1. Evaluating Wi-Fi Location Estimation Technique for Indoor Navigation
Roshmi Bhaumik
Today’s topic is indoor location estimation in a Wi-Fi network

2. Outline Motivation Problem Definition Related work Our Contribution
Experiment
Case Study
Conclusion and Future work

3. Motivation Wireless communications Indoor location aware applications
Chaska city wide Wi-Fi available from July 2004
Minneapolis city wide Wi-Fi proposal , 2006
Indoor location aware applications
Indoor Navigation aid for visually impaired
Navigation through Exhibition Halls and Museums
Challenges of indoor Wi-Fi positioning
Noisy characteristics of wireless channels
Multi-path fading
Navigating through public spaces
For visually impaired or people unfamiliar to the surrounding
Healthcare
Tracking patients carrying emergency alarm systems or tracking expensive hospital telemetric devices.
Retail
Track shopper behavior, provide store/ product location for shoppers and offer promotions based on shopper’s location.
Security
Access control of wireless devices and detecting device positions
Entertainment
Location based gaming

4. Motivation(2)
Evaluating Wi-Fi based Indoor Location Estimation techniques for indoor navigation
Location aware solution (client)
Location based Server
Noisy characteristics of wireless channels
Bluetooth devices, cordless phones, microwaves etc can cause interference as they all operate in the same band as Wi-Fi devices, namely 2.4 GHz
Water and human bodies absorb RF signals at that frequency
Multi-path fading
Due to multiple reflections, refractions and attenuation paths for signals , a transmitted signal can reach the receiver through different paths and thus have different amplitude and phase.
Changes in the environment, such as temperature and humidity can affect the received signal strength
Core Common Technology
Indoor Positioning Technology
System Architecture of Location Based Service

5. Outline Motivation Problem Definition Related work Our Contribution
Experiment
Case Study
Conclusion and Future work

6. Problem Definition Given Evaluate Objective Constraints
Wi-Fi Localization scheme
Evaluate
Given localization scheme for indoor navigation system
Objective
Fulfilling accuracy requirements of navigation application
Constraints
Using existing Wi-Fi Access Points (APs)
Experiment carried out in typical building

7. Outline Motivation Problem Definition Related work Our Contribution
Experiment
Case Study
Conclusion and Future work

8. Related Work Requirements of positioning for indoor navigation:
Accuracy
Integrity- issue alarm in case of large estimation errors
Availability (Coverage)
Continuity of service (Location Estimation response time)
(Swiss Federal Institute of Technology 2003)

9. Related Work (2) Evaluation criteria: Location labeling
System performance
Architecture
Cost
Hightower et al, 2001 [14]
Algorithm vs. accuracy
Performance vs. accuracy tradeoff
Fault-tolerance
P. Prasithsangaree et al, 2002 [11]

10. Outline Motivation Problem Definition Related work Our Contribution
Experiment
Case Study
Conclusion and Future work

11. Our Contribution Limitations of related work:
Did not consider effect of different motion patterns on accuracy
Our Contribution:
Evaluate the positioning accuracy with different types of movement
Other lessons learnt
Case Study: Indoor navigation aid for the visually impaired

12. Outline Motivation Problem Definition Related work Our Contribution
Experiment
Case Study
Conclusion and Future work

13. Experiment Design
Floor Map
Stationary
Stop and Go
Smooth about a point
Smooth uniform motion
Change in direction
Ekahau
Software
Track
controlled
movement
Calibration
Constant parameters:
Device Scan Interval (500msec)
Accurate Mode
Variable parameters:
Location Update Interval
( msec)
Location
Estimates
Number of visible Access Points (APs) –
Significant benefit when going from 1 to 2 or from 2 to 3 APs but little incremental benefit beyond 3 APs.
Placement of APs-
Placement should be such that it avoids symmetric coverage
Calibration density –
Higher density provides better accuracy only if sample points show variation in signal intensities
Irregular distribution should be avoided
Device speed –
For devices move very fast, density of calibration should be low for optimal latency/ accuracy ratio
Speed above 1m/s should not be continuously used for average error of 2m
Device scan interval –
Interval of reading RSSI values from the client device
For average error of 2m need scan interval smaller than 2500 ms
Environmental –
Open areas vs corridors
Metal bodies around cause radio interference
Accuracy
Calculations
Error measured= Euclidean distance between gold standard and estimated location

14. Movements for a point 1) Stationary: Gold standard= A
10 Readings taken after 2min
2)Stop and Go
Gold standard= B
10 Readings taken within 30 sec
3)Smooth about a point
Gold Standard= B
Start taking readings at A and end at C A A B
Change in wireless infrastructure
Addition, removal or displacement of location of APs
Change in direction of antenna
Change in transmission power of APs
Proximity to APs
Stability of calibration is better near transmitters
Change in radio environment
New walls, cubicle changes, large containers, furniture
Busy hour with lots of people and non busy hours with empty spaces eg. Malls, shipping/ receiving areas etc.
< =3ft
<= 3ft A B C
Slow motion
Quick motion
Static

15. Smooth Motion on Tracking Rail
4) Smooth uniform motion
Gold standard= points on the tracking rail
Readings taken using the Ekahau Accuracy Tool
Error calculation is done by the tool
Slow motion
Tracking rail

16. Direction Change 5) Changes in direction
a) Turning back 180 degrees on the path
Gold Standard = Point B
Readings taken at B after turning
b) Intersections with ambiguous signals
Readings taken at B for all types of movement for a point A B A B
Slow motion
Tracking rail
Turn back

17. Elliot Hall First Floor

18. Elliot Hall Third Floor

19. Ekahau Positioning Software
Ekahau Client (on every mobile device)
Retrieves signals from visible Access Points through the network cards
Ekahau Manager
For site calibration, adding logical areas, tracking and accuracy analysis
Ekahau Positioning Engine (EPE)
Server stores calibration data
Calculate location estimates
Algorithm: Maximum A Posteriori Estimate using Bayes Rule
Applications retrieve positioning data through YAX protocol /Java SDK

20. Results

21. Comparing Types of Motion: Floor 1
Actual X,Y
Stationary
- Error (ft)
Smooth slow about point -Error (ft)
Averaged X,Y – Error (ft)
Stop and go -Error (ft)
703,79
0.78
7.0
6.97
9.79
703,848
6.33
11.82
6.94
11.06
703,969
11.58
18.43
15.55
11.73
703,1268
3.17
10.92
9.36
10.68
702,1452
5.53
5.64
4.50
5.98
457,1419
2.25
3.5
3.22
4.37
780,1485
6.89
6.41
5.34
26.50
Average
5.22
9.10
7.41
11.44

22. Comparing Types of Motion: Floor 1
Actual
Stop go
Smooth Avg
Stationary

23. Comparing Types of Motion: Floor 3
Average Accuracy variation with
type of motion: (1, 2, 3 &5a)
Actual
Smooth -5.8
Stationary -5.5
Stop n Go- 10.2
Turn around – 6.8
9.7,9.1,10,9.2
7.2,7.2,17.5,7.7
8.5,6.5,21.2
2.3,2.3,5.6
0.7,3.3,0.7,4.1
6.5,4.8,6.2,6.2

24. Smooth uniform motion: Floor 1
Use Ekahau Accuracy Tool
Smooth uniform motion in a straight line along the tracking rail
Error vector shown in red color
Discuss the effect of inaccuracy at the intersection with ambiguous signal pattern with added animation
Floor
Min
Max
Average
90%
Med
Elliot-1st units in ft
0.4
26.2
8.8
16.8
7.6

25. Smooth uniform motion: Floor 3
Readings taken using Ekahau Accuracy Tool
Discuss the effect of turning back 180 here by adding animation
Floor
Min
Max
Average
90%
Med
Elliot -3rd units in ft
0.8
16.8
7.2
12.4
8.1

26. Ambiguous Intersection
All types of movement for a point are considered
Average Error worse than 12 ft for this specific location
Discuss the effect of inaccuracy at the intersection with ambiguous signal pattern with added animation
Actual X,Y
Stationary
- Error (ft)
Smooth slow about point -Error (ft)
Averaged X,Y – Error (ft)
Stop and go -Error (ft)
703,969
11.58
18.43
15.55
11.73

27. Analysis
Accuracy requirements are met for the listed types of movements except :
Stop and go
Direction change at ambiguous intersection
System does not respond to quick changes and error increases is due to overestimation
Ambiguous signal patterns cause large errors

28. Other lessons learned Calibration Stability: Stable over 4 months
Transferability: Calibration done with one client, used to track any other supported clients with same accuracy
Tested with Cisco, Orinoco and Dell’s built-in WLAN cards and Toshiba PDA
Calibration is transferable as EPE performs normalization of RSSI* values from different network cards and devices
*Radio Signal Strength Indicator

29. Other lessons learned Tracking on multiple floors
Adjacent floor maps linked at certain positions (e.g. start of staircase)
Latency (5-15 sec) depends on device speed.
Connection points between floors are not specified
Correct floor detected with greater latency (2 min)
Floors are significant barriers to signal propagation
Points in adjacent floors have different signal patterns

30. Other Lessons learned
Accurate mode meets accuracy requirement and latency is 5 sec
Variation of LUI* does not completely control the effect of history
At low RSSI, changes in signal strength with distance is very small. This causes signal aliasing and reduces accuracy
*Location Update Interval

31. Outline Motivation Problem Definition Related work Our Contribution
Experiment
Case Study
Conclusion and Future work

32. Case Study: Indoor Navigation Aid for the Visually Impaired
We used EPE to build an Indoor Guidance System (IGS) meant to help the visually impaired to find their way inside buildings
BUILDING DATABASE
Spatial data: X,Y
Attribute data: Physical & Logical spaces
CLIENT
Location Aware Application
Audio/ Visual Output
Wi-Fi Sensors
USER
LOCATION SERVER (EPE)
Positioning Model: floor maps and calibration data
Information Flow Diagram

33. Existing Work Building Database
Data Entry interface: building and floor data
Audio/Visual output : a list of surrounding features, orientation and distance from current user position
User input needed to get current user position

34. My Additions
Infer current user location in building database coordinates
Multiple floor tracking and transferable calibration
Improve accuracy and latency of location estimate
Background thread to collect location & error estimates
Average location estimates over specified time intervals
Add location buoys spaced 15 ft
Reduce computation
Describe the surrounding with finer granularity

35. IGS screen shot

36. Application Design Guidelines
During calibration, sample points should be rejected if RSSI is below certain threshold
Averaging location estimates over optimum time interval is better than getting a single instantaneous estimate
Accurate mode gives better location estimates for normal walking speeds
Accuracy will vary depending on the type of movement
The signal strength variability at a fixed position causes a location error that cannot
be eliminated if a single measure is executed.

37. Conclusions and Future Work
EPE location estimation scheme performs well for an indoor navigation application but some areas can be improved:
Calibration stability and transferability is good
Tracking over multiple floors works well
Accuracy is good for smooth movement in a straight line and stationary state
Accuracy is poor when signal strengths are low
Differences from LOCADIO:
LOCADIO – transitional probabilities are functions of 1. Shortest distance path 2. Human walking speed distribution
3. Still/ Moving inference.
The last two are combined to obtain probability of a given speed given moving or still.
The probability of a given speed is combined with the probability of transitioning to a node based on shortest distance path.
My approach- Pre-calculate the transitional probability matrix based on moving and still ( ie 2 matrices). Based on moving or still inference
Switch between the two matrices.
The “moving” matrix may be independently updated based on speed inference -> using another HMM.
Example in tracking people, one HMM could be assigned to (x,y) and another to speed.
The results of the speed inference can be used to update transitional probabilities for (x,y).

38. Conclusions and Future work(2)
Accuracy is poor if the path of movement has a number of intersections, specially with ambiguous signal pattern
Using directional APs for asymmetric coverage
Accuracy is poor for stop and go motion
Using some method to detect still and moving state
Use above to choose the transition probabilities
Related work –LOCADIO[7]
Further tuning of the application can achieve better accuracy for indoor navigation
Future work:
Implement own location detection scheme.
Incorporate suggested improvements

39. Keywords Wi-Fi : Wireless fidelity, IEEE802.11b
APs: Access Points, Within the range (~50ft) of an AP, the wireless end-user has a full network connection with the benefit of mobility.
RSSI : Received Signal Strength Indicator
Bayesian networks :A probabilistic graphical model . The nodes represent variables and directed arcs represent conditional dependencies between variables.
Stochastic model :A mathematical model which contains random (stochastic) components or inputs; consequently, for any specified input scenario, the corresponding model output variables are known only in terms of probability distributions in contrast to a deterministic model
Machine learning :A method for creating computer programs by the analysis of data sets.
IGS: Indoor Guidance system, a navigation aid developed for the Low Vision Lab, Psychology department, UMN

40. References
Papers:
1. P. Bahl and V. N. Padmanabhan, “RADAR: An In-Building RF-Based User Location and Tracking System,” Proceedings of IEEE Infocom 2000, March 2000, pp. 775–784
2. P. Myllymaki, T. Roos, H. Tirri, P. Misikangas, and J. Sievanen, “A Probabilistic Approach to WLAN User Location Estimation,” Proceedings of the 3rd IEEE Workshop on Wireless LANs, September 2001, pp. 59–69.
3. R. Battiti, A. Villani, and T. Le Nhat, “Neural network models for intelligent networks: deriving the location from signal patterns,” in Proceedings of AINS2002, (UCLA), May 2002.
4. Castro, P., et al. A Probabilistic Room Location Service for Wireless Networked Environments. in Ubicomp 2001.
5. Ladd, A.M., et al. Robotics-Based Location Sensing using Wireless Ethernet. in Eighth International Conference on Mobile Computing and Networking
6. John Krumm, “Probabilistic Inferencing for Location”, Proceedings of the 2003
Workshop on Location-Aware Computing, October 2003.
7. John Krumm and Eric Horvitz, "LOCADIO: Inferring Motion and Location from Wi-Fi Signal Strengths", First Annual International Conference on Mobile and Ubiquitous Systems: Networking and Services (Mobiquitous 2004), August 2004
8. D. Fox, J. Hightower, H. Kautz, L. Liao, and D. Patterson. "Bayesian techniques for location estimation" in Proceedings of The 2003 Workshop on Location-Aware Computing, October2003.
9. L. R. Rabiner, "A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition," Proc. of the IEEE, Vol.77, No.2 pp , 1989.
10. J.A.Tauber.Indoor Location Systems for PervasiveComputing
11. P. Prasithsangaree, P. Krishnamurthy, and P. K. Chrysanthis, "On Indoor Position
Location With Wireless LANs ," The 13th IEEE International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC 2002), Lisbon, Portugal, September 2002.
12.
13.
14. Jeffrey Hightower and Gaetano Borriello.Location systems for ubiquitous computing.IEEE Computer,August2001.
Slides: P1.”Graphical Models on Manhattan: A probabilistic approach to mobile device positioning” presented by Petri Myllymäki and Henry Tirri
P2. faculty.cs.tamu.edu/dzsong/teaching/ fall2004/netbot/Yutu_Liu_Robot.ppt

41. Commercial Solutions Ekahau Aeroscout WhereNet Newbury 3.5 ft 3.1 ft
Avg. Accuracy claims
3.5 ft
3.1 ft
10ft
Tracking Range
WLAN covered site
Latency claims
Low(?) -
< 5 sec
Software API
Open API : SDK (Java) or TCP protocol
Open API for 3rd party applications
Available to partners for building applications
Vendors need to integrate it in their applications
Client software
yes
none
Proprietary Hardware
none( any Wi-Fi enabled client)
Summary
Software based solution using Wi-Fi technology for location estimation
Hardware and software based solution using Wi-Fi and Active RFID technology
Hardware & software based solution based on Wi-Fi technology for WLAN security , access control etc.
There are a lot of such new solutions. These are the more popular solutions in the market. These are all applicable to indoor positioning using Wi-Fi technologies. The comparison parameters are based on requirements for the navigation application.

42. Wi-Fi Positioning Theory
Static Location Estimation with Wi-Fi
Nearest Neighbor [1]
Minimum Euclidean distance in signal space between real time and offline entries (x,y,z,ssi (i=1..N))
weighted average of k nearest neighbors
Average accuracy ~ 10 ft
Neural Networks [3]
Dependencies between signal and location are not easily modeled by an Multi-Layer Perception neural network
Bayesian Approach – Maximum A Posteriori (MAP) Estimate [2,4,5]
The probabilistic model assigns a probability for each possible location l, given the observation o.
Nearest Neighbor (RADAR[1])
database with entries (x,y,z,ssi (i=1..N)) recorded in an offline-phase
weighted average of k nearest neighbors in signal space to estimate location
N= number of visible APs
Neural Network
The fact that a basic method like the k -nearest-neighbors achieves similar results
is an indication that the complex dependencies between signals and positions are not
easily modeled by an MLP neural network.
Kalman filter: all the conditional distributions of the HMM model are continuous linear-Gaussian
HMM: In discrete case, distributions are represented as sparse matrices
Particle Filter: Approximate inference is obtained using particle filtering technique
Focuses computation on areas where most probability mass lies.

43. Wi-Fi Positioning Theory (2)
Tracking motion (recursive estimate)
Kalman Filter [6,8]
linear relation between measurement and state vector
Gaussian noise
Extended Kalman Filter [6,8]
nonlinearity dealt by updating linearization
not suitable for large disturbances
Hidden Markov Model (HMM) [6,9]
all states must be explicitly represented
separate HMM for each subset of state variable
apply Viterbi Algorithm to find most probable path
Particle Filter [6,8]
current state => N weighted state samples
samples updated using SIR after each new measurement
focus on state space regions with high probability
Markov assumption- current state depends only on immediate previous state

44. Ekahau: Theory
Location estimation is posed here as a machine learning problem
Calibration data is used to build a model to predict location given real-time signal measurements using probabilistic methods
Tracking uses a Hidden Markov model (HMM) where the locations lt are the hidden unobserved states. We compute maximum probability path l1…ln , given a sequence (history) of observations o1, …on using the Viterbi algorithm
During calibration signal strength measurements are collected at sample points at known locations
Radio signal strength measurements are performed by the mobile client device and sent to the a server
The server estimates the client location based on real-time measurements from the client at that location and the calibration data, using probabilistic methods
Tracking with history improves accuracy as signal aliasing is eliminated by using position history

45. Theory (2)
p(o|l): likelihood function/conditional probability of obtaining observations o at location l. (from calibration)
p(l): prior probability of location l (can add user profiles, tracking rails etc. to improve accuracy)
p(o): normalizing constant
Using a loss function and posterior distribution, p(l|o), we can get a optimal estimator t for the location variable.
Average accuracy obtained varies between 3ft to 10ft depending on actual implementation
likelihood function / conditional probability of obtaining observations o at location l. This is determined at calibration phase by taking empirical observations at known location

46. Hidden Markov Model HMM: λ = <N,M,{Pi},{aij},{bi(k)}>
N the number of states (S1 S2 .. SN)
M the number of possible observations
O = O1 O2 .. OT is the sequence of observations
Q = q1 q2 .. qT is the notation for a path of states
{P1, P2, .. PN} The initial probabilities
P(q0 = Si) = Pi
The state transition probabilities
P(qt+1=Sj | qt=Si)=aij
The observation / emission probabilities
P(Ot=k | qt=Si)=bi(k)

47. Viterbi Algorithm
Find most probable path given a sequence of observations O1, …On i.e. argmax P(Q| O1 O2… OT) ?
By induction:
The most probable path to Sj has Si* as its penultimate state where
i*=argmax dt(i) aij bj (Ot+1) i
= The Probability of the path of Length t with the maximum chance of doing all these things:
…OCCURING
and
…ENDING UP IN STATE Si
…PRODUCING OUTPUT O1…Ot
By Induction: The most prob path with last
two states Si Sj is the most prob path to Si ,
followed by transition Si → Sj
The most probable path to Sj has
Si* as its penultimate state
where i*=argmax δt(i) aij bj (Ot+1) i
Dynamic Programming is used here
Here location is the hidden state and signal strengths from APs are the observed states