Dynamic Clustering for Acoustic Target Tracking in Wireless Sensor Networks

1
Dynamic Clustering for Acoustic Target Tracking
in Wireless Sensor Networks

• Wei-Peng Chen, Jennifer C. Hou and Lui Sha
• Department of Computer Science
• University of Illinois at Urbana-Champaign

2
Acoustic Target Tracking System

• Air dropped inexpensive acoustic sensors and
  wireless nodes for mobile vehicle tracking

3
Outlines

• Background knowledge
• Motivations for dynamic clustering
• Protocol skeleton
• Analysis simulation results
• Conclusions

4
Acoustic Localization Energy Based Approach

• Magnitude of signal decays with propagation
  distance exponentially

log of magnitude
log of distance (inch)
5
Acoustic Tracking Architecture

• Cluster structure is suitable to the tracking
  task
• Hierarchical sensor network A cluster consists
  of
• A cluster head (CH) high capability sensor
• Sensors
• Tracking steps within a cluster
• 1. CH sound detection
• 2. CH classification
• 3. CH broadcast REQ (energy signature)
• 4. Sensors matching and reply energy
• 5. CH localization
• 6. CH report the results to a sink

6
Outlines

• Background knowledge
• Motivations for dynamic clustering
• Protocol skeleton
• Analysis simulation results
• Conclusions

7
Problems for Static Clusters

• Static cluster
• Fixed membership, coverage area, size of cluster
• Not robust in terms of fault tolerance
• Cannot share data in different
• clusters
• Generate redundant results at
• different clusters
• Create contentions between
• different clusters

8
Dynamic Clustering Follow the Target
9
Dynamic Clustering Follow the Target
10
Challenges

• (I1) Only one CH (preferably closest to the
  target) is active
• (I2) Only a sufficient number of sensors respond
  to determine the target location when receiving
  REQ
• (I3) REQ, REP and report packets do not incur
  collisions
• Goal to mitigate the contention problem in the
  tracking system

11
Outlines

• Background knowledge
• Motivations for dynamic clustering
• Protocol skeleton
• Analysis simulation results
• Conclusions

12
Protocol Skeleton

• Distance Calibration and Tabulation
• Cluster Head Volunteering
• Sensor Replying
• Reporting Tracking Results

13
Phase I Distance Calibration and Tabulation

• Each sensor exchanges the position with its
  neighbor
• Each sensor constructs Voronoi diagram (locally)
  and response tables used in the tracking phase
• The first phase is only executed initially and
  the results are stored in the tables

14
Voronoi Diagram

• Each pair of energy readings from two sensors
  determines a half plane that contains the target
• The intersection area of all the half planes ?
  Voronoi diagram
• A sensor detects the maximal energy ? the target
  is in its Voronoi cell
• Voronoi diagram will be used in constructing the
  response table

Sensor A
Sensor B
15
Construction of Response Table at a CH

• Purpose each CH estimates the probability of a
  target closer to itself than other CHs, Pr(id)
• d estimate of the distance from the target to
  itself ?
• The possible location of the target is a circle
  with radius d
• The probability of CHi becoming active ? Pr(id)

16
Determining Pr(id) 3 Cases in Voronoi Diagram
17
Construction of Response Table for Sensors

• Sensor Sj determines Pr(jri?j) prob. of target
  closer to Sj than any other sensors or CHs
• Given ri?j , the ratio of energy from CHi to its
  own energy, the possible location for the target
  is a circle
• Use the same counting technique to determine
  Pr(jri?j)

18
Phase II Cluster Head Volunteering

• Goal select the CH closest to the target with
  high probability
• Twophase random delay based mechanism to
  implicitly determine a leader
• CHi set a back-off delay before sending REQ
• Two phase broadcast mechanism
• 1st Energy(short) 2nd Signature(long)
• During the back-off period, if a CH overhears a
  energy pkt with larger energy or a signature pkt,
  it cancels its transmission otherwise, ignores
  the overheard pkt
• To reduce as many potential competitor as
  possible in the first phase
• To avoid long signature packet get corrupted

19
Phase III - Sensor Replying

• A sensor Sj set a back-off delay, D, before
  replies
• When the timer expires, Sj replies if
• (i) its energy is largest among the overheard
  pkts
• (ii) it is one of the Voronoi nbr of the sensor
  reporting the largest energy
• To collect the replies around the Voronoi cell
  where the target locates

20
Phase IV- Reporting Tracking Results

• The active CH generates the localization result
  when
• (i) the timer expires or
• (ii) CH receives sufficient replies
• including the REP with max energy and REPs
  surrounding the max REP
• A simple localization technique take the
  position of the sensor with the largest energy as
  the estimation of the target

21
Outlines

• Background knowledge
• Motivations for dynamic clustering
• Protocol skeleton
• Analysis simulation results
• Conclusions

22
Two Simplifications in Analysis

• Back-off delay
• Square deployment

23
Analysis (Cont.)

• Number CHs as CH1, CH2 (d1ltd2ltltdN)
• 3 cases in the two-phase broadcast mechanism
• C1. CH1 sends energy pkt first
• C2. CHi, i gt 1 sends energy pkt first but CH1
  sends energy pkt successfully
• C3. Energy pkts of CH1 and CHi collide
• Lemma 1 C2 and C3 occur only if
• Lemma 2 in C2, CH1 will still become leader if
• Wran Wmin ?in C2, CH1 always becomes leader

24
Calculation of Probability of C3

• Use prob. of C3 as the upper bound that CH1 can
  not become the leader
• If Wmax0.1 sec, WranWmin0.1 msec, slot time
  20 us, ?
• Pr(C3) 5.3e-5

25
Simulation Scenario

• Using SensorSim from UCLA (built on ns-2)
• 36 CHs and 288 sensors in 180180m2 field
• One sink is at (0,0)
• Two deployment strategies Square Random
• Performance comparison
• Static cluster
• Full-fledged version of proposed approach
• 2 phases broadcast mechanism
• Back-off based on response table
• V.1 One phase and no response table
• V.2 No response table

26
Simulation Results-Square
27
Simulation Results-Random
28
Conclusions

• A light-weighted, self-organized dynamic
  clustering protocol for target tracking is
  proposed
• Only one cluster is formed with high probability
• We can effectively reduce contentions between
  sensors and render more accurate estimate results

29
(No Transcript)
30
Future Works

• Incorporate more accurate localization tracking
  methods
• Integrate dynamic clustering with information
  quality driven routing protocols
• We have implemented part of the protocol on a
  testbed using Berkeley motes and PC104

31
Acoustic Localization - Delay Based Approach

• use on-set (starting) time of signal to measure
  the propagation delay between a pair of sensors
• Use multilateration to do localization
• () No deed for microphone gain calibration
• () Can use less sensors
• (-) Require accurate time synchronization
• (-) Only can track gun-shut type of signal
• (-) Difficult to find the on-set point accurately

32
Acoustic Localization Energy Based Approach

• Magnitude of signal decays with propagation
  distance
• Compare readings at different sensors Maximum
  likelihood method
• Use the position of sensor closest to the target
  as the estimation of target position

33
Maximum Likelihood Method
34
Case 3 - Determining Pr(ir)

• For each point p on the circle with radius d
• if p in Voronoi cell of CHi
• gain
• else if p in Voronoi cell of CHk and p is not
  within inner circle of k
• loss
• Return gain/(gainloss)

CHk
35
Real Measurements from Motes
36
Testbed Setup
Acoustic Sensor Target/Tracker Base Router Sin
k Cluster
37
Results Single Cluster

• 9 sensors, each is separated by 18 in.
• Target at 22 positions, 10 tests for each
  position
• Avg. error 8.37 in
• In bound error ltlt out bound error