1

The Effects of Ranging Noise on Multihop Localization:

An Empirical Study

Kamin WhitehouseJoint With: Chris Karlof, Alec Woo, Fred Jiang, David Culler

IPSN ‘054/24/05

2

Introduction

Ranging Localization

Single-hop Multi-hop

3

Introduction

Ranging Localization

Single-hop Multi-hop

“Noisy Disk”

4

Introduction

Ranging Localization

Single-hop Multi-hop

“Noisy Disk” Unit Disk Connectivity Guassian Noise

Design and comparison Optimal solutions Cramer-rao bounds Algorithmic proofs Empirical parameters

Prediction gap Difference between

predicted and observed error

dmax

σ

5

Introduction

Loca

lizat

ion

Err

or

EmpiricalDeployment

NoisyDisk

PredictionGap

6

Methodology

Loca

lizat

ion

Err

or

EmpiricalDeployment

NoisyDisk

Model B Model C

Significant

Dominant

Sufficient

7

Outline

Deployment Setup Simulation Methodology Comparisons and Analysis

8

Ultrasound Hardware

Circuitry derived from the Medusa node

Cricket’s RF envelope Millibots reflective cone

9

Radio (RSS)

Chipcon CC1000 similar fidelity to WiFi In our experiments, 2m

std error near 20m range RFIDeas: 2m std error near

2m range RFM DR3000 and TR1000:

6m std error near 6m range

10

DV-distance Algorithm

True distance to anchor is approximated by shortest-path distance

Representative of large class using shortest path or bounding box Zig-zag makes paths

longer Noise makes paths

shorter

[16]

11

Ultrasound Deployment

49 nodes on a paved surface

13x13m area 4 anchor nodes Randomized grid

topology Distributed

implementation 7 executions Median

localization error of 0.78m

12

Signal Strength Deployments

49 and 25 node topologies in a grassy field

50x50m area Median localization

error ~4.3 and 13.4m Comparable to GPS

13

Outline

Deployment Setup Simulation Methodology Comparisons and Analysis

14

Traditional Simulation

Ranging estimates are generated using parametric functions

Noisy Disk

Parameters σ and dmax must be estimated from data

[16]

15

Parameter Estimation

[14]

Maximum Range: dmax

Error: σ

16

Statistical Sampling For each in

simulation, randomly choose

Data set includes ranging failures

Can be divided into two components Sampled Noise Sampled Connectivity

±

RangingFailures

17

Data Collection

18

Data Collection

Traditional Data Collection Low spatial resolution

Single pair of nodes at a single orientation

Single path through space

Our Data Collection For each , ~400 empirical

readings taken within 0.05m Represents wide range of

node, antenna, and orientation variability

Captures variability due to dips, bumps, rocks, etc

19

Outline

Deployment Setup Simulation Methodology Results and Analysis

20

Experimental Setup

2 Connectivity and noise components Unit Disk connectivity (D) Gaussian noise (G) Sampled connectivity (S) Sampled noise (S)

Hybrid Simulations (C/N)

Unit Disk

Sampled Conn

No Noise Gaussian Noise Sampled Noise

D/N

S/N

D/G D/S

S/SS/G

21

Experimental Setup

Unit Disk

Sampled Conn

No Noise Gaussian Noise Sampled Noise

D/N

S/N

D/G D/S

S/SS/G

Loca

lizat

ion

Err

or

D/N

S/N

D/G D/S

S/SS/G

Deployment

22

49 Node RSS Experiment

Unit Disk

Sampled Connectivity

No Noise Gaussian Noise Sampled Noise

D/N

S/N

D/G D/S

S/SS/G

D/N D/G D/S S/N S/SS/G Deployment

D/N

S/N

D/N S/N

23

49 Node Ultrasound Experiment

Unit Disk

Sampled Connectivity

No Noise Gaussian Noise Sampled Noise

D/N

S/N

D/G D/S

S/SS/G

D/N D/G D/S S/N S/SS/G Deployment

D/N D/G D/S

D/N D/G D/S

24

Non-disk like Connectivity

25

Non-disk like Connectivity

Less constraints on location

Reduced connectivity can cause more “zig-zag” in the shortest paths This increases

shortest-path distance

[16]

26

49 Node Ultrasound Experiment

Unit Disk

Sampled Connectivity

No Noise Gaussian Noise Sampled Noise

D/N

S/N

D/G D/S

S/SS/G

D/N D/G D/S S/N S/SS/G Deployment

27

Non-Gaussian Noise

28

Non-Gaussian Noise

The shortest-path algorithm selectively chooses underestimated distances

Heavy-tailed noise can decrease shortest path distance

[16]

29

25 Node RSS Experiment

Unit Disk

Sampled Connectivity

No Noise Gaussian Noise Sampled Noise

D/N

S/N

D/G D/S

S/SS/G

D/N D/G D/S S/N S/SS/G Deployment

Significant

30

Conclusions

Non-disk like connectivity Non-Gaussian noise

Methodology A deployment is required to evaluate predictive

ability of a model