Tracking and Collaborative Signal Processing

1
Tracking and Collaborative Signal Processing

• Wireless Ad-hoc Sensor Networks
• EE 206A
• Louane Kuang
• Jonathan Hui

2
Outline

• Basics of Ad-hoc Sensor Networks
• Relatively immobile
• Severely power constrained
• Large scale
• Embedded processing capabilities
• Sensors
• Acoustic/seismic
• Infrared, magnetic, imaging

3
Topics of Presentation

• Tracking and Collaborative Signal Processing
• Applications
• Battlefield tactical
• Environmental monitoring

4
Paper Topics

• Source Localization Beamforming
• Information-Driven Dynamic Sensor Collaboration
• Detection Classification
• Tracking and Reasoning with Relations

5
Detection, Classification and Tracking of Targets

• Detection and tracking of a single target
  requires participation and handoff by several
  nodes
• Target classification is needed for simultaneous
  detection and tracking of multiple targets
• Inter-node cooperation is inherent in detection,
  tracking and classification algorithms and can be
  achieved through collaborative signal processing
  (CSP)

6
CSP

• A single node covers a limited field, therefore
  more than one sensor nodes need to cooperate to
  process space-time signals together to obtain a
  global view
• Distributive processing-raw signals are processed
  locally while only transmitting requested higher
  level information
• Goal-oriented, on-demand processing-information
  is forwarded and processing takes place only upon
  request, otherwise, nodes enter an
  energy-conserving standby mode
• Information fusion-the data exchanged farther
  away are of lower bandwidth than data exchanged
  with closer neighbors
• Multi-resolution processing-resolution of
  sampling depends on the required CSP task

7
Detection

• The output of a detector is sampled periodically
• An output higher than a false alarm threshold
  signals an event
• The threshold is calculated using noise output
  from the detector and is dynamic to new noise
  readings
• Upon detection of an event, data about the event
  is sent to manager nodes that includes the time
  when the threshold was first exceeded, the time
  when closest point of approach (CPA) is achieved,
  the signal detected during CPA, and the entire
  duration of detector outputs remaining above the
  threshold

8
Localization

• Energy measurements from multiple (4 or more)
  nodes are used
• More accurate localization requires time
  synchronization, which is costly in low-power
  sensor nets
• Assumptions made by beamforming and other
  coherent localization algorithms may not hold in
  field environments
• Eliminates the need to exchange time series data,
  which may consume too much energy

yi(t) energy reading of ith sensor r(t)
unknown coordinates of source ri coordinates of
ith sensor s(t) unknown target signal energy ?
decay exponent
9
Localization contd

• Find yi(t)/yj(t) for all pairs of i and j
• this eliminates s(t) and defines a circle which
  contains r(t)
• Estimate r(t) using
• (x, y) is the target coordinate
• (oi, x, oi, y) center of the circle
• ?i radius

10
Tracking

• Geographic positions of nodes are more important
  than high level addresses (ex. a simple way to
  approximate location of a target is the position
  of the node that detected the strongest signal
  from it)
• A geographic region is divided into cells and
  manager nodes selected from nodes in a cell
  coordinate sensing in the cell
• (a) Nodes that detect a target are called active
  nodes and the cell they are in is called the
  active cell, active nodes report sensing data of
  a target to their manager nodes
• (b) Manager nodes use localization algorithms to
  find the target position
• Manager nodes use past target positions to
  predict future target locations
• According to the predicted positions, new cells
  are formed in regions the target is likely to
  enter, some will be activated
• If the a new cell detects the target, a handoff
  occurs between the new active cell and the
  previous one, steps (a) to (e) is repeated by the
  new cell

11
Classification

• Single node classification algorithms
• k-nearest neighbor (kNN)
• maximum likelihood (ML)
• Support Vector Machine (SVM)
• Classifiers chosen to maximize differences
  between target classes
• Power spectral density (PSD) of time series data
• Data generated from seismic and acoustic readings
  for binary classification of tracked and wheeled
  vehicles

12
Classification contd

• x x ? ?N set of feature vectors
• ?1,?2,?,?m set of m target classes, ?c is a
  class
• p(?c) prior probability that x ? ?c
• p(?cx) posterior probability for ?c given x
• x ? ?c if p(?ix) gt p(?jx) for all j ? i
• approximate using gi(x) gt gj(x) if p(?ix) gt
  p(?jx) for j ? i , gi(x) is a discriminant
  function

13
k-NN

• pk is a set of prototypes
• Find distance from test vector to every prototype
• Identify k prototypes closest to test vector
• Combine to generate the appropriate class label
  for the test vector
• Not scalable to increasing prototypes

14
Maximum Likelihood

• Likelihood function
• Gi(x?i)
• ?i mi1, , miP, ?i1, , ?iP mean and
  covariance parameters of the P Gaussian mixture
  densities for a class ?i
• Discriminant function
• gi(x) Gi(x?i)p(?i)
• p(?i) can be approximated by the number of
  training vectors for the class ?i

15
SVM

• Linear classifier
• is the symmetric SVM kernel representation
• a set of nonlinear transformations that
  map the input vector (N-dimension) to a feature
  space (M- dimensions, where M gt N)
• Each class uses a uniquely trained SVM whose
  output gives an approximation of p(?ix) for a
  class ?i

16
Classification Data

• Binary classification between wheeled and tracked
  vehicles using
• Low bandwidth seismic data
• Wideband acoustic data

17
Sensing, Tracking, and Reasoning with Relations

• Relations refer to spatial or temporal
  connections between objects and or environmental
  features
• Relations allow the mapping of high-level user
  queries to low-level signal processing that
  minimizes the use of resources
• Large-scale behaviors and relations of objects
  relative to its environment or to other objects
  may be easier to ascertain than exact object
  position or motion
• Simple global queries can be answered without
  active data collection by aggregating the partial
  information of each node and then storing it
  locally

18
Example of Uses of Relational Sensing

• Who is the leader? (positional relation)
• Am I surrounded? (geometric relation)

s1 f, e2 and e3 form a counter clockwise
triangle (CCW) s2 f, e3 and e1 form a counter
clockwise triangle (CCW) s3 f, e1 and e2 form a
counter clockwise triangle (CCW) Therefore, e1,
e2 and e3 form a CCW enclosing f, which is indeed
surrounded
19
Kinetic Data Structure (KDS)

• Incremental update is a more efficient way to
  track the attribute values of a target
• Objects are allowed to move as long as the
  relations among them stay valid
• Certificates- elementary relations that certify
  the value of an attribute
• KDS- data structure designed for maintaining data
  about objects that move incrementally using
  several support certificates
• A KDS algorithm is used to find alternate
  certificates (relations) that will support an
  attribute when sensors cannot support current
  ones or the current relations have failed
• The goal is to find certificates that change
  incrementally and locally according to coherence
  of motion
• More certificates implies a quicker computation
  of attribute values, but it also means a greater
  likelihood of certificate failure requires more
  processing to fix

20
KDS Contd

• The KDS model incorporates the costs for sensing
  and communication in sensor nodes
• KDS is useful for
• Coordinating groups of sensors during target
  tracking
• Motion prediction of the target to facilitate
  formation of tracking groups
• Creating and maintaining clusters of moving nodes
• Directing communication routes throughout the
  sensor net either as a relay for outside user
  nodes or for sensor nodes within the net

21
Probabilistic Relational Reasoning

• KDS needs to be enhanced with tolerance for
  uncertainty in sensing target location
• The belief state of the system regarding target
  location is represented as probability density
  function which is translated into a set of
  weighted particles, each particle represents a
  position and the corresponding weight gives the
  probability
• The distribution has to take into account not
  only the target location, but other attributes
  associated with the target
• A distribution is factored into parts each
  represented by a particle with independent
  uncertainties

22
RDBN

• Dynamic Bayesian Networks (DBN) can be used to
  model dependencies resulting from the various
  states an object goes through during motion and
  it is adapted to the sensor environment
• Relational Dynamic Bayesian Networks (RDBN) are
  used to deal with uncertainties and change
  occurring in the relations between objects , in
  the identities of the objects and in the number
  of participating objects
• RDBN can be integrated with KDS
• KDS algorithm finds certificates with confidence
  specified by the RDBN model
• The RDBN belief state representation can be
  modified by the KDS so that its belief state can
  be more easily matched to good certificates and
  improve its accuracy

23
Issues

• Variability in data
• Sufficiently accurate time synchronization and
  position of sensors are difficult to obtain even
  with GPS
• Doppler shifts due to motion may create spectral
  variations that inhibit accurate classification
  of targets
• Data used to train classifiers may not resemble
  actual data obtained in the field

24
Keys to Tracking

• Leverage the distributed computing environment
  with respect to
• Sensor networks enable dense spatial sampling
• Asynchronous
• Optimization
• Information Gain
• Resource Cost

25
Information-Driven Dynamic Sensor Collaboration

• Collaborative Signal Information Processing
  (CSIP)
• dynamically determine
• who should sense
• what needs to be sensed
• who the information must be passed on to

26
Assumptions about Sensors

• Have local sensing communication range
• Physical phenomenon of interest
• Can locally estimate cost of sensing/processing/co
  mmunicating
• Monitor power usage

27
(No Transcript)
28
Tracking Scenario

• Moving vehicle in two-dimensional space
• No road constraint
• No prior knowledge can be exploited
• Vehicle accelerates/decelerates between sensors
• Many sensors potentially make simultaneous
  observations
• Potentially (Flood network with information)

29
Sensor Selection

• Wish to incrementally update the belief by
  incorporating measurements of other nearby
  sensors
• Not all sensors provide useful information that
  improves estimate
• Task is to select an optimal subset of available
  sensors and optimal order of how to incorporate
  these new measurements

30
Collaboration

• Detection quality
• Track Quality
• Scalability
• Survivability
• Resource Usage

31
Information Driven Sensor Querying (IDSQ)

• Bayesian Estimation problem
• x - target we wish to estimate
• zi - sensor measurement (at location i)
• p(x z1,, zj-1) - current estimate
• p(x z1,, zj-1,zj)-new estimate based on latest
  measurement zj
• select sensor j that provides greatest
  improvement at the lowest cost

32
IDSQ Continued

• Optimization Problem
• M(p(x z1,,zj)) ???Utility(p(x z1,,zj))
  - (1-?)?Cost(zj)
• ?Utility() - information utility measure
• characterizes usefulness of data provided
• ?Cost() - Cost of resources
• cost of obtaining information (link bandwidth,
  transmission latency, power reserve)
• ? - relative weight of utility versus cost

33
Information Utility

• Examples of what ? could be defined as
• Information-Theoretic Measure Entropy
• Mahalanobis Distance Measure
• Measures on Expected Posterior
• Apply one of the above to a simulated measurement
  incorporated into belief state

34
Information Utility Entropy

• Natural choice for ?Utility() is statistical
  entropy (measures randomness of random variable)
• Smaller the entropy the more certain we are about
  the random variable
• For example ?Utility() -Hp(x)

35
Information Utility Mahalanobis Distance

• Works well when the current belief state is well
  approximated by a Gaussian distribution
• xj is the position of sensor j
• x is the mean of the belief (target position
  estimate)

36
Sensor Selection continued
37

• Estimation error for Nearest neighbor selection
• Estimation error for Mahalanobis Distance
• Estimation error for minimizing entropy

38
Additional Considerations

• Sequential versus Concurrent Information exchange
• node-to-node versus leader-to-leader
• Parallel information exchange
• Tracking Robustness
• sensor placement density
• sensing range
• communication range

39

• Sources
• Chen, J.C. Kung Yao Hudson, R.E. Source
  localization and
• beamforming. IEEE Signal Processing Magazine,
  vol.19, (no.2),
• IEEE, March 2002. p.30-9.
• Dan Li Wong, K.D. Yu Hen Hu Sayeed, A.M.
  Detection,
  classification, and tracking of targets. IEEE
  Signal Processing Magazine, vol.19, (no.2),
  IEEE, March 2002. p.17-29.
• Feng Zhao Jaewon Shin Reich, J.
  Information-driven dynamic
• sensor collaboration. IEEE Signal
  Processing Magazine, vol.19,
• (no.2), IEEE, March 2002. p.61-72.
• Guibas, L.J. Sensing, tracking and reasoning
  with relations. IEEE Signal Processing Magazine,
  vol.19, (no.2), IEEE, March 2002. p.73-85.
• Sri Kuma Feng Zhao David Sheperd.
  Collaborative Signal and Information Processing
  in Microsensor Networks. IEEE Signal Processing
  Magazine, vol.19, (no.2), IEEE, March 2002.
  p.13-14.