A Device-Independent Method for Object Localization based on WiFi RSSI Fingerprinting

values that are captured from L APs at the time t and the location. There are two critical problems need to be considered for training and testing phases, those are sele[r]

A Device-Independent Method for Object Localization

based on WiFi RSSI Fingerprinting

Thi Thanh Thuy Pham1,2, Trung Kien Dao2

Thi Lan Le2 Thi Ngoc Yen Pham2 Anh Duc Nguyen2, Thi Tam Duong3

1 University of Technology and Logistics Thuan Thanh, BacNinh

2 Hanoi University of Science and Technology No 1, Dai Co Viet Str., Hai Ba Trung, Ha Noi, Viet Nam

3 Hanoi University of Mining and Geology - 18 Pho Vien, Duc Thang, Bac Tu Liem, Ha Noi

Received: October 26, 2015; accepted: August 26, 2016

Abstract

Indoor object localization based on WiFi signals has been researched increasingly in the last decade Most of reported works concentrate on (1) building an optimal model for relationship between WiFi signal features and the object positions or (2) applying a fingerprinting method for learning and matching the best position candidates In this paper, we proposed a new combined method which uses both of above approaches for WiFi-based object localization system From our robust path-loss model, the distances between mobile user and APs (Access Points) are calculated A new radio map of distance features instead of RSSI (Received Signal Strength Indicator) values is defined in order to make the radio map independent from the WiFi receivers The matching methods of KNN and SVM are applied to estimate the mobile user position Some comparative experiments on life environment are conducted and promising results are gained on our proposed system

Keywords: Dataset, Multi-modal system, Localization, Identification, Camera network, WiFi

T

(1) (2)

(Received Signal Strengt

T : WiFi RSSI Fingerprinting, Radio map, Object localization, path-loss model

1 Introduction *

Most of WiFi-based localization systems pay

attention on processing the signal features which is

transmitted from target node to nearby APs in order to

find its location There are some popular signal

features can be used to calculate the mobile device

location, such as TOA (Time of Arrival) or TOF (Time

of Flight), TDoA (Time Difference of Arrival) and

RSSI Among these, RSSI is the most common feature

for WiFi-based object localization However, there

still exist many challenges need to be considered to

utilize this signal feature The nature of media space,

* Corresponding author: Tel.: (+84) 915.651.748

Email: thanh-thuy.pham@mica.edu.vn

including the number of APs, the obstacles, motion and direction of mobile devices (target nodes) not only causes the multi-path and non-line-of-sight (NLOS) signal propagation but also high signal attenuation and scattering Those factors change over time, hence both spatial and temporal challenges must be considered for object indoor localization based on RSSI

Two main approaches are reported for estimating location of mobile user in WiFi-based localization systems: the path-loss model and fingerprinting method The former is based on propagation or path-loss model to convert RSSI values that are transmitted

from mobile device to surrounding APs to physical

distances From these values, the coordinate of mobile

user is then calculated by geometry techniques such as

trilateration/multi-lateration or triangulation The

positioning precision of this approach much depends

on the path-loss model so many efforts have been done

for this The log-normal model in [1] is the simplest

path-loss model This model presents a simple

mapping of RSSI to the distance, therefore it does not

fully reflect the complex nature of the indoor

environment and need more improvements as in [2],

[3]

In general, mobile user localization based on

path-loss model is still an open problem It is not easy

to model the relationship between RSSI and the

relative distance in indoor environment Another

approach is fingerprinting method which is very

popular recently with higher positioning accuracy

This method includes two phases of training or

building fingerprint database and testing The radio

map is a core component of fingerprint database It is

a set of reference points with their known positions and

RSSI values captured from nearby APs A testing

sample will be mapped to the fingerprinting database

to return the corresponding location of the mobile

device [3], [4]

In this paper, we propose a new-defined radio

map of the distance features instead of RSSI values

An optimal model for RSSI-distance relationship is

applied into testing phase in order to calculate the

distance values from the RSSI features These values

will be mapped to fingerprint database by KNN or

SVM to find the most suitable locations for each target

node of mobile user

The rest of paper is organized as follow Section

II presents a combined positioning system of

RSSI-distance model and fingerprinting map with KNN and

SVM The comparative experiments in life

environment and the results are shown in section III

Conclusion and future directions will be finally

denoted

3 Proposed system

The flowchart of our WiFi-based object

localization system is illustrated in Fig 1 There are

two main phases of training and testing in the diagram

The first phase is processed off-line with the

construction of radio map to make fingerprint

database Normally, a radio map contains reference

point coordinates and corresponding RSSI values

However, in our proposed system, RSSI values are

replaced by distance values from each reference point

(RP) to all available APs In testing phase, a mobile

device continuously scan signals from nearby APs and

sends corresponding RSSI values to a server We

propose a robust path-loss model to calculate the distance values from mobile device to all available APs based on RSSI values These distance values will

be matched to fingerprint database by methods of KNN or SVM to find the best candidates for mobile user locations

Fig 1 The testing environment

3.1 The path-loss model

In previous work we proposed a path-loss model [5] to estimate the distance between Wifi receivers and APs This model is constructed based on the fact that the signal strength of a radio wave is attenuated when traveling through a certain environment We begin with the empirical model which is widely used in previous work as follow:

where is known signal power in dBm unit at a reference distance , is signal power at an unknown distance , and is the path-loss exponent indicating the rate at which the path loss increases with distance

By practical experiments we can determine the parameters , and The distance between a certain location and AP can be finally estimated by the given RSSI

Equation (2) is a propagation model describing the relation between RSSI and distance in the environment without obstacles between AP and the receiver When walls and floors are taken into account, the attenuated model becomes:

(2)

Where is the number of walls and floors lying between AP and the receiver, is the thickness of the

ith wall/floor, is the angle of arrival corresponding

to the ith wall/floor, and is the attenuation factor per wall/floor thickness unit In general case, can be

extended to be dependent specifically on each

wall/floor

Equation (2) is a deterministic model that means

the uncertainty of RSSI at a distance is not taken into

account To address this limitation, we proposed to use

a probabilistic model In reality, given the RSSI , the

distance might not be exactly the value calculated

from (Eq.2) but in within a range around this value,

which is denoted by To be more precise, will be

the nominate value of the distance with highest

probability Given a RSSI , the distribution of the

distance is assumed to follow the normal (or Gaussian)

distribution with median will be:

(3)

where is the standard deviation, which is also a

function of For simplicity, is assumed to be

related to by a following linear function:

where is a constant, which also needs to be

determined by experimental data

3.2 The fingerprinting method

The radio map in fingerprinting method is

defined as follow:

(5)

coordinate of the ith reference point and

is fingerprinting matrix, with n being

the number of training samples at each reference point

values that are captured from L APs at the time t and

the location

There are two critical problems need to be

considered for training and testing phases, those are

selection of signal feature and matching algorithm We

need to find out a robust and discriminative

representation of signal feature which is useful for later

matching or classification A good classifier also plays

an important role in giving better final results for the

system

In this paper, we propose to use the distance

feature instead of RSSI which is usually used in

WiFi-based localization systems In training phase, it is

defined as the distance from the i th RP to the Lth

AP (Fig 2) and it is a stationary and predefined value

so it is much more stable than RSSI

The radio map in Eq.5 then has a fingerprinting

includes distance samples from the

i th RP to L APs This results to a reliable and stable

radio map even in case some APs may be inactive at a certain point of time Furthermore, the cost for setting and updating the radio map is much lower than usual

It is only rebuilt when we deploy new APs and RPs or discard them from the WiFi-based localization system Additionally, the update information is just distance data rather than RSSI which is much more expensive

in collecting and selecting the representative value for each RP

Fig 2 The radio map with a set of positions of reference points (RP) and the distance values

from each RP to L APs

In testing phase, the RSSI values scanned from nearby APs by mobile device will be converted to the corresponding distance figures by the above mentioned path-loss model They will be compared with the training data to find the best matches The matching methods that are used in this paper are KNN and SVM These are popular solutions for regression

or classification We propose to use them in order to show the comparative results for different input features of RSSI or distance on two distinctive classifiers

KNN regression: in KNN, prediction for a new

instance is based on its nearest neighbors in the training data There are three main ingredients associated with this method, those are (1) the similarity measure (distance) between instances or objects; (2) the number of neighbors to be taken in prediction; (3) the weight of the neighbors; Euclidean and Manhattan distances are two common geometric measures, in which Euclidean is the most used in WiFi-based localization system [6], [7] In this paper, we evaluate KNN method on both of these measures to show the comparative results

In the proposed radio map, each RP is

L dimensional space In learning phase, we store all

these training data D with their dependent variables In

this case, the dependent variables are equivalent to the positions of RPs in the environment In prediction,

for a new instance z and for each instance d in D, we compute the similarity between d and z by Euclidean

(Eq.6) or Manhattan (Eq.7) distance measures

(6)

(7)

A set NB(z) of the nearest neighbors of z with

|NB(z)|=k is also determined and then the estimated

location for z is calculated To find out an optimal k,

we test on the empirical data with k in range of (1:200)

by an error function (Eq.8) for each k

(8)

Where is the estimated position and y is true

position Finally, the predicted location of z is

calculated by the weighted sum of the k neighbors

(Eq.9)

(9)

where w shows the weights that are chosen by

(Eq.10)

(10) where and are constants used to define the

curve of the exponential functions; t is the time when

the mobile device scans a WiFi signal and belongs

to the time of WiFi signal scanning at each

corresponding RP in fingerprinting phase

SVM classifier: In this paper, we apply

non-linear and multi-class SVM for classifying a new

instance z In non-linear SVM, kernel trick is used to

maximum-margin hyperplanes in order to create

non-linear classifiers from non-linear ones It is defined by a

nonlinear kernel function which maps the input data

into high-dimensional feature space so as to an easier

separation is made in new space

The localization problem is now represented by

with N

being the number of reference points or the classes,

is equivalent to the class labels and is a

fingerprinting matrix of the distances from the i th

reference point to L APs and

Let is a kernel function, and let

Then the decision function of the max-margin

classifier using kernel function is given by (Eq.11)

(11)

subject to

(12)

The chosen kernel function is RBF kernel (radial basis function) (Eq.13)

(13) For our localization problem, the classification belongs to multiple classes, with the number of classes

is equal to the quantity of RPs Therefore, in this case,

we propose to use one-versus-one multiclass SVMs [8] with the idea of training one SVM per pair of class

classes To classify a new test example of z, all SVMs

are assessed, and a strategy of majority voting is done

to find out the target class label (Eq.14)

(14)

4 Experimental results

4.1 The testing environment and data collection

The experiments are carried out on the 8 th floor

of an 11-story building at the MICA International Research Institute There are 20 WiFi APs distributed

at different positions on three floors of 8, 9 and 10 (Fig 3)

Fig 3 The distribution of APs in testing environment

To collect the RSSI fingerprints, we develop an Android platform application for mobile device to automatically scan Wifi signals from surrounding APs, with a scanning period of 2, 3, 5, 10 or 60 seconds In order to collect the fingerprints, the collector should hit the marker button in the application before each turn

to mark the straight path he takes The collected data can be either written into an XML file or uploaded to database if the mobile device has the Internet connection Because the coordinates of the markers are known, the coordinates of the scanned positions between every two markers are then interpolated by a Windows application developed by us Finally, these positions are saved into a binary file as the Wifi

fingerprint database for the 8 th floor

The binary file is updated periodically in order to

collect the latest fingerprints which contain the

changes of the environment Thus, in matching phase,

besides the weight based on distance, we also consider

the fingerprint changes according to time This means

the older a fingerprint is, the lighter the weight (Eq 10)

of that fingerprint is Updating fingerprint database

frequently will make our localization system more

precise and reliable

In our experiment, we run the application with a

period of 2 seconds on different devices such as

Samsung Galaxy S I9000, Samsung Galaxy Note

N7000, Samsung Galaxy Tab GT-P7500, Huawei

tablet The 700 and 1200-fingerprint versions are

shown in (Fig.4) with 700 and 1200 RSSI samples

being collected respectively

Fig 4 The RSSI fingerprinting map of (a) 700 samples

and (b) 1200 samples

4.2 The results and evaluation

- The results from KNN method: According to

the calculation from the error function (Eq.8), the

optimal k should be 9 Figure 5 and Table 1

demonstrate the comparative results between RSSI and

distance features with path-loss model when using

Euclidean and Manhattan measures for fingerprinting

map of 1200 samples

Table 1 Localization results of KNN using Euclidean and

Manhattan measures for RSSI and distance samples

Maximal Error

Average Error

Error at Reliability

of 90%

Euclidean-Distance 23.70 3.15

5.37

5.91

Manhattan-Distance 20.59 3.13

6.10

6.40

The first conclusion can be drawn from our

results is that matching the query point using the

path-loss model does remove the dependence of fingerprint

on WiFi signal receivers Despite the fact that fingerprint data is collected by several devices the localization system gives reliable localization results Second, while Euclidean formula shows a greater domination in Error at Reliability of 90%, the Maximal Error decreases considerably when using Manhattan measure to compute the differences between query point and the fingerprints

Figure 6a and Table 2 show the localization results when using 700 newest fingerprints and total

1200 fingerprints with Euclidean distance being used

We can see that using newest fingerprints can effectively reduce the Maximum Error and Average Error of the localization system but when using the combination between new and old data, it gives a better Error at Reliability at 90\% Therefore, we can utilize this trade-off depending on the purpose of the experiment

- The results from SVM classification: Figure 6b

and Table 3 show the localization results of SVM when using 700 newest fingerprints and total 1200 fingerprints As we can see, when using 700-fingerprint version, SVM proves a higher reliability compared to when using the 1200-fingerprint one It can be concluded that the efficiency of SVM will decrease due to the changes in environment Therefore, in both cases the localization results of SVM are less reliable than that of KNN algorithm

Table 2 Localization results of KNN using Euclidean and

Manhattan measures for RSSI and distance samples

Maximal

Error

Average Error

Error at Reliability

of 90%

1200 Fingerprints 23.70 3.15

5.37

700 Fingerprints 19.76 3.06

5.91

Table 3 Localization results of SVM for distance samples

Maximal Error

Average Error

Error at Reliability

of 90% SVM 1200

Fingerprints 46.33 4.77

6.68

SVM 700 Fingerprints 15.62 2.83

6.08

Fig 5 The testing results with Euclidean and Manhattan measures applied to RSSI and distance samples: (a)

localization result, (b) the distribution of positioning error and (c) localization reliability

Fig 6 The localization results, distribution of positioning error and localization reliability for two fingerprinting

versions using (a) KNN with Euclidean measure and (b) SVM method

5 Conclusion and future works

In WiFi fingerprint-based object localization, a

well-defined radio map plays critical role for

positioning performance of the system In this paper,

not RSSI but distance samples used in radio map will

make it not only more robust and reliable but also

decrease the labor and time in collecting fingerprints

Experimental results show that building radio map

based on distance samples helped to increase the

invulnerability of the system to environment changes

In the future, an integrated experiment will be

carried out The WiFi-based positioning method will

be combined with vision-based solution to enhance

reliability of our localization system

Acknowledgments

This work is supported by National Foundation

for Science and Technology Development, Vietnam

(NAFOSTED) in the project under grant number

102.04-2013.32

References

[1] Rappaport, T.S., Wireless communications: principles

and practice Vol 2 1996: prentice hall PTR New

Jersey

[2] Lourenço, P., et al A received signal strength

indication-based localization system in Control &

Automation (MED), 2013 21st Mediterranean Conference on 2013 IEEE

[3] Chen, Z., et al., Fusion of WiFi, smartphone sensors and landmarks using the Kalman filter for indoor localization Sensors, 2015 15(1): p 715-732

[4] Zhu, J.Y., et al Spatio-temporal (ST) Similarity Model for Constructing WIFI-based RSSI Fingerprinting Map for Indoor Localization in International Conference on Indoor Positioning and Indoor Navigation 2014

[5] Dao, T.-K., T.-T Pham, and E Castelli A robust WLAN positioning system based on probabilistic propagation model in Intelligent Environments (IE),

2013 9th International Conference on 2013 IEEE [6] So, J., et al., An improved location estimation method for wifi fingerprint-based indoor localization International Journal of Software Engineering and Its Applications, 2013 7(3): p 77-86

[7] Farshad, A., et al A microscopic look at wifi fingerprinting for indoor mobile phone localization in diverse environments in Indoor Positioning and Indoor Navigation (IPIN), 2013 International Conference on 2013 IEEE

[8] Hsu, C.-W and C.-J Lin, A comparison of methods for multiclass support vector machines Neural Networks, IEEE Transactions on, 2002 13(2): p

415-425