Transcript SPIE_07_Submarine

Defense and Security Symposium 2007
A Sparse Undersea Sensor Network Decision Support System
Based on Spatial and Temporal Random Field
Dr. Bo Ling
Migma Systems, Inc.
Dr. Mike Traweek
Office of Naval Research
Dr. Tom Wettergren
Naval Undersea Warfare Center
April 10, 2007
Presentation Outline
- Problem Statement
- Random Field Estimation of Undersea Sensor Network
- Forward & Backward Mapping Mitigation
- Spatial & Temporal Layering Discrimination
- A Simulation Tool for Target Detection and Tracking
- Conclusion
Problem Statement
No Overlapping
In a passive submarine detection
system, we need to consider:
- Number of sensors required in
the sparse sensor network
- Probability of submarine
detection when a submarine
might be detected by only one
sensor at one particular time
instance
- False alarm mitigation
A Large Surveillance Region
Problem in Motion Based Detection
In a sparse sensor network, targets
can be detected and tracked over a
period of time.
As a target moves, although it may
not be detected by any sensors at
certain time instance, it will be
detected by a number of sensors
over a period of time.
When positive and negative reports
co-exist, it is difficult to determine if
targets actually exist!
Basic System Architecture
Although targets in a surveillance region may
not always be detected, as they move, sensors
can collectively detect, classify and track them.
Random Field for Undersea Sensor Network
A typical random field model can be
described by the following difference
equation:
Random Field Estimation
In random field modeling setting, it is always assumed that
Z = {Zij | (i, j)  } is stationary and Gaussian distributed
with known or empirical mean and variance.
Gaussian Random Field (GRF) or Gaussian-Markov
Random Field (RMRF) modeling has been widely used in
image processing.
In undersea sensor network, the assumption of stationary
process and Gaussian or Markov distribution needs to be
carefully verified.
Our Approach for Random Field Estimation
Based on the continuation of sensor outputs, the random field zf
over two consecutive samples must be similar provided that the
sampling rate is large enough or the sampling time is small enough.
Mathematically, suppose and
are the random fields at the
sampling time instants k and k+1. Then, based on the continuation
property, the quantity of ||
- || is small.
Minimize:

FBMM Technology (Patented)
Forward & Backward Mapping Mitigation
It can be used to reduce false reports
while keeping the positive reports.
A Simulated Network
2020 sq. miles
Detections after 2 hours
Two targets at 180 m/h
Both true and false detections
co-exist
200 sensors
Difficult to Find Targets Visually
What an operator
will see after twohour monitoring.
Are there any targets?
Random Field Estimation
Random field estimated using
our LMI-based optimization
Mathematical Morphological Operation
Morphological operators can be
applied to reduce false detections
Backward Mapping
Backward Mapping for refined
situation awareness.
STLD Technology (Patented)
Spatial & Temporal Layering Discrimination
STLD technology applies temporal patterns
to further reduce the false reports.
Gap Statistics Based Clustering
Gap Statistic (Robert Tibshirani, Guenther Walther, Trevor Hastie,
“Estimating the number of clusters in a data set via the gap statistic”, J.
R. Statist. Soc. B (2001), 63, Part 2, pp. 411-423) is a technique used to
estimate the number of clusters in the data.
Individual Clusters in Temporal Patterns
Temporal & Spatial Mitigation
Detections in each cluster are
checked for both temporal and
spatial patterns.
Temporal Pattern - True
detections must show temporal
trend
Spatial Pattern - True
detections must be relatively
close spatially
False Reports Reduction Using STLD
After FBMM Processing
After STLD Processing
STLD
FBMM
Combine FBMM & STLD
Target Synthetic Tracks
Groundtruth
Synthetic Tracks
A Simulation Tool
Conclusion
- In a sparse undersea network, targets can be
detected and tracked based their motion
- The mixture of both positive and negative reports
makes it difficult for an operator to determine
whether or not targets are actually present
- Our patented Forward & Backward Mapping
Mitigation technology and Spatial & Temporal
Layering Discrimination technology can help
operator make better decisions