of magnitude data reduction, coupled to activity detection and tracking  Compression: 100 –10,000X data reduction over raw collections  Demonstrated 1000X (TBytes GBytes)  Scalability: Adaptively assigned computational resources to regions of interest  Hardware : Ground station cluster embedded on-board processing, optimized for size, weight and power  Resolution and Accuracy: Geo-located, super-resolution motion imagery on “stationary” background  Output: Areas classified by content, with vehicles/dismount tracks and track behavior analysis  Query: Fast real-time interactive search, query, and playback (TiVo-like)
  • 5.
    Lawrence Livermore NationalLaboratory LLNL-PRES-559252 5 Conventional camera stitch Subpixel merge Static corrections Dynamic, variable shape corrections Large residual-from-background with static-only corrections
  • 6.
    Lawrence Livermore NationalLaboratory LLNL-PRES-559252 6
  • 7.
    10x20x10 = 2000xbackground compression Lawrence Livermore National Laboratory LLNL-PRES-559252 7 θ θ Persistent flight loops:>10x re-use of background 80 background frames per quadrant compress to 4 cubic control images: 20x reduction Standard encoders compress 10x spatially <0.1% mover pixels 2000x compression; Net ~ 1000X data shrinkage: data dependent
  • 8.
    ROI Selection VehiclesROI’s: Automatically Generated Lawrence Livermore National Laboratory LLNL-PRES-559252 8 Analyst Prioritized/ Vehicles – high Roads – medium Buildings/Field – low - Scene segmented
  • 9.
    Stabilized Imagery TracksActivity/Events NW Threats* 10 Events* 103 Tracks* 106 Geo-regist. Pixels 1013 Lawrence Livermore National Laboratory LLNL-PRES-559252 9 1 100 200 300 400 500 600 50 100 150 200 250 300 350 400 450 Analyst Cueing Threat Assessment • Activity classification and scenario discrimination • Visual template for merging of multi-sensor data 2 100 200 300 400 500 600 50 100 150 200 250 300 350 400 450 20 100 200 300 400 500 600 50 100 150 200 250 300 350 400 450 (state | state ) t 1 t P  Raw Pixels/ Day 1015 Forensics Predictive
  • 10.
    Single Mission View:Handoff Detection Lawrence Livermore National Laboratory LLNL-PRES-559252 10 Multiple Mission View: Complex Queries With filters for • Detecting activity • Removing spurious tracks • Height mapping • Time stamping events • …
  • 11.
    Lawrence Livermore NationalLaboratory LLNL-PRES-559252 11 Pattern-of-life using traffic heat-maps
  • 12.
    Lawrence Livermore NationalLaboratory LLNL-PRES-559252 12 Wide area video data processing for forensic backtracking and real time anomaly detection  Orders of magnitude data reduction  Spatially accurate , high resolution video exploitation and activity analysis – Mapped onto 3D terrain  On-board video analysis and streaming – Merged multi-sensor data product playback at the desktop  Rapid search architecture with network extraction & pattern of behavior from – Varying spatial and temporal time domains  Potential application areas: − Coastal/Border Surveillance, Law Enforcement, Disaster Imagery, Wideland Firefighting, Search and Rescue, Large Crowd Events … Single image Multiple images