Predictive Unmanned Persistent Surveillance for Harbor Security Peter Wubbels

1
Predictive Unmanned Persistent Surveillancefor
Harbor SecurityPeter Wubbels
2
The Problem

• Monitoring security systems requires constant
  vigilance
• Amount of incoming data to analyze is enormous
• Difficult for a single human user to track both
  local and global phenomena
• Honolulu Harbor

3
Our Approach

• Generalized, adaptive framework for intelligently
  modeling data

PUPS
Sensors
Interface
Feature Extraction
Local
Knowledge DB
Local
Algorithmic Analysis
Remote
Predictive Model
Alert
4
PUPS Sensors

• Automatic Identification System (AIS)
• Ships broadcast identifying information,
  location, speed, cargo, etc.
• Two ceiling mounted cameras
• Each camera is connected to a DSP board
• Motion detection algorithms running on DSP board
• Sends back motion feature packets to PUPS system

5
PUPS Main System

• Feature Extraction
• Convert incoming data into generalized features
• AIS Ship location, speed, heading, size, type
  and time of day
• Video Supplement AIS
• Knowledge Database
• Set of a priori rules that define known undesired
  behaviors
• Restricted Areas
• Unsafe combinations of ship speed and size

6
PUPS Main System

• Algorithmic Analysis
• Initially investigated Growing Hierarchical
  Self-Organizing Maps (GHSOM)
• Now using Continuous Density Hidden Markov Models
  (CDHMM)
• CDHMM
• Automatically build a model of state transitions,
  discretized to desired resolution
• Given an initial state, what are the probable
  next steps?
• Current model trained on two weeks of data.
  Contains 72,000 states, each with an average of 3
  transitions

Next State 1
.25
Initial State
.75
Next State 2
7
PUPS Interface
Anomalies
8
PUPS Interface
Prediction
9
PUPS Interface
Live Video
10
Current Work

• Real-world data is not clean
• Building and maintaining model
• Motion detection algorithm optimization
• Camera control
• Improving interface
• Predicting destinations
• Finding new ways to use framework
• Contacting experts in the industry

11
Future Directions

• Current work is done on a small harbor
• Imagine a much larger system, able to model paths
  on a global scale
• Compare past/most common paths with ships
  current path
• Flag anomalous/dangerous behavior
• Include other sensors and data
• Seismic buoys
• Depth charts
• Underwater microphones

12
Future Directions

• Model not only state transitions, but also
  patterns of behavior over much longer
  distances/times
• Expert systems trained from human analysis of
  naval data
• Predictive Analysis for Naval Deployment
  Activities (PANDA)

13
Questions?
14
Thank You