AHLoS

1
AHLoS

• Review of Dynamic Fine-Grained Localization in
  Ad-Hoc Networks of Sensors (A. Savvides, C. Han,
  M. Srivastava)
• Lewis Girod
• 22 May 2001

2
Points to consider when evaluating localization
work
3
Ranging Assumptions

• Important points to consider about ranging
• Simulation or Experiment?
• If simulation, where did the data come from?
• Real data how was it collected? (see
  experiment below)
• Modeled data What are the models?
• If experiment, what was the experimental setup?
• Indoors? Outdoors?
• Multipath environment?
• Obstructed environment?
• Beacon placement?

4
Algorithmic Assumptions

• Points to consider about algorithms
• Power efficiency.. Many things consume power
• Ranging experiments
• Communication of intermediate results
• Communications overhead due to noise, collision,
  etc
• Centralized vs. distributed
• Varies on a continuum
• No one-size-fits-all solution.. Either can be
  appropriate
• Iterative estimation often part of distributed
  solution
• Intermediate results must be communicated.. Can
  be costly
• Convergence time is critical to lowering power
  requirements
• Convergence often affected by environmental
  conditions

5
AHLoS
6
RSSI Findings

• RSSI is not useful for fine-grained localization
• RSSI ? 1/rn, where n depends on
• Environment grass vs concrete
• Multipath constructive/destructive interference
• Antenna height
• Fading
• OBS/LOS
• Time-varying long term dependence on environment
• Motion can help here, not always possible with
  sensor nets
• Precision was a few meters at best

7
Ultrasound Findings

• Medusa node 5 pairs of transceivers, pointed
  in different directions
• Ultrasound time-of-flight system
• Claim multipath variations can be filtered out
• It is true that the variations caused by
  multipath can be filtered out easily by taking
  multiple measurements
• Does not solve problems with persistent effects
• constructive interference
• lightweight obstructions to LOS (e.g. foliage)
• Heavy obstructed conditions (NLOS)

8
Algorithmic Findings

• Atomic Multilateration
• Requires ranges to 3 beacons (4 beacons to
  simultaneously estimate speed of sound)
• Maximum Likelikood estimate
• Minimum Mean Square Estimate
• Constraint equations from distance estimates
• Squaring distance functions results in set of
  linear eqns
• Speed of sound just another unknown
• Important underlying assumption that errors in
  data fit a gaussian error model

9
Collaborative Multilateration

• Collaborators must see 3 neighbors that are
  either beacons or are other collaborators
• i.e. subgraph of nodes with degree ? 3
• System of distance equations.. Nonlinear
• Solve with gradient descent or simulated annealing

10
Iterative Multilateration

• Combine Atomic and Collaborative multilateration
• Small fraction of nodes (beacons) know their
  position
• All-to-all ranges measured
• Estimate positions of unknown node with
  multilateration preference for atomic because it
  is more accurate and computationally less
  expensive
• Newly estimated nodes now considered beacons,
  iteratively estimate positions of other nodes
• Node density requirements to achieve degree of 3
• Requirements are on node degree (at least 3 or 4)
• Analytical result for uniform distributions on
  plane, 10m range, 0.016 nodes/m2 provides 98
  probability of degree 3 connectivity

11
Experimental Results

• Experiments with 9 Medusa nodes and a PC
• No results showing error in localization
• No description of environment (tabletop?)
• Power model does not consider sound emitter
• Only considers radio cost, processor cost
• Analysis only considers radio cost
• No results describing convergence properties
• How many iterations does it take to get the
  answers?

12
Centralized vs Distributed

• Argues for distributed solutions using simulation
• Fully Centralized energy cost linear with
  network size
• Fully Distributed cost independent of network
  size
• This is not surprising
• A more interesting question relates to
  convergence
• Given convergence properties, what is the correct
  degree of centralization?
• What is the optimal cluster size
• Where a cluster-head does a local centralized
  computation
• Balance cost of transmitting data estimates
  with cost of iteration

13
Unanswered Questions

• Data on power consumption? How much do acoustics
  cost relative to GPS?
• Accuracy of Position Estimates
• Convergence properties
• Sensitivity of above to degree and type of error?
• More detail on experimental setup