RADAR: An InBuilding RFbased User Location and Tracking System

1
RADAR An In-Building RF-based User Location and
Tracking System

• P. Bahl and V.N. Padmanabhan
• Microsoft Research (2000)

Presented by Conan Noronha
2
Localization

• Context
• - Physical, Symbolic, Absolute, Relative
• Resolution or Granularity
• Real Time Component
• Active or Passive
• - Passive burdens the device to be located with
  computation

Where am I?? Where's the stuff around me!!
3
Alternative Technologies

• InfraRed
• - Limited Range
• - High Infrastructural Costs
• - Interference From Sunlight
• eg. Active Badge System
• Pulsed DC Magnetic Fields
• - High Precision
• - Limited Range
• - Expensive
• Global Positioning System (GPS)
• - No Reception Indoors
• - High Cost, Power Size

4
The Driving Force Behind This Technology

• Exploit the existing infrastructure of an RF
  wireless LAN, to achieve localization
• - Uses a Radio Map constructed from signal
  strength
• - Overcomes the multi-path problem
• (Especially at short range)
• RF Signal Strength varies as a
• function of distance.
• - High range
• - Scalable
• - Easy Deployment
• - Low Maintenance

5
The ExperimentalSetup

• Base Stations provide overlapping coverage
• - Pentium PCs running BSD 3.0
• Mobile units transmit
• -Pentium Laptops running Windows 95
• Signal Strength(SS) SNR
• - SS is a stronger function of distance
• SS varies as a function of User Orientation
• Two Methods
• - Empirical Method
• - Signal Propagation Model

6
Variation of Signal Strength With Distance
7
The Empirical Method

• Off-line Phase Constructs the Radio Map
• SS from various locations are recorded
• - Each of the 4 directions at 70 distinct
  locations
• - gt 20 samples per location-orientation pair
• Stored in tuples of the form (x, y, d, ssi, snri)
  for I 1,2,3
• Form the search space for the Nearest Neighbor(s)
  in Signal Space algorithm (NNSS)
• Real-time Phase Simulated from off-line data
• Pick any one location-orientation reading
• Find its nearest neighbor in signal space from
  amongst the remaining 69 4 readings
• NNSS gt Sqrt((ss1-ss1)2 (ss2-ss2)2
  (ss3-ss3)2)is minimum

8
Results of the Empirical Method

• Compared with
• - Random Selection
• - Strongest Base Station
• Note Error is the Euclidean Distance in
  physical space

9
Improving On The Results

• Uses the k-NNSS algorithm
• - There will be multiple close neighbors in
  signal space
• - The error vectors in physical space will
  average to the actual point
• - Deciding the value of k is critical

N1

• Eg. k 5
• 22 improvement at 25th percentile
• 9 improvement at 50th percentile
• Large k values include far neighbors
• Nearest k neighbors in signal space may not be
  physically nearby.

G
T
N2
N3
3 Nearest Neighbors T - True Location, G -
Guess N1,N2,N3 - Neighbors
10
How Important Is Orientation?

• Compute the maximum SS at each base station
  across the four possible orientations for each
  location Emulates no obstruction

k 1 6 9 at 25th 50th percentiles k
2 to 4 48 28 at 25th 50th
percentiles Distinct k neighbors cause
averaging to be more effective
11
How many samples are needed?

• Off-line samples
• 40 is almost as good as 70
• Real-time samples
• Improvement
• 1 30
• 2 11
• 3 4

12
Mobile User Tracking

• 4 Samples per Second
• Reduce the problem to locating a stationary user,
  by using a sliding window of 10 samples
• Median error of 3.5 meters

13
What's wrong with the Empirical Method?? The
Off-line Dataset!!
14
Radio Propagation Model

• Issues The Multi-path problem
• - Reflection
• - Scattering
• - Diffraction
• Causes
• - Layout
• - Construction Material
• - Electrical Links
• - Objects People
• Model Options
• - Rayleigh Fading Model All signals have equal
  strength - Unrealistic
• - Rician Distribution Model Highly Complex
• - Wall Attenuation Factor Simple, yet flexible

15
The Wall Attenuation Factor Model

• P(d0) Power at some reference point, distance
  d0 away
• nW Path Loss exponentNumber of Walls
• C Threshold above which nW makes no difference

Actual readings
Readings corrected for walls
16
How well does WAF work?
4.3m resolution at 50th percentile Compare with
2.94m Empirical 1.86m at 25th
percentile Compare with 1.92m

• Common values can bring down setup costs

17
Further Improvements Possible

• Use movement pattern information
• Environmental profiling
• - Multiple data sets improves reliability

18
Questions ??

• Thank you !!