Automated Intruder Tracking using Particle Filtering and a Network of Binary Motion Sensors

1
Automated Intruder Tracking using Particle
Filtering and a Network of Binary Motion Sensors

• Jeremy Schiff
• EECS Department
• University of California, Berkeley
• Ken Goldberg
• IEOR and EECS Departments
• University of California, Berkeley
• http//www.cs.berkeley.edu/jschiff
• Supported by NSF Grants 0424422/0535218

2
Outline

• Introduction
• Related Work
• Problem Formulation
• Setup and Assumptions
• Particle Filtering
• Results
• Simulation
• Experimental
• Conclusion/Future Work

3
Motivation

• New class of technologies due to 9/11
• Automated Security
• Wireless Sensor Networks
• X10 PIR sensors - 25
• Robotic Webcams
• Pan, Tilt, Zoom
• 500 Mpixels/Steradian
• Increased computer processing speeds
• Enables Realtime Applications

4
Goal and Approach

• Wish to secure an environment
• Low Cost Binary Sensors
• X10 25
• Optical Beam
• Floor Pad
• Manufactured in China
• Noisy triggering pattern
• Refraction
• Use sensor triggering patterns to accurately
  localize an intruder

5
Intuition

• Utilize Sensor Overlap Information

6
Intuition

• Utilize Sensor Overlap Information

7
Outline

• Introduction
• Related Work
• Problem Formulation
• Setup and Assumptions
• Particle Filtering
• Experiments
• Simulation
• Real-world
• Conclusion/Future Work

8
Related Work

• Pursuer/Evader Games
• Using line-of sight optical sensors
• Isler, Kannan, Khanna 2004
• Avoid being seen by evader
• Bandyopadhyay et al. 2004
• Tracking Worn Devices
• Track Infrared Beacon
• Shen et al. 2004
• Dynamic Shipment Planning using RFIDs
• Kim et al. 2005

9
Related Work II

• Video Tracking Systems
• Micilotta and Bowden 2004
• Multiple Classes of Sensors
• Multiple exclusive modes
• Cochran, Sinno, Clausen 1999
• Fuse data of multiple sensor types
• Jeffery et al. 2005
• Automated Camera Control
• Song et al. 2005

Virtual Devices
Physical Devices
10
Related Work III

• Probabilistic Tracking Approaches
• Kalman Filtering
• Kalman 1960
• Extended Kalman Filtering
• Lefebvre, Bruyninckx, De Schutter 2004
• Particle Filtering
• Book Thrun, Burgard, Fox 2005
• Arulampalam et al. 2002

11
Related Work IV

• Multiple humans controlling a camera
• Song and Goldberg 2003
• Song, Goldberg and Pashkevich 2003
• Panorama Generation
• Song et al. 2005
• Art Gallery Problem
• Shermer 1990
• Urrutia 2000

12
Outline

• Introduction
• Related Work
• Problem Formulation
• Setup and Assumptions
• Particle Filtering
• Experiments
• Simulation
• Real-world
• Conclusion/Future Work

13
Setup and Assumptions

• Room Geometry
• List of nodes and edges

• Discretize space
• Discretize time

14
Setup and Assumptions II

• Intruder occupied world-space cell j in iteration
  
• Sensor i triggered during iteration

• Sensor i experienced refraction period in
  iteration

15
Setup and Assumptions III

• Three Conditional Distributions
• Trigger while experiencing refraction
• Trigger from intruder
• Trigger from no intruder

16
Output

• Estimated intruder location
• Objective
• Minimize error between ground truth and
  estimation.

17
Characterization

• Per sensor type
• Grid over sensor space
• Determine
• Refraction period
• False Negative Rate
• False Positive Rate

18
Deployment

• Convert to world-space
• Overlay grid
• Transformed point to Cells

19
Deployment II

• Determine potential non-zero characterization
  cells via convex hull
• Inverse Distance Weighting
• Interpolation according to distance
• Determines values for cells without readings
  inside convex hull

20
Particle filters

• Non-Parametric
• Sample Based Method (Particles)
• Particle Density Likelihood
• Tracking requires three distributions
• Initialization Distribution
• Transition Model (Intruder Model)
• Observation Model
• Determines

21
Example
22
Example
23
Intruder Model

• State
• Position, Orientation, Speed, and Refracting
  Sensors
• Euler Integration for position
• Gaussian Random Walk for new speed and
  orientation
• Orientation change inversely proportional to
  speed
• Deterministic refraction periods
• Rejection Sampling to enforce room geometry

24
Intruder Model II

• Time between iterations
• Empirically determined constants

25
Intruder Model - Example
Example state at iteration 0
26
Intruder Model - Example
Accepted state for iteration 1
27
Intruder Model - Example
Example state at iteration 1
28
Intruder Model - Example
Accepted state for iteration 2
29
Intruder Model - Example
Example state at iteration 2
30
Intruder Model - Example
Rejected state for iteration 2
31
Intruder Model - Example
Example state at iteration 2
32
Intruder Model - Example
Rejected state for iteration 2
33
Intruder Model - Example
Example state at iteration 2
34
Intruder Model - Example
Accepted state for iteration 2
35
Sensor Model

• Evidence is vector of which sensors are
  triggering
• Triggering of sensors independent given intruder
  state implies
• If sensor refracting
• Otherwise

36
Outline

• Introduction
• Related Work
• Problem Formulation
• Setup and Assumptions
• Particle Filtering
• Experiments
• Simulation
• Real-world
• Conclusion/Future Work

37
Simulation Setup

• 22 Optical Beams
• Perfect
• Optimal Performance

• 14 Triangular Motion Sensor
• Perfect Imperfect

38
Simulation Results

• Example Path
• Ground Truth
• Red Circles
• Estimations
• Grey Circles

39
Simulation Results II

• Baseline Estimate
• Perfect Optical-Beam Sensors

40
Simulation Results III

• Perfect Triangular Motion Sensors
• Imperfect Triangular Motion Sensors

41
Simulation Results IV

• Error over Time 4 Sec. Refraction, Imperfect
  Sensors
• Density - 8 Sec. Refraction, Imperfect Sensors

42
In-Lab Results

• 8 Passive Infrared Sensors
• X10
• 8 second refraction time
• Room 8x6 meters
• .3 m /Cell dimension
• Sampled every 2 seconds
• 1000 Particles

43
In-Lab Results II
44
Outline

• Introduction
• Related Work
• Problem Formulation
• Setup and Assumptions
• Particle Filtering
• Results
• Simulation
• Experimental
• Conclusion/Future Work

45
Conclusions

• Real-time Tracking System
• Binary Sensors with Refraction Period
• Particle Filtering for Sensor Fusion
• Conditional Probability Models
• Models
• Intruder Velocity
• Room Geometry
• Sensor Characterization

46
Future Work

• Effects of varying different components
• Number Particles
• Types of sensors
• Spatial arrangements of sensors
• Multiple intruders
• Decentralize
• Vision Processing
• Other applications
• Warehouse Tracking

47
Thank You

• Jeremy Schiff jschiff_at_cs.berkeley.edu
• Ken Goldberg goldberg_at_ieor.berkeley.edu
• URL www.cs.berkeley.edu/jschiff