A Mobile Indoor Positioning System Founded on

Convolutional Extraction of Learned WLAN Fingerprints

Avi Bleiweiss

BShalem Research, Sunnyvale, U.S.A.

Keywords:

Indoor Positioning System, WLAN Fingerprint, K-Means Clustering, Convolutional Extraction, KNN.

Abstract:

The proliferation of both wireless local area networks and mobile devices facilitated cost-effective indoor

positioning systems that obviate the need for expensive infrastructure. We explore a ﬂoor-level, indoor lo-

calization system to predict the physical position of a mobile device holder in an ofﬁce space by sensing a

ﬁngerprint of signal strength values, received from a plurality of wireless access points. In this work, we

devise an instructive model that tailors elemental algorithms for unsupervised ﬁngerprint learning, and resorts

to only using a single-layer convolutional neural-network, succeeded by pooling. We applied our model to a

ﬁngerprint-based dataset that renders large multi-story buildings, and present a detailed analysis of the effect

of changing setup parameters including the number of hidden nodes, the receptive ﬁeld size, and the stride

between extracted features. Our results surprisingly show that classiﬁcation performance improves markedly

with a sparser feature extraction, and afﬁrms a more intuitive gain, yet milder, as any of the number of features

or the tile size increases. Despite its simplicity, the positional accuracy we attained is sufﬁcient to provide a

useful tool for a location-aware mobile application, purposed to automate the mapping of building occupants.

1 INTRODUCTION

One of the more prominent technologies to provide

occupancy information in commercial buildings is In-

door Positioning Systems (IPS). The expansion of

location-aware mobile computing to indoors, bene-

ﬁts many real-world consumer applications, includ-

ing emergency responder, adaptive control of condi-

tioning and lighting, store navigation, and augmented

reality. For outdoors, location-based services (LBS)

typically utilize Global Positioning Systems (GPS)

that provide relatively accurate and robust positioning

solution. However, GPS require unobstructed line-

of-sight to the orbiting GPS satellites and indoors, its

signal becomes substantially compromised and pre-

sumed practically unﬁtting for resolving ﬁne-grain

object locations. A variety of alternatives have been

proposed for indoor operation ranging from visual

through infrared and ultrasound to acoustic (Ruoxi

et al., 2014) sensing. Albeit being fairly matured,

these techniques are vulnerable to environmental dis-

ruptions and therefore require costly custom hard-

ware. On the other hand, the pervasive nature of radio

frequency (RF) signals spurred extensive research of

IPS founded on wireless networks along with WiFi

enabled mobile devices (Ching et al., 2010), to lever-

age an ever-growing and widespread infrastructure.

In recent years, location ﬁngerprinting meth-

ods that harness existed wireless local area net-

work (WLAN) have been proposed for indoor spaces

(Kaemarungsi and Krishnamurthy, 2004). Most of

WLAN-based positioning systems in indoor environ-

ments rest on the Received Signal Strength Indicator

(RSSI) measure – the higher the RSSI power level,

the better the quality and speed of communications.

Typically, WLAN deployment performs a site sur-

vey on a rectangular grid of indoor locations to cap-

ture RSSI values from a multitude of dispersed wire-

less access points (WAPs), and obtain maximal space

coverage by overlapping transmittance beams. The

vector of RSSI intensities associated with each grid

point is termed the location ﬁngerprint, and a set of

predetermined ﬁngerprints then formalizes a training

database that maps all the grid nodes for locality of

reference. A mobile device held by an indoor area

occupant, captures the signal strengths from all wire-

less access points, and creates an RSSI vector sample

that is further compared to each of the database ﬁnger-

prints. The position of the person thus corresponds

to the location correlated with the most similar ﬁn-

gerprint of the database map (Marques et al., 2012)

(Zhou and Shi, 2009). Location accuracy, deﬁned as

214

Bleiweiss, A.

A Mobile Indoor Positioning System Founded on Convolutional Extraction of Learned WLAN Fingerprints.

DOI: 10.5220/0005685702140223

In Proceedings of the 5th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2016), pages 214-223

ISBN: 978-989-758-173-1

(a) Point spread. (b) Line path.

Figure 1: Physical positional distribution of training data: each mobile device location is expressed in two-dimensional

coordinates, shown for each of the ﬂoors of the three UJI university buildings in both point spread and line path formations.

the error distance from the actual position, largely de-

pends on the prediction algorithm used and the dataset

size (Chen et al., 2006). A real-world application that

utilizes a WiFi-based positioning system to track con-

struction site workers, reports a favorable location ac-

curacy of under ﬁve meters (Woo et al., 2011).

The ﬁngerprint calibration data collected in an of-

ﬁce building is often annotated with a ground-truth

location of the mobile device. However, to enable

context-aware services that scale well to large wire-

less infrastructures, mobile devices rather construct

their own spatial representation based on sequences

of unlabeled RSSI data. This requires the device to

simultaneously estimate both its location and the en-

vironment mapping (SLAM), corollary to practices in

the domain of robotics. To model WiFi signal propa-

gation in space, the work by Ferris et al. (Ferris et al.,

2007) is one technique that builds RSSI maps at run-

time without resorting to location training labels. By

reducing the high dimensional vector of values em-

anated from all the WAPs in the environment to a two

dimensional embedding space, the method facilitates

an immediate translation to extract an objective pair

of longitude and latitude coordinates of the device.

In our work, we use a pre-compiled SLAM

dataset, UJIIndoorLoc, publicly accessible from the

UCI Machine Learning Repository (UCI, 2014) and

known to date as the largest collection of WiFi ﬁnger-

prints. UJIIndoorLoc comprises distinct training and

validation sets, and is speciﬁcally targeted to provide

a reference platform for comparing results of indepen-

dent research that evaluates ﬁngerprint-based indoor

localization methodologies. The dataset represents

three buildings of the Jaume I University (UJI) cam-

pus, each of either four or ﬁve ﬂoors, with a surface

coverage exceeding 10

(Torres-Sospedra et al.,

2014). Figure 1 illustrates the physical positional dis-

tribution of the training data collection. Each mobile

device location is formulated as a pair of longitude

and latitude variables, with occupancy displayed for

each of the ﬂoors of the three target buildings, and

shown in both point spread and line path formations.

A remarkable research effort has been devoted to

learning features from unlabeled input data for clas-

siﬁcation objectives, often typiﬁed by deploying in-

creasingly complex algorithms and training a multi-

layer of representations, one layer at a time. For an-

alyzing system performance, each of the layers is pa-

rameterized by the number of features to learn, the

location coordinates of where features are computed,

and the designed encoding scheme of system inputs

and outputs. In this paper, we explore these parame-

ter choices in seeking higher accuracy rate for predict-

ing a ﬂoor-level location of a mobile device indoors,

but rather pursue an economical single-layer, convo-

lutional neural-network (CNN) architecture, trained

by simple and primitive unsupervised-learning (Duda

et al., 2001) methods. Prior work on analyzing single-

layer CNN centered primarily on imaging benchmark

datasets (Coates et al., 2011), and to the best of

our knowledge, the system we propose is the ﬁrst to

incorporate this methodology for evaluating indoor

positioning systems. Our proposed solution seam-

lessly consolidates both information retrieval (Man-

ning et al., 2008) and unsupervised machine learning

algorithms, as information retrieval (IR) is rapidly be-

coming the dominant form of data source access. Our

work closely leverages IR practices and follows ef-

ﬁcient similarity calculations directly from the well

known Vector Space Model (Salton et al., 1975).

Most schemes of feature learning have revolved

around single-layer models that are cascaded to build

a deeper hierarchy. Typically, the basic building block

of a feed-forward CNN alternates between ﬁlter banks

and a down-sampling layer, and amongst the many

modules a CNN is composed of, the unsupervised

A Mobile Indoor Positioning System Founded on Convolutional Extraction of Learned WLAN Fingerprints

215

(a) w = 6. (b) w = 8. (c) w = 10.

Figure 2: Visualization of centroids learned by k-means clustering of the UJIIndoorLoc validation set, with 25 randomly

selected tiles from each of the unlabeled WiFi ﬁngerprints, and shown for K = 100 as a function of the receptive ﬁeld size, w.

learning algorithm appears to be the most probed.

However, recent studies considered the performance

impact of other system level parameters that directly

affect the design tradeoffs of a CNN architecture, in-

cluding the number of hidden nodes, the receptive

ﬁeld size, and the sampling stride. The work by Jarrett

et al. (Jarrett et al., 2009) proved that introducing non-

linearity following the ﬁlter banks, produces sparse

features that are more suitable for subsequent pool-

ing and shown to be the most single important fac-

tor to improve object recognition accuracy. Similarly,

over-specifying a larger number of distinctly sampled

patches, is the most inﬂuential parameter governing

visual categorization results (Nowak et al., 2006).

The main contribution of our work is demonstrat-

ing that the CNN design considerations we laid out

may, in fact, be principal to the algorithm efﬁcacy in

learning WiFi ﬁngerprint features for improving the

accuracy rate of predicting the ﬂoor-level location of

a human-held mobile device. Potentially, more im-

portant even than the selection of the unsupervised

learning algorithm itself. To further commit to this

assertion, we use the simple k-means clustering al-

gorithm that requires no tuning parameters and has

not been widely adapted for deep feature learning.

Surprisingly, our analysis attributes more weight to

the choice of feature stride and increased sparsity for

ameliorating indoor location prediction, in contrast to

the common intuition broadly perceived in the image

understanding domain that a denser formulation mer-

its a higher object detection rate. The rest of this paper

is organized as follows. In section 2, we describe our

feature learning framework that incorporates k-means

clustering, coupled with a hard and a soft activation

function versions. Section 3 outlines the process ﬂow

of ﬁngerprint feature extraction prescribed in a pa-

rameterized single-layer CNN architecture, leading to

our majority-voting based k-nearest neighbor, base-

line classiﬁcation method. We then present our eval-

uation methodology for analyzing WiFi ﬁngerprint-

ing in using the UJIIndoorLoc (UCI, 2014) dataset,

and report extensive quantitative results of our exper-

iments, in section 4. We conclude with a discussion

and future prospect remarks in section 5.

2 FINGERPRINT LEARNING

In our feature learning framework, we view the

dataset of WiFi ﬁngerprint vectors as a matrix W ∈

mxn

, where m are the distinct, indoor mobile-device

locations, and n the RSSI measurements captured

from all the WAPs in all the site buildings. We then

deﬁne a ﬁngerprint tile as a contiguous subset of a ﬁn-

gerprint vector with a dimension w and of d channels.

Conventionally, w is referred to as the receptive ﬁeld

size and for ﬁngerprint data, d is ﬁxed and set to one.

Each ﬁngerprint tile is then represented as a vector

x ∈ R

of RSSI intensity values, where N = w · d.

Our ﬁngerprint learning process proceeds in sev-

eral stages. First, we extract random tiles from unla-

beled training ﬁngerprints and construct a dataset of l

randomly sampled tiles X = {x

(1)

(2)

,...,x

(l)

}, where

(i)

∈ R

. Then, every tile, x

(i)

, is optionally normal-

ized by subtracting the mean and dividing by the stan-

dard deviation of the tile vector elements. After tile

normalization, to discover features (Rajaraman and

Ullman, 2011) from unlabeled WiFi ﬁngerprint data,

we employ the exceptionally efﬁcient and simple to

tune k-means clustering algorithm that is used exten-

sively in the domain of computer vision. The k-means

unsupervised learning procedure takes the dataset X

and produces a function f : R

→ R

that maps an in-

put tile vector x

(i)

to a new feature vector of K dimen-

sionality, where K is an algorithm control parameter

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

216

that sets the number of clusters to generate, and the

feature of the mapped vector is denoted as f

. The

k-means algorithm uses the Euclidean distance mea-

sure and learns K centroids, c

(k)

, from the input tiled

data, X. Figure 2 provides visualization of bases, or

centroids, learned by k-means clustering of the UJI-

IndoorLoc validation set, with 25 randomly selected

tiles from each of the unlabeled WiFi ﬁngerprints, and

shown for K = 100 as a function of an increased re-

ceptive ﬁeld size, w. We consider two versions of the

feature mapping function f (Coates et al., 2011). A

standard 1-of-K, hard assignment encoding scheme

(x) =

(

1 if k = argmin

( j)

− xk

0 otherwise,

(1)

and a non-linear soft mapping that offers a sparse out-

put, more suitable to the CNN pooling stage, and is

governed by the following equation

(x) = max{0,µ(z) − z

}, (2)

where z

= kx−c

(k)

and µ(z) is the mean of the ele-

ments of z. Function f transforms an input tile x ∈ R

to a new representation y = f (x) ∈ R

that we use as

a learned feature extractor and apply it to our labeled

training ﬁngerprints, for classiﬁcation. Hereon, to tell

them apart, the activation functions are referred to as

k-means hard and k-means soft, respectively.

3 CNN EXTRACTION

In this step, we extract features from equally spaced

tiles that cover an entire input ﬁngerprint vector of the

labeled training dataset, and further reduce layer di-

mensionality by pooling features together over speci-

ﬁed regions. We apply either the hard or the soft ver-

sion of our k-means feature extractor, f : R

→ R

, to

a multitude of ﬁngerprint tiles, each of length w, and

compute a compact representation y ∈ R

for each

tile. The deﬁnition of a single layer architecture for

convolutional ﬁngerprint extraction ensues by deploy-

ing the function f to any number of tiles, uniformly

selected from the entire scope of an input ﬁngerprint

vector. This process is graphically staged in Figure

3. Distinctly, given a single channel, WiFi ﬁngerprint

vector of n RSSI intensity elements, construed as tiles

of a receptive ﬁeld size w each, and are evenly spaced

by a stride s of signal strength values, we formalize

the interpretation y as (

(n−w)

+1), K-dimensional fea-

ture vectors, each computed for a ﬁngerprint tile.

The extracted feature vectors, y

(i)

, are succes-

sively pooled over two evenly sized half-space bins by

computing the sum of all the y

(i)

contributing in each

















 



  



 









   



 



Tile

Representation

Classification

Features

Input

Fingerprint

mapping

pooling

Figure 3: Convolutional feature extraction using a receptive

ﬁeld size w and a stride s. Evenly distributed WiFi tiles,

(i)

, of the input ﬁngerprint vector of n RSSI intensity com-

ponents, are mapped to K-dimensional feature vectors to

form a new ﬁngerprint interpretation. Following a standard

practice in deep feature learning, the mapped vectors are

then sum-pooled over a half-space local region to derive a

feature vector Φ of dimensionality 2K, we further use for

classiﬁcation. For clarity, the stride shown for the input ﬁn-

gerprint vector is greater than the receptive ﬁeld size, but in

practice, the step s is almost always smaller than w.

region, to form the feature vector Φ of dimensionality

2K that we use for classiﬁcation. Given t, the number

of tile samples in a ﬁngerprint of size n, then the rela-

tionship t · K  n often holds to justify pooling for di-

mensionality reduction of features, a key step to merit

efﬁcient classiﬁcation computation. To our pooled,

2K-dimensional feature vectors, Φ

(i)

, constructed for

each training WiFi ﬁngerprint and a label, we apply

a majority-voting based k-nearest neighbor (Cormen

et al., 1990), baseline classiﬁer to evaluate our system

cross-validation accuracy for predicting ﬂoor-level in-

door location of a mobile device.

4 EMPIRICAL EVALUATION

To validate our system in practice, we have imple-

mented a software library that realizes the analysis of

WiFi ﬁngerprinting in several stages. After collect-

ing and cleaning the archived indoor data, we learn

unlabeled ﬁngerprints by performing k-means clus-

tering (Kaufman and Rousseeuw, 1990) on randomly

selected tiles drawn from the validation dataset. Us-

ing a single-layer CNN architecture, we then extract

features from labeled training ﬁngerprints for classiﬁ-

cation. Our learning framework behavior is primarily

governed by the setting of the number of features, K,

the receptive ﬁeld size, w, and the stride, s. We report

our experimental results on the impact of modifying

A Mobile Indoor Positioning System Founded on Convolutional Extraction of Learned WLAN Fingerprints

217

Table 1: Positional distribution of humanly held mobile devices, broken down by individual ﬂoors of each of the experimented

UJI campus buildings, for both the training and validation subsets. NA signiﬁes a non-existent ﬁfth ﬂoor in a building.

Dataset Building 1

Floor 2

Floor 3

Floor 4

Floor 5

Floor Total Mean SD

Training 1 1059 1356 1443 1391 NA 5249 1049.8 172.6

2 1368 1484 1396 948 NA 5196 1039.2 239.1

3 1942 2162 1577 2709 1102 9492 1898.4 605.4

Validation 1 78 208 165 85 NA 536 134 63.2

2 30 143 87 47 NA 307 76.8 50.2

3 24 111 54 40 39 268 53.6 33.8

these parameters by performing cross-validation.

4.1 Experimental Setup

Our work exploits the R programming language (R,

1997) to acquire the raw UJIIndoorLoc dataset from

the UCI Machine Learning Repository (UCI, 2014),

and fosters cleanup to serve useful in our software

environment. The extensive WiFi ﬁngerprint dataset,

purposed to evaluate indoor positioning systems, is

multivariate and every row of the data frame com-

prises 529 columns that coalesce features of differ-

ent categories, each represented as a vector, possibly

combining real, integer, and boolean element types.

A set of intensity values leads off to form the WiFi

ﬁngerprint vector. A ﬁngerprint is described by 520

integer elements, each spans the [−104,0] decibel-

milliwatts (dBm) range, corresponding to the weak-

est (−104dBm) and exceptionally strong (0dBm) sig-

nal when contributed by a discovered wireless access

point, or set uniformly to +100dBm to indicate an

undetected WAP. Individual RSSI measurements both

inside [−45dBm,0dBm] and under −95dBm are rare

and retain inconsequential percentage of the total in-

tensity values captured (Torres-Sospedra et al., 2014).

Each ﬁngerprint rendition follows with an asso-

ciated set of six real-world device location proper-

ties, to be either partially or in its entirety predicted

in the classiﬁcation process. This output label vector

comprises three-dimensional positional coordinates

including longitude and latitude that are measured in

meters and reference the Universal Transverse Mer-

cator (UTM) grid, and an altitude value designated by

the ﬂoor enumeration in a building ∈ [0,4]. Along

with relational space identiﬁers that further expand

the ﬁngerprint physical position and incorporate the

building id ∈ [0, 2], a categorical deﬁnition of the in-

door closed area that encompasses the wireless trace

e.g. a classroom, a lab, or an ofﬁce, and a location

proximity indicator denoting the held device of being

either inside the speciﬁed region or outside and right

in front of the door of the prescribed space perimeter.

Lastly, a global class of attributes correlate user

and mobile device information to the acquired ﬁnger-

print data, and in addition provides an accurate and

device independent time-stamp for the WiFi record-

ing that took place. Eighteen individuals in total,

user id ∈ [1,18], participated in the process of collect-

ing the training samples, each identiﬁed by a phys-

ical height that is dereferenced via a separate table.

This adds a ﬁne-grained, spatial positional dimension

to the device, found to directly impact its RSSI read-

ings (Kaemarungsi and Krishnamurthy, 2004). Mo-

bile device wise, twenty unique phone models, pop-

ulated with twenty ﬁve distinct Android OS versions,

were deployed to obtain the ﬁngerprint data. As a

three dimensional table facilitates the essential map-

ping from an itemized device list onto a phone model,

the installed system-image version, and the associated

user enumeration. User id 0 is uniquely assigned for

capturing the validation ﬁngerprint collection and its

occurrences relate several different phone models.

Quantitatively, the UJIIndoorLoc dataset is split

into a training subset of 19,937 ﬁngerprint instances,

and an exclusive validation part of 1,111 WiFi record-

ings. Training held-device locations regularly corre-

spond to pre-deﬁned reference points, whereas vali-

dation wireless readings were gathered from arbitrary

coordinates inside the buildings, without performing

any user tracking to better emulate a real-world indoor

positioning system. Notably, the validation collection

has several ﬁelds intentionally unlabeled in each of its

records, including the afﬁliated closed space and user

properties. Table 1 shows positional distribution of

humanly held mobile devices, broken down by indi-

vidual ﬂoors of each of the targeted UJI campus build-

ings, for both the training and validation subsets. The

ﬁrst and the second subject buildings are each of four

ﬂoors with more evenly spread training ﬁngerprints,

as the third building has ﬁve stories and exposes a

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

218

Table 2: Unique RSSI values presented in a ﬁngerprint tile dataset, X

,shown as a function of increased receptive ﬁeld size.

Receptive Field Size 2 4 6 8 10 12 14 16 18

Unique Fingerprints 341 927 1506 2110 2664 3194 3737 4171 4620

higher standard deviation. For this study, we are pri-

marily interested in the system performance of classi-

fying ﬂoor-level device location, hence longitude, lat-

itude, and space properties of the output feature vector

are unsubstantiated. Similarly, analysis related to user

and device attributes is outside the scope of our cur-

rent implementation and is deferred to future work.

4.2 Experimental Results

In this section we report experimental results of eval-

uating our indoor positioning system, using the UJI-

IndoorLoc validation subset to cross-validate its train-

ing companion. Our testing methodology commences

by training unnormalized WiFi ﬁngerprints in a sin-

gle layer CNN, employing both the hard and soft ver-

sions of the k-means feature mapping function f , as

we vary our system parameters K, w, and s. We then

train a baseline k-nearest neighbor (KNN) classiﬁer

and test it on the validation set. In KNN, we compute

the Euclidean-squared distance between a validation-

ﬁngerprint mapped vector against each learned train-

ing vector. Our selected k = 10 most nearest training

samples are sorted in a non-descending order, and by

a majority rule we derive a score for all the building-

ﬂoor pairs. This score is further accumulated and av-

eraged across validation ﬁngerprints, singled out for

each of the building-ﬂoor pairs, and the matching pair

corresponds then to the highest average scoring. Our

experiments ascribe a discrete value set to each of our

model parameters, K, w, and s. Respectively, we use

representations of 10, 25, 50, 100, and 250 learned

features, a series of receptive ﬁeld sizes ∈ [2,18] in

increments of two, and stride values of 1, 2, 4 and 8.

A close observation at the validation raw ﬁnger-

prints reveals a highly sparse data structure with a

fairly large proportion of about 91% of vector ele-

ments are assigned the value of +100dBm, signifying

many wireless access points are undiscovered at the

designated device locations. This review warranted

the generation of many thousands of ﬁngerprint tiles

to challenge the dispersed WiFi feature data, and en-

sure that k-means produces statistically reasoned clus-

ters. Hence, in learning our ﬁngerprint feature rep-

resentation, for every system setting of the receptive

ﬁeld size we sampled 25 random tiles from each of

the RSSI recording vectors of the unlabeled validation

set, and constructed our collection of tile datasets, X

where i ∈ 1,2,...,T , and T = 9, the number of system

choices for setting w. For 1,111 WiFi ﬁngerprints, this

yielded per dataset a total of 27,775 tiles to learn their

centroids from, each with a corresponding dimension-

ality of the speciﬁed receptive ﬁeld size. From an al-

ternate perspective, the uniqueness of values obtained

from WiFi signal measurements in a tile dataset, X

, is

also vital to clustering robustness. Table 2 depicts the

number of distinct RSSI vectors in each of the des-

ignated tile datasets, X

, as a function of our exper-

imental set of receptive ﬁeld sizes, shown in a non-

descending order. Given our discrete choices for vary-

ing the number of features, K, we hereon report our

results using receptive ﬁeld sizes that are greater than

four, with no less than 1,500 differentiable tiles.

For succinctness, we use the compact notation

λ(K, w,s) to describe our parametrically driven, IPS

implementation model that exploits a single layer

CNN. Using the UJIIndoorLoc validation dataset, we

evaluate our system for ﬂoor-level device location

pairing, by only varying one model parameter at a

time, while keeping the other two variables uniformly

constant and assigned to a prescribed default value.

First, we modiﬁed the number of features, or cen-

troids, K, in conjunction with ﬁxing the receptive

ﬁeld size to six RSSI length, and setting the fea-

ture extracting spacing between ﬁngerprint tiles to

one, generally considered optimal for convolutional

learning systems. However, systems for learning fea-

tures from two dimensional images use a stride s > 1

to step across patches and avoid excessive computa-

tional cost. In contrast, for the WiFi ﬁngerprint data,

traversing the tiles is of a considerably reduced linear

time complexity, and lets us strike a more reasonable

balance between process running time and algorithm

robustness. Figure 4(a) and Figure 5(a) show the ef-

fect on system average accuracy as the experimentally

prescribed, number of learned centroids increases, ap-

plying both the hard and soft versions of the k-means

activation function, f , respectively. For visualization

conciseness, positional performance is categorized by

the subject building id, rather than by each and every

pair of building and ﬂoor combinations. In learning

more centroids, the soft algorithm attains higher posi-

tional accuracy almost consistently for all the three

buildings. The mild performance decline observed

for K = 250 is mostly attributed to constraint clus-

tering due to our highly sparse tiling datasets. On the

other hand, k-means hard performed as expected for

A Mobile Indoor Positioning System Founded on Convolutional Extraction of Learned WLAN Fingerprints

219

0.5

0.55

0.6

0.65

0.7

10 25 50 100 250

Accuracy

Features

Bldg1 Bldg2 Bldg3

(a) Number of features, K.

0.55

0.6

0.65

0.7

0.75

6 10 14 18

Accuracy

Receptive Field Size

Bldg1 Bldg2 Bldg3

(b) Receptive ﬁeld size, w.

0.55

0.6

0.65

0.7

0.75

0.8

0.85

1 2 4 8

Accuracy

Stride

Bldg1 Bldg2 Bldg3

Figure 4: Floor-level, system average accuracy using the hard k-means feature mapping function. Shown for varying each of

the model governing parameters, (K, w, s), and each categorized by the subject building id.

0.5

0.55

0.6

0.65

0.7

10 25 50 100 250

Accuracy

Features

Bldg1 Bldg2 Bldg3

(a) Number of features, K.

0.55

0.6

0.65

0.7

0.75

6 10 14 18

Accuracy

Receptive Field Size

Bldg1 Bldg2 Bldg3

(b) Receptive ﬁeld size, w.

0.55

0.6

0.65

0.7

0.75

0.8

0.85

1 2 4 8

Accuracy

Stride

Bldg1 Bldg2 Bldg3

Figure 5: Floor-level, system average accuracy using the soft k-means feature mapping function. Shown for varying each of

the model governing parameters, (K, w, s), and each categorized by the subject building id.

the third building, but rather exposes an inconsistent

accuracy drop for both building ids, one and two. No-

tably, the 3

building occupies ﬁve ﬂoors and owns

almost half of the training ﬁngerprint data (Table 1).

Next, we stepped discretely along the second axis

of our learning model, λ(K,w, s), and quantiﬁed the

related incremental effect on our ﬂoor-level, posi-

tional classiﬁcation rate. We sought after leveraging

the k-means algorithm to learn larger receptive ﬁelds

and possibly reduce feature sparseness in a tile that

overlaps an enlarged extent of a raw WiFi ﬁngerprint.

Clearly, this necessitates an expansion of the state

space for the learning algorithm to operate on, and

thereof mandates an increased number of features to

learn. For this experiment, we tried receptive ﬁeld

sizes of 6, 10, 14, and 18 to ensure the tile datasets

are of the highest distinctive wireless recordings. We

held the stride to one RSSI length, and selected a rea-

sonably high count of 100 centroids. An overview of

our performance results for both the hard and soft ver-

sions of our k-means feature mapping is shown in Fig-

ure 4(b) and Figure 5(b), respectively. Both variants

of f behave almost identically, though the soft version

is more modestly stable, and trend towards elevating

location matching accuracy as the receptive ﬁeld size

increases, with the exception of the third target build-

ing that displays a relatively ﬂat performance.

Lastly, we evaluated the impact of altering the

stride parameter, s, on predicting the indoor location

of a humanly held, mobile device. For this experi-

ment we varied the stride over 1, 2, 4, and 8 consec-

utive RSSI elements, ﬁxed the number of centroids

to learn to 100, and set the receptive ﬁeld size w to

6. The summary of our performance results for both

the hard and soft versions of our activation function,

f , is shown in Figure 4(c) and Figure 5(c), respec-

tively. Surprisingly, our results challenge a basic intu-

ition that despite an apparent reduction of the tile sam-

pling rate, we rather demonstrate a striking upward

accuracy trend as we increase the step size between

ﬁngerprint tiles. Indifferent to the mapping function

form, ﬂoor-level accuracy for building one, for exam-

ple, depicts a marked climb from 0.58 to 0.79, or a

36% gain, as the stride treads a full extent from 1 to 8

RSSI units. In contrast to a much milder performance

gain observed when varying any of the centroid count

or the receptive ﬁeld size, model parameters.

In addition to reporting average accuracy per sub-

ject building, we were interested in evaluating our

absolute system performance of device positional

matching for the broader and ﬁner composition set

of all the possible building-ﬂoor pairs, p

, where

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

220

(a) K = 100, w = 6, s = 1.

(b) K = 100, w = 14, s = 1.

Figure 6: Confusion matrices for all the valid thirteen building-ﬂoor combination pairs, using k-means hard learning with

preferred model parameters, (K,w, s).

(a) K = 100, w = 6, s = 1.

(b) K = 100, w = 14, s = 1.

Figure 7: Confusion matrices for all the valid thirteen building-ﬂoor combination pairs, using k-means soft learning with

preferred model parameters, (K,w, s).

1 ≤ b ≤ 3 and 1 ≤ r ≤ 5, excluding the non-existent

and p

pairing choices, as prescribed by the UJI-

IndoorLoc dataset. Figure 6 and Figure 7 depict con-

fusion matrices for all predicted against actual thir-

teen building-ﬂoor combination pairs, p

, assessing

both the hard and soft designs of the feature mapping

function, f , respectively. As the parameters of our

learning model, λ(K,w,s), are ﬁxed to a limited set

of preferred values we obtained from analyzing our

building scoped performance data. Evidently, results

for stride s = 8 are by far the better achieving with a

slight edge towards the soft activation version.

Selecting a set of model parameters entails thereof

a location-pairing accuracy tradeoff, we further quan-

tify against the implied computational cost. Figure 8

shows the normalized running time of feature learn-

ing, as we vary each of the governing parameters,

(K, w, s), and respectively contrasting the hard with

the soft k-means mapping functions. Execution time

tends to rise fairly close to linear as the number of

centroids to learn increases, whereas a local minimum

is evident for w = 14 and immediately thereafter a

steep non-linear leap, as we step through the exper-

imental receptive ﬁeld sizes. Incrementing the stride,

s, consistently trends a notable commensurate decline

of running time. Overall, and as expected, the soft

feature mapping tracks well the hard version behav-

ior, albeit running slower and at a proportional scale.

5 CONCLUSIONS

We have demonstrated the apparent potential in de-

ploying a learning architecture comprised of a sin-

gle convolutional layer for extracting WiFi ﬁngerprint

features, to predict ﬂoor-level position of a human-

held mobile device, indoors. We conducted exten-

sive experiments using the recently introduced UJI-

IndoorLoc dataset, and assessed the effect of varying

neural-network parameters on location matching ac-

curacy. Despite an extremely simple learning algo-

rithm, k-means clustering, each of the model controls

we tested including centroid count, receptive ﬁeld

size, and stride, conferred either a milder or a signiﬁ-

cant impact on our system classiﬁcation performance.

One of the major challenges of our work was the

highly sparse context of the WiFi ﬁngerprint tiles with

a majority of undetected, wireless access points. Con-

A Mobile Indoor Positioning System Founded on Convolutional Extraction of Learned WLAN Fingerprints

221

0.2

0.4

0.6

0.8

10 25 50 100 250

Normalized Running Time

Features

K-means Hard K-means Soft

(a) Number of features, K.

0.2

0.4

0.6

0.8

6 10 14 18

Normalized Running Time

Receptive Field Size

K-means Hard K-means Soft

(b) Receptive ﬁeld size, w.

0.2

0.4

0.6

0.8

1 2 4 8

Normalized Running Time

Stride

K-means Hard K-means Soft

Figure 8: Normalized compute running time of feature learning. Shown for varying each of the model governing parameters,

(K, w, s), and respectively contrasting the hard with soft k-means activation functions.

sequently, some of the more ubiquitous and essential

intuitions that more centroids to learn and denser fea-

ture extraction beneﬁt performance greatly, become

considerably unsubstantiated. While conﬁrming that

more features and larger receptive ﬁeld size conclude

in a relative benign performance gain, the most sur-

prising result of our work is the striking behavior of

ascending positional accuracy as the sampling of the

ﬁngerprint tiles becomes coarser, by increasing the

stride parameter. This result may seem inexplicable

at ﬁrst observation, but the larger stride is likely to

sample more unique RSSI values to avoid underﬁt-

ting, and hence the improved location identiﬁcation.

Respectively, sparse sampling also merits a propor-

tional decline in system-level computational cost.

To the best of our knowledge, and based on lit-

erature published to date, we are unaware of indoor

positional systems with similar goals to evenhandedly

contrast our results against. The creators of the UJI-

IndoorLoc dataset (Torres-Sospedra et al., 2014) have

provided a basic reference implementation of an IPS

that employs a Euclidean distance based KNN classi-

ﬁer with k = 1. The classiﬁer operates on the origi-

nal unpacked WiFi ﬁngerprints of 520 elements each,

drawn out directly from the database. Unlike our

system that commences by performing convolutional

feature extraction, and utilizes pooling to produce an

O(n/K) compact tile representation of much reduced

dimensionality that is used subsequently for classi-

ﬁcation. Assuming that performing CNN feature

learning is a rather infrequent event in our pipeline,

our classiﬁcation phase is presumably more efﬁcient

computationally, and furthermore it scales well to

large WiFi infrastructures of thousands of wireless ac-

cess points and tens of thousands ﬁngerprint records.

For cross-validation, the reference implementation re-

ports a success rate of %89.9 for correctly identifying

a ﬁngerprint location in a building-ﬂoor pairing. De-

spite our fairly rigorous pooling scheme, to predict an

identical positional target, we recorded a respectful

0.79 accuracy of ﬂoor-level location matching, for a

sparse tile sampling with a stride s = 8.

A direct progression of our work is to evolve our

positioning system to ultimately predict the ﬁner, user

location components of longitude and latitude in con-

junction with the third dimension of the ﬂoor enumer-

ation in the target building. Similarly, we ﬁnd a use-

ful value-add in extending our work to better under-

stand the implications of the user height and the mo-

bile device type on our overall, system performance

rate. Given the limited vocabulary of RSSI values,

exploring a bag of words (Baeza-Yates and Ribeiro-

Neto, 1999) format for a more compact ﬁngerprint

representation, holds the prospect to ameliorate over-

all compute efﬁciency. We look forward to advance

our study and incorporate a more adaptive sampling

of the ﬁngerprint tile data for feature extraction, rather

than using a uniform stride, and further improve our

system accuracy. Incorporating a more comprehen-

sive linear-SVM classiﬁer, as an additional option to

our baseline KNN model that is exclusively trained

on a single parameter k, is pertinent to our work to

possibly enhance positional matching performance.

We hope that this work and others to follow, will

provide researchers a larger foundation for comparing

results of different learning algorithms that key off an

identical, indoor location dataset.

ACKNOWLEDGEMENTS

We would like to thank the anonymous reviewers for

their insightful suggestions and feedback.

REFERENCES

Baeza-Yates, R. and Ribeiro-Neto, B., editors (1999). Mod-

ern Information Retrieval. ACM Press Series/Addison

Wesley, Essex, UK.

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

222

Chen, M. Y., Sohn, T., Chmelev, D., Haehnel, D., High-

tower, J., Hughes, J., LaMarca, A., Potter, F., Smith,

I., and Varshavsky, A. (2006). Practical metropolitan-

scale positioning for GSM phones. In Ubiqui-

tous Computing (UbiComp), pages 225–242, Orange

County, CA.

Ching, W., Rue, J. T., Binghao, L., and Rizos, C. (2010).

Uniwide WiFi based positioning system. In Tech-

nology and Society (ISTAS), pages 180–189, Wollon-

gong, Australia.

Coates, A., Lee, H., and Ng, A. (2011). An analysis of

single-layer networks in unsupervised feature learn-

ing. In JMLR Artiﬁcial Intelligence and Statistics,

pages 215–223, Fort Lauderdale, FL.

Cormen, T. H., Leiserson, C. H., Rivest, R. L., and

Stein, C. (1990). Introduction to Algorithms. MIT

Press/McGraw-Hill Book Company, Cambridge, MA.

Duda, R. O., Hart, P. E., and Stork, D. G. (2001). Unsu-

pervised learning and clustering. In Pattern Classiﬁ-

cation, pages 517–601. Wiley, New York, NY.

Ferris, B., Fox, D., and Lawrence, N. D. (2007). WiFi-

SLAM using gaussian process latent variable models.

In International Joint Conference on Artiﬁcial Intelli-

gence (IJCAI), pages 2480–2485, Hyderabad, India.

Jarrett, K., Kavukcuoglu, K., Ranzato, M., and LeCun, Y.

(2009). What is the best multi-stage architecture for

object recognition? In International Conference on

Computer Vision (ICCV), pages 2146–2153, Kyoto,

Japan.

Kaemarungsi, K. and Krishnamurthy, P. (2004). Model-

ing of indoor positioning systems based on location

ﬁngerprinting. In IEEE Computer and Communica-

tion Societies (INFOCOM), pages 1012–1022, Hong

Kong.

Kaufman, L. and Rousseeuw, P. J., editors (1990). Finding

Groups in Data: An Introduction to Cluster Analysis.

Wiley, New York, NY.

Manning, C. D., Raghavan, P., and Schutze, H. (2008). In-

troduction to Information Retrieval. Cambridge Uni-

versity Press, Cambridge, United Kingdom.

Marques, N., Meneses, F., and Moreira, A. (2012). Combin-

ing similarity functions and majority rules for multi-

building, multi-ﬂoor, WiFi positioning. In Indoor Po-

sitioning and Indoor Navigation (IPIN), pages 1–9,

Sydney, Australia.

Nowak, E., Jurie, F., and Triggs, B. (2006). Sampling strate-

gies for bag-of-features image classiﬁcation. In Euro-

pean Conference on Computer Vision (ECCV), pages

490–503, Graz, Austria.

R (1997). R project for statistical computing. http://www.r-

project.org/.

Rajaraman, R. and Ullman, J. D. (2011). Mining of Massive

Datasets. Cambridge University Press, New York,

NY.

Ruoxi, J., Ming, J., and Costas, J. S. (2014). Sound-

Loc: Acoustic method for indoor localization with-

out infrastructure. Computing Research Reposi-

tory, Human-Computer Interaction, arXiv:1407.4409.

http://arxiv.org/abs/1407.4409.

Salton, G., Wong, A., and Yang, C. S. (1975). A vector

space model for automatic indexing. Communications

of the ACM, 18(11):613–620.

Torres-Sospedra, J., Montoliu, R., Martinez-Uso, A., Ar-

nau, T. J., Avariento, J. P., Benedito-Bordonau, M.,

and Huerta, J. (2014). UJIIndoorLoc: A new

multi-building and multi-ﬂoor database for WLAN

ﬁngerprint-based indoor localization problems. In In-

door Positioning and Indoor Navigation (IPIN), Bu-

san, Korea.

UCI (2014). Machine learning repos-

itory - UJIIndoorLoc data set.

http://archive.ics.uci.edu/ml/datasets/UJIIndoorLoc.

Woo, S., Jeong, S., Mok, E., Xia, L., Choi, C., Pyeon, M.,

and Heo, J. (2011). Application of WiFi-based indoor

positioning system for labor tracking at construction

sites: A case study in Guangzhou MTR. Automation

in Construction, 20(1):3–13.

Zhou, J. and Shi, J. (2009). RFID localization algorithms

and application: a review. Journal of Intelligent Man-

ufacturing, 20(6):695–707.

A Mobile Indoor Positioning System Founded on Convolutional Extraction of Learned WLAN Fingerprints

223