Tracking Moving Devices with the Cricket Location System

1
Tracking Moving Devices with the Cricket
Location System

• Adam Smith, Hari Balakrishnan, Michel Goraczko,
  and Nissanka Priyantha
• Mobisys, 2004
• Ku Dara

2
Contents

• Introduction
• Tracking Algorithm
• Hybrid Architecture
• Evaluation
• Conclusion

3
Introduction(1/4 )

• Location-aware application

Direct users to their desired destinations on
active map
Lacation sensors provide position information to
a moving robot
Players can move in the real world game like Doom
or Quake
4
Introduction(2/4)

• Location-awareness system

Indoor environment
Outdoor environment
Active Badge(infrared)
GPS (Global Positioning System)
Active Bat(ultrasonic)
Hiball tracking system (infrared LED)
Whisper system(audio)
Cricket(RF,ultrasonic)
5
Introduction(3/4)

• Indoor location architecture

Infrastructure receivers
Infrastructure trasmitters
DB
receiver
trasmitter
Cricket system
Active Badge, Active Bat
6
Introduction(4/4)

• Comparison
• Active mobile architecture
• Cost high, scalability low, performance high
• Require a network infrastructure to connect the
  deployed receiver to central database
• raising privacy concern
• Passive mobile architecture
• Scalability high, cost low, performance low
• Independent privacy concern

7
Tracking algorithm(1/5)

• Three Componets of tracking algorithm
• Lesat-squares minimization(LSQ)
• Use LSQ to reset the bad EKF state
• Extended Kalman filter(EKF)
• Predicte next devices state from samples
• Correct the prediction each time new distance
  sample is obtained
• Outlier rejection
• Bad distance sample eliminate

t,p,d t current time p known position of the
beacon or receiver d distance btwn mobile device
and known beacon or receiver F a good position
estimate
8
Tracking algorithm(2/5)

• LSQ(Least Squares Minimization)
• If mobile devices were static, a standard way to
  solve the problem of estimating is by minimizing
  the sum of the squares of the error terms
  corresponding to each distance sample
• LSQ is complex
• LSQ does not always produce a good estimate
• Use to initializing and reseting the Kalman
  filter (bad state)

A
B
9
Tracking algorithm(3/5)

• EKF(Extended Kalman Filter)
• Using a state vector with six components
• Three position components(x,y,z), three velocity
  components( )
• Use the most recent distance sample and internal
  state to project ahead and produce an estimate of
  F of where the device might be in the next
  time-step
• P model velocity and higher-order derivatives
  are zero
• PV model acceleration and higher order
  derivatives are zero
• Multi-modal filter combining the output states
  of PV and P models

10
Tracking algorithm(4/5)

• Outlier Rejection
• Wherein egregiously bad distance samples are
  eliminated

Outlier Rejection
bad samples
outlier elliminated samples
IF( ?)
True?eliminate False?accept
r residual(guess-actual measurement) ?empirical
ly-selected parameter
11
Tracking algorithm(5/5)

• Tracking algorithm

Measurements
t current time p known position of the beacon
or receiver d distance btwn mobile device
and known beacon or receiver
transmitter
EKF next position extimate
Current estimate distance correct
12
Hybrid Architecture(1/3)

• Problem
• Bad EKFstate
• Extremely different estimate
• Passive mobile system has a higher probability of
  reaching a bad state
• Rarely happen in active mobile

1
2
3
1
2
3
123
123
123
1
12
123
Recent value
Old value
Passive mobileone sample, inaccurate
active mobilemultiple samples, accurate
13
Hybrid Architecture(2/3)

• Solution
• Normal state passive mode use for Scalability,
  user-privacy
• Bad Kalman filter state active mode use

14
Hybrid Architecture(3/3)

• Solution

Compute distance samples
RF message
Beacon use a simple CSMA scheme with randomized
back off to avoid RF collisions
ActiveChirp
receiver
Reset EKF
Internal state
15
Evaluation(1/4)

• Experimental setup
• Crickets h/w and s/w
• Computer-controlled Lego train with Cricket
  attached to the moving train
• Cricket listener to the train, beacons to the
  ceiling
• Six different speedmodel a range of realistic
  pedestrian speeds

16
Evaluation(2/4)

• Error CDF(speed0.78m/s)

• Passive
• Multi-modal(MM)
• 90, error 30cm
• LSQ
• 90,error 50cm (poor)
• Precision MM gt LSQ
• Active
• 90, error 10 cm
• High precision

Occurrence
error(cm)
17
Evaluation(3/4)

• Error CDF(speed1.43m/s)

• Hybrid
• 0.43 m/s 90, 20 cm
• Speed increase, LSQ low precision

High precision
Active-MultiModal
Hybrid-EKF-PV
Passive-MultiModal
Passive-LSQ
Low precision
Act 10cm
P-LSQ 85cm
Hv 45cm
P-MM 65cm
18
Evaluation(4/4)

• Median error
• Hybrid close to active mode

Passive 25cm below
Hybrid 15cm below
Active 5cm below
19
Conclusion(1/1)

• Hybrid architecture
• Preserve the scalability and privacy advantages
  of the passive mobile
• Improving tracking precision