Information-Driven Dynamic Sensor Collaboration for Tracking Applications

1
Information-Driven Dynamic Sensor Collaboration
for Tracking Applications

• Feng Zhao, Jaewon Shin, James Reich
• Xerox Palo Alto Research Center
• Presented by Wei Wei

2
Outline

• Motivation
• Introduction to tracking problem
• Information-driven sensor selection
• Validation
• Representation of belief state
• Conclusion

3
Motivation

• Sensor network is ideally suited for tracking
  moving phenomena due to
• spatial coverage
• multiplicity in sensing aspect and modality
• Detection, classification and tracking of moving
  events requires non-local collaboration among
  sensors
• Aggregation of multitude of sensor data can
  improve accuracy
• Informed selective collaboration of sensors, in
  contrast to flooding data requests to all
  sensors, can reduce latency
• Sensor collaboration can minimize bandwidth
  consumption (energy savings) and mitigate risk of
  network node/link failures
• Energy constrained dynamic sensor collaboration

4
Motivation Energy Constrained Dynamic Sensor
Collaboration

• Dynamically determine
• Who should sense
• What needs to be sensed
• Whom the information must be passed on to
• Tracking dynamically invoke regions of a network
  informed by motion prediction
• Large-scale event monitoring active sensors
  around which there has been a significant change
  in physical measurement
• Disambiguate multiple interpretation of an event
  selectively aggregate multiple sources of
  information to improve detection accuracy, or to
  actively probe certain node

5
Approach

• Base decision for sensor collaboration on
• information constraint
• constraint on cost and resource consumption
• Using measures of information utility, sensors in
  a network can exploit information content of data
  already received to optimize utility of future
  sensing actions, thereby efficiently managing the
  scarce communication and processing resources

6
A Tracking Scenario
Report position of the vehicle every 5 seconds
7
Assumptions of Tracking

• No road constraint, no prior knowledge of
  possible vehicle trajectories can be exploited
• Vehicle can accelerate or decelerate in between
  the nearest sensors
• Focus on sensing collaboration during the tracing
  phase, ignore detection phase, gloss over details
  of routing query into region of interest
• One leader node is active at any moment. It
  selects and routes tracking information to next
  leader
• Every node knows certain information about their
  neighbors location, sensing ability, etc.

8
Leadership Transfer

• Leadership is transferred from node to node
• Leader calculates
• Current belief based on measurements so far
• usefulness of data provided by sensor j
• cost of obtaining information, characterized by
  link bandwidth, transmission latency, node
  battery power reserve. Cost of handling current
  belief state off to sensor j, acquiring data at
  sensor j, and combining data with current belief
• Based on calculation, select best sensor as next
  leader and transfer current belief to next leader

9
Entropy
.11
.11
.11
H(X)2.2
.11
.11
.11
.11
.11
.11
0
0
0
H(X)0
0
0
1
0
0
0
10
Posterior Distribution
.02
.02
.02
.02
.02
.84
.02
.02
.02
?
?
?
?
?
?

?
?
?
?
?
?
?
?
?
?
?
?
11
Posterior Distribution Derivation
12
Expected Entropy, Information Utility and Cost
Function
13
Cost function An Example
14
Validation
15
Validation-cont.
16
Representation of Belief State
Parametric Compact, limited classes of
distributions
Nonparametric When sparse, can be represented
efficiently
Nonparametric Samples
17
Related Work

• Information utility is used in computer vision
  and robotics to manage sensing. Sensor selection
  is done centrally. Geman96
• Expected utility measures for decentralized
  sensing system based on local node.
  (communication cost is not explicitly
  considered.) Manyika Drrant-Whyte 1994.
• Byers 2000 uses a simple step or sigmoid
  function to describe utilities of each node,
  without explicitly modeling of network spatial
  configuration.
• This paper consider both spatial configuration
  and communication cost.

18
Conclusion

• Formulate distributed tracking problem as an
  information optimization problem
• Information-driven approach balances information
  gain provided by each sensor with cost associated
  with acquiring the information
• Sensor network can seek to make informed decision
  about sensing and communication in energy
  constrained environment

19

• Thank you!