>
Large-Scale Localization
from Wireless Signal Strength
Presented
by Wen-Chih, Hsieh
Outline
Introduction
Concerned Issue
Experiment Design
Parameter Learning
Discussion and Future
Work
Location Estimation
System Application
City-scale service
AT&T, Google,
Microsoft
Specialized web search
Navigation and
Near-by-friend finding
Activity Recognition
Home rehabilitation
Indoor scale
Recognizing complex
behaviors of people by analyzing their motion trajectories
Goal of this Research
To develop a location
estimation system that is large
scale and long-term
Indoor & Outdoor
Coverage
Existing location estimation
system covers indoor or outdoor locations, but not both.
Minimum calibration
It is not
feasible to collect accurately
labeled training data for every location in large-scale coverage area.
System Requirement
Minimum hardware requirements
Users should not be
required to carry special hardware, in order to use a system.
Privacy-observant
Most users are not
willing to be tracked by system in some situation.
Equipment Selected
To meet Minimum hardware
requirements and Privacy-Observant
Wireless networking
infrastructures
802.11 access points(APs)
GSM(Global System for
Mobile Communications) cell tower
How to meet other requirements?
Goal of Bayesian
Localization
To estimate posteriors
over a person00 location, xt, conditions on all sensor measurements
obtained through time t
Particle filter
Using sets of weighted
samples to represent and propagate the posterior.
1) We have a prior of uniform
weighted particle
At this point, we have
m unique samples
2) Particles are weighted based
on the sensor measurement and resampled according to weight to generate
our posterior.
We still have m
samples, but they are all equally
weighted and not necessarily unique
Particles are again unique and
equally weighted
3) Particles are passed through
our motion model to generate a new posterior
Graph-based Location
Estimation System
Important Sampling
Particles with their location
Weight by measure likelihood
Existing WIFI sensor
models:
Signal Propagation
models
Use path loss to
determine likelihoods based
upon distance from
AP
More generalize
Fingerprinting models
location-specific statistics
Require far more training
data
Hierarchical Bayesian
Sensor Model
AP00 location
Signal strength measured at a reference distance from AP
Path loss exponent modeling the degree to which signal strength decreases
with distance from the AP
Learning sensor
model from unlabeled data
Conclusion
Improve sensor model
and add new APs using unlabeled data
Work both outdoors
and inside multi-story buildings
Hierarchical
Bayes model
1
2
3
4
5
GSM : Global System for Mobile
Communications
6
7
8
9
10
11
12
13
14
download Large-Scale Localization from Wireless Signal Strength
