Collaborative Signal Processing

1
Collaborative Signal Processing

• CS 691 Wireless Sensor Networks
• Mohammad Ali Salahuddin
• 04/22/03

2
Introduction to CSP

• Need for CSP
• Features
• Distributed Processing
• Goal-Oriented on demand processing
• Information Fusion
• Multi-resolution Processing

3
Space-Time Sampling and Space-Time Cells

• Space and Time for useful information
• Moving object in a region corresponds to a peak
  in the spatial signal field that moves with time

4
Detection and Tracking

• Tracking a single target in a distributed network
• Routing protocols being used in the Sensor
  Networks
• Data-Centric
• Node-Centric
• Geographic-Centric (UW)
• Geographic region divided in spatial cells
• Some nodes are designated Manager nodes

5
Single Target
6
Single Target (contd.)

• Assumptions
• Potential target may enter the region via one of
  the four corners, four cells A, B, C and D
• Nodes in each of the four corner cells are
  activated to detect potential target
• Nodes run an energy detection algorithm, sampling
  at a priori fixed sampling-rate
• Suppose a target enters cell A

7
Single Target (contd.)

• Algorithm
• Some or all nodes in A detect the target, send
  the CPA and energy detector output to the manger
  nodes at N successive time intervals
• At each time instance, the manager nodes
  determine the location of the target from the
  energy detector outputs.
• Manager node use locations of the target at N
  successive time instants to predict the location
  of the target at MltN future time intervals
• The predicted positions of the target are used to
  create new cells that the target is likely to
  enter
• Once the target is detected in one of the new
  cells, it is designated as the new active cell
  and the node in the original active cell (cell A)
  may be put to sleep
• The above steps are repeated for the new active
  cell and so on, thus
• forming the basis of detecting and tracking a
  single target

8
Multiple Targets

• Two scenarios
• Multiple targets sufficiently separated in space
  or time i.e. they occupy distinct space-time
  cells
• Then the same procedure as described for single
  targets can be used, initialing and maintaining a
  separate track for each target
• Multiple targets are not separated in space or
  time
• Needs classification algorithms that operate on
  target signatures to classify them
• Needs temporal processing, FFT
• The output from the classification, active
  classifiers, are reported to the manager node as
  opposed to the energy detector outputs
• Then the same procedure as described for single
  targets can be used, initialing and maintaining a
  separate track for each target

9
Signal Processing

• Detection
• Target Localization
• Tracking

10
Detection

• An event is detected when the detector output
  exceeds a threshold value
• Noise component in the detector may be modeled as
  a Gaussian random variable whose mean and
  variance can be determined from the statistics of
  the background noise
• The Threshold is dynamically adjusted according
  to the noise variance of the detector output in
  order to maintain a CFAR
• Detector outputs below CFAR are used to update
  the Threshold
• Output parameters sent to the manager node
  consists of
• Time when the detector output exceeds the
  Threshold
• Time of maximum CPA
• Detector output of CPA time
• Time when the detector output falls below the
  threshold

11
Target Localization

• Need for Complex Algorithms
• Accurate localization methods based on time-delay
  estimation require accurate synchronization among
  nodes, which comes at a high cost
• Exchange of time series data among nodes, as
  required by some algorithms, consume too much
  energy to be feasible
• Algorithm (4 or more nodes)
• Energy-based algorithm, which assumes an
  isotropic exponential attenuation for the target
  energy source
• where, yi(t) is the energy reading at the ith
  sensor, r(t) denotes the unknown coordinates of
  source with respect to a fixed reference, ri are
  coordinates of ith sensor, s(t) is the unknown
  target signal energy, alpha is the decay
  component

12
Target Localization (contd.)

• The algorithm first computes the ratio
  yi(t)/yj(t) for all pair of sensors to eliminate
  the unknown variable s(t)
• The unknown r(t) is estimated by solving a
  nonlinear least square problem of the form
• where, (x,y) are the unknown coordinates of the
  target, (oi,x, oi,y) are the center coordinates,
  and pi is the radius of the circle associated
  with the ith ratio

13
Target Localization (contd.)
14
Target Tracking

• Given target locations at time instants in the
  past, it is possible to fit data samples into a
  dynamic model to predict future target locations
• For single moving targets, data can be fitted
  into a linear or polynomial model using a least
  square fit
• Is complicated when it comes to tracking multiple
  targets as targets can cross paths resulting in
  data association problems
• Classification algorithms can provide a solution
  to the above problem

15
Target Classification

• Needed for tracking multiple targets
• Feature Vector
• Categories
• Building block
• p (wc x) p (wc) p (x wc)
• p(x)

16
Future Research

• To study the feasibility of introducing parallel
  computations in classification in CSP (Self)
• Choosing between DATA fusion and DECISION fusion

17
References

• KWHS D. Li, K.D. Wong, Yu H. Hu and A.M.
  Sayeed, Detection, Classification and Tracking
  of Targets in Distributed Sensor Networks,
  Department of Electrical and Computer
  Engineering, University of Wisconsin-Madison
• S A. Sayeed, Collaborative Signal
  Processing, MobiComm 2003
• R T.S. Rognvaldsson, Introduction to
  classification, Learning and Self- Organizing
  Systems, January 21, 2001
• DS A. DCosta and A. M. Sayeed, Collaborative
  Signal Processing for Distributed Classification
  in Sensor Networks, Department of Electrical
  and Computer Engineering, University of
  Wisconsin-Madison

18
References (contd.)

• M D. J. Mackay, Introduction to Gaussian
  Processes, Department of Physics, Cambridge
  University
• R P. Ramanathan, Location-centric Approach
  for Collaborative Target Detection,
  Classification, and Tracking, University of
  Wisconsin- Madison

19
Question and Comments