Improved PIR Target Localization in Region Based Distributed Sensor Networks

1
Improved PIR Target Localization in Region Based
Distributed Sensor Networks

• Ahtasham Ashraf

2
Outline

• Introduction
• UW CSP Approach
• Basics of PIR Detection/Localization
• Motivation
• Table Based Approach
• Results
• Conclusions

3
Introduction

• A framework for performing PIR Target
  Localization over a distributive, ad hoc wireless
  sensor network is presented.
• Traditionally, if a PIR sensor reports a target
  detection event, the location of the target will
  be positioned at the intersection of the
  line-of-sight of the PIR sensor and the road.
• This coordinate can be computed in advance and
  stored in a table to be looked up during run
  time.
• A potential drawback of this approach is that
  when more than 1 PIR detection are made at the
  same time, it would be difficult to choose an
  appropriate target location.
• A novel target localization method using
  directional polarized infrared (PIR) sensors is
  given in this work.
• In this work, we developed an empirical PIR
  localization method using ground truth of
  training data
• We have implemented this algorithm and compared
  the PIR localization and tracking results using
  real-life sensor network time series.

4
UW CSP Approach

• IDEA
• Track multiple targets moving through a
  distributed sensor network using multiple
  modalities.
• Sensor Network is divided into regions
• Dynamically or statically created
• May overlap with each other.
• Multi-Modal Detection
• Active region Sensors performs the native multi
  modal detections.
• Classification
• Time series used to classify targets.
• Helps in Multiple Target Tracking.
• Localization
• Manager Nodes perform MM Localization
• Tracking
• Makes Predictions Current Position Estimates
  from Localization results.
• Region Management
• New regions are created as the target escapes the
  current region.

5
Basics of PIR Det/Localization

• Heat generating Objects generate Infrared
  radiation. Strongest at a wavelength of 9.4mm.
• The Pyroelectric (PIR) sensor is made of a
  crystalline material that generates electric
  charge when exposed to heat in the form of
  infrared radiation.
• constant false alarm rate (CFAR) detection method
  for detection from the energy time series.
• ?(n) Mean , ?(n) variance of received energy
  y(n) and at time n, then threshold
• ?(n) ?(n-1) C . ?(n-1)
• C constant chosen to yield constant PFA

6
Basics of PIR Det/Localization

• Perpendicular Projections of the sensor node
  positions onto the road gives the target location
  at time.
• Can be implemented as Pre-Computed Table.
• Technique is prone to a few of problems even in
  the presence of single target.
• ? If Multiple Sensors detect a single target, the
  position estimate becomes ambiguous.
• Multiple reporting sensors can be widely spaced
  apart as well.
• ? One can get a lot of false detections due to a
  number of factors.
• Heat generating body crossing.
• Bad Sensor.

7
Motivation

• We assume perfect Sensor Orientation.
• There can be Multi-Sensor Detections for a single
  target.
• This enables us to give more precise estimate of
  target position due to overlap.
• More Sensor Detections for same target mean more
  accurate Estimate.
• This also implies smaller Measurement Covariance
  Matrix (R).

8
Table Based Approach
 

• We have used the Spatial and Multi-Modal
  information.
• Gives more accurate estimates.
• Provides us a probability measure for the
  estimate of each PIR detection pattern.
• Gives Single position estimate for multiple PIR
  sensor detections.
• Uses Training Data.
• Sitex02 Data
• Single vehicle and matching ground truth data for
  15 sensors.
• Runs AAV3, AAV6, AVV9, DW3, DW6, DW9, DW12.
• We know the sensor positions, road coordinates
  and time stamped ground truths from GPS.
• But this ground truth(? 10m) is not used in the
  localization process at the run time.
• Approach
• The idea is to put some constraints on the
  spatial separation of detecting sensor position
  and the vehicle ground truth. If they are widely
  spaced apart then its a false detection.
• If acoustic modality (with a good PD) is not
  showing any detection, it means that the vehicle
  is at least 30-40 meters away from the current
  sensor and the PIR detection is false.

9
Table Based Approach

• Look at all possible patterns of PIR detections
  for all sensor nodes
• and accumulate the ground truths corresponding
  to each pattern for all the sets of data
  mentioned above.
• These GT are weighed to get the final position
  estimate (X,Y).
• The last column is the Covariance Matrix for the
  ground truths available for that pattern.
• In some cases there will be no (X,Y) estimate
  available as that particular pattern never
  happened in the training data.
• Similarly, there will be (X,Y) estimates
  available for some rows with more than a single
  1 due to some sensors which are quite close to
  each other.
• But this multiple occurrence of 1 can also
  happen due to noisy detection events.
• There are a few valid patterns incase of a single
  vehicle situation and the rest are results of
  spurious detections.

 Table 1. Detection Patterns and the Localized
position estimate.
10
Time Series Plots Detections
Sensor Orientation Problem This is the ideal
Orientation
Result of Bad Orientation
False PIR Detections
PIR, Acoustic Energies, PIR Detections and Ideal
Detections for the 15 sensor nodes of AAV3 data
set. Its clear that at later stage the PIR shows
false detections for sensor 6, but its negated by
Acoustic Modality
11
(No Transcript)
12
Probability Measure of correctness

• We have come up with a probability measure of the
  correctness of the PIR detection decision.
• So for the situation when we have a positive PIR
  detection we can arrange the results into a
  matrix such that

Four Possibilities P(Target/Acoustic Det) N4/(
N4 N3 ) P(Target/ No Acoustic Det) N2/( N2
N1 ) P(Noise/Acoustic Det) N3/( N4 N3
) P(Noise/ No Acoustic Det) N1/( N1 N2 )
13
Sensitivity Specificity

• Similarly, for a set of data, we can find out the
  Sensitivity and Specificity of the PIR detection
  results. For this particular setup we can
  similarly come up with a matrix

 
Now we have Sensitivity (N4/(N4 N2))
x100 Specificity (N4/(N4 N3)) x100
 
These are a good measure of looking at the
detection results. The goal is to achieve 100 of
these measures.
14
Results
Histograms of PIR Loc Errors (meters) using Old
and The Table based Approach.
I have observed some very good improvements
against large PIR localization errors, besides
reducing the small ones as well.
15
Conclusions

• Advantages
• Higher accuracy.
• Simple implementation.
• Immune to noise. 
• It also promises some future framework for
  research in this direction.
• We plan to extend this approach for the case of
  multiple targets.
• We believe that it can greatly help in
  automatically indicting the presence of multiple
  targets in a region, which of course has been one
  of the main issues in the Sensor Network Signal
  Processing.
• The use of PIR Table Based Approach also opens a
  gateway of research in the direction of Region
  Detection.
• If the PIR detections are made more reliable then
  these can greatly help in reducing the false
  Acoustic Detections, which occur due to
  inevitable noisy acoustic time series