WiDeep: WiFi-based Accurate and Robust Indoor Localization System using Deep Learning

Abstract

Robust and accurate indoor localization has been the goal of several research efforts over the past decade. Due to the ubiquitous availability of WiFi indoors, many indoor localization systems have been proposed relying on WiFi fingerprinting. However, due to the inherent noise and instability of the wireless signals, the localization accuracy usually degrades and is not robust to dynamic changes in the environment. We present WiDeep, a deep learning-based indoor localization system that achieves a fine-grained and robust accuracy in the presence of noise. Specifically, WiDeep combines a stacked denoising autoencoders deep learning model and a probabilistic framework to handle the noise in the received WiFi signal and capture the complex relationship between the WiFi APs signals heard by the mobile phone and its location. WiDeep also introduces a number of modules to address practical challenges such as avoiding over-training and handling heterogeneous devices. We evaluate WiDeep in two testbeds of different sizes and densities of access points. The results show that it can achieve a mean localization accuracy of 2.64m and 1.21m for the larger and the smaller testbeds, respectively. This accuracy outperforms the state-of-the-art techniques in all test scenarios and is robust to heterogeneous devices.

Full text

WiDeep: WiFi-based Accurate and Robust Indoor
Localization System using Deep Learning
Moustafa Abbas
Dept. of Comp. and Sys. Eng.,
Alexandria University
Alexandria, Egypt
m.abbas@alexu.edu.eg
Moustafa Elhamshary
Dept. of Comp. and Cont. Eng.,
Tanta University
Tanta, Egypt
mostafa.elhamshary@f-eng.tanta.edu.eg
Hamada Rizk
Dept. of Comp. Sci. and Eng.,
EJUST, Alexandria, Egypt
& Tanta University, Tanta, Egypt
hamada.rizk@ejust.edu.eg
Marwan Torki
Dept. of Comp. and Sys. Eng.,
Alexandria University
Alexandria, Egypt
mtorki@alexu.edu.eg
Moustafa Youssef
Dept. of Comp. and Sys. Eng.,
Alexandria University
Alexandria, Egypt
moustafa@alexu.edu.eg
Abstract—Robust and accurate indoor localization has been the
goal of several research efforts over the past decade. Due to the
ubiquitous availability of WiFi indoors, many indoor localization
systems have been proposed relying on WiFi fingerprinting.
However, due to the inherent noise and instability of the wireless
signals, the localization accuracy usually degrades and is not
robust to dynamic changes in the environment.
We present WiDeep, a deep learning-based indoor localization
system that achieves a fine-grained and robust accuracy in
the presence of noise. Specifically, WiDeep combines a stacked
denoising autoencoders deep learning model and a probabilistic
framework to handle the noise in the received WiFi signal and
capture the complex relationship between the WiFi APs signals
heard by the mobile phone and its location. WiDeep also intro-
duces a number of modules to address practical challenges such
as avoiding over-training and handling heterogeneous devices.
We evaluate WiDeep in two testbeds of different sizes and
densities of access points. The results show that it can achieve
a mean localization accuracy of 2.64m and 1.21m for the larger
and the smaller testbeds, respectively. This accuracy outperforms
the state-of-the-art techniques in all test scenarios and is robust
to heterogeneous devices.
Index Terms—WiFi, Deep learning, indoor, localization, finger-
printing
I. INTRODUCTION
As people spend most of their time indoors, academia and
industry have recognized the value of the indoor localization
problem and have devoted much effort and resources into
solving it [1], [2]. Due to the wide-spread coverage of WiFi
and the support of the IEEE 802.11 standard by the majority of
mobile devices, most proposed indoor localization systems are
WiFi-based including propagation- and fingerprinting-based
techniques [3]–[10]. Propagation-based techniques, e.g. [3],
[4], [11], aim to model the relation between the received
signal and distance without site surveying. Despite their ease
of deployment without the need for prior calibration, these
This work has been supported in part by a grant from the Egyptian National
Telecommunication Regulatory Authority(NTRA).
techniques do not work well with heterogeneous phones
and their accuracy is usually less than fingerprinting-based
techniques.
On the other hand, fingerprinting techniques leverage the
recorded WiFi APs signatures (i.e. fingerprints) to estimate
the device location. Typical fingerprint-based WiFi localiza-
tion techniques work in two phases: The first one is the
offline phase (i.e., calibration) during which the received
signal strength (RSS) readings from the multiple access points
(APs) installed in the area of interest are recorded at known
locations. Then, in the tracking phase, RSS measurements
from the detected APs at an unknown location are matched
against the stored fingerprints to estimate the best location
match either deterministically, e.g. [12], or probabilistically,
e.g. [6]. Fingerprinting-based techniques are widely adopted
due to their relatively good accuracy. Practically however,
the deployment of such techniques faces major challenges
due to the inherent noise in the wireless signals that affects
localization accuracy [13], [14]. Therefore, many systems have
been proposed to address these challenges over the years, e.g.
[6], [15]–[17]. Probabilistic techniques such as [6], [18] can
counter the inherent wireless signal noise in a better way than
deterministic techniques [12]. However, they usually assume
that the signals from different access points are independent
to avoid the curse of dimensionality problem [19]. This leads
to coarse-grained accuracy. Hybrid techniques, e.g. [15], [16],
[20]–[22], leverage the sensors that are available on high-
end smartphones to combat the wireless channel noise. Other
techniques, e.g. [8], [23], leverage the detailed channel state
information obtained from specialized WiFi chips to combat
the noise. However, both of the last two categories are not
supported by the vast majority of mobile devices, limiting their
ubiquitous deployment.
In this paper, we propose WiDeep: a WiFi-based indoor
fingerprinting localization system that can achieve robust and
high accuracy tracking in the presence of device heterogeneity.

To do this, WiDeep builds on deep learning to automatically
capture the non-linear and correlated relation between the dif-
ferent access points at different fingerprint locations, without
assuming access points’ independence as in current probabilis-
tic techniques. However, leveraging a deep network alone as in
[8] may not lead to the required performance in the presence
of device heterogeneity which can be considered as a form of
noise. To ensure the robustness and the generalization ability
of the system in this challenging scenario, we adopt a deep
network model utilizing stacked denoising autoencoders to
robustly extract a good representation of the relation between
the noisy WiFi scans and the different fingerprint locations.
Furthermore, WiDeep also employs a regularization technique
to avoid model over-fitting and boost the robustness of the
system.
During tracking, the output of the deep learning models is
fused using a probabilistic framework to further handle the
noise in the input signal.
We implemented and deployed WiDeep on different Android
phones and evaluated its performance in two different testbeds:
a 629m2university building and a 65m2residential apartment.
The two buildings have different layouts and WiFi APs den-
sities. Our results show that WiDeep can achieve a consistent
mean accuracy of 2.64m and 1.21m in the two testbeds
respectively under different scenarios. This is better in mean
accuracy than traditional fingerprinting techniques and other
basic deep learning techniques by at least 29.8% and up to
168%. This accuracy is maintained under significant decrease
in the access points density as well as under heterogeneous
devices, highlighting WiDeep promise as a robust and accurate
indoor localization technique.
The rest of the paper is organized as follows: Section II
presents an overview on how WiDeep works and its mathe-
matical model. Section III presents the details of the WiDeep
system. We evaluate the system performance in Section IV.
Finally, sections V and VI discuss related work and conclude
the paper respectively.
II. SYSTEM OVERVIEW AND MATHEMATICAL MODEL
A. System Overview
Fig. 1 shows the system architecture. The system runs in two
phases: offline training phase and online localization phase.
During the offline phase, the system builds Ndeep neural
networks corresponding to Nfingerprint training points (i.e.,
a deep neural network for each reference point). This helps in
scaling the system to large areas as well as keeping the model
size small and easier to train.
To collect the training data, the Signature Collector module
is used to scan for the APs and their associated signal strengths
at the different fingerprint locations in the area of interest.
These measurements are opportunistically transfered to the
WiDeep server in the cloud. The Preprocessor module is
used to transform the WiFi measurements to fit the format
required in the deep network training model as well as perform
the required normalization. The preprocessed data is then
fed to the Noise Injector module to corrupt the original
Signature Collector
Online
Offline
Preprocessor
Noise Injector
Model Trainer
RSS Collector
Preprocessor
Loc
c
1 Loc
c
2 Loc
c
N
Model 1
Model 2
Model N
Probabilistic Localizer
Probabilistic Estimator
Deep
p
Model Estimator
Fig. 1: WiDeep system architecture.
measurements by injecting artificial noise to the collected WiFi
scans. This helps not only in simulating the distorted wireless
channel but also in reducing the model over-fitting by forcing
the model to learn the inherent feature of the data. For each
fingerprint point, the original and corrupted noisy data are
forwarded as input to the Model Trainer module. This module
is responsible for creating and training a stacked denoising
auto-encoders deep model corresponding to each fingerprint
point. Moreover, it enhances the robustness of the model by
avoiding over-training. Finally, all the trained models for the
different fingerprint locations are stored to be used during the
online localization phase.
During the online localization phase, the user is tracked
in realtime. The process starts by scanning for the APs and
their RSSs at the unknown user location. This data is first
preprocessed (shaped and normalized) to fit with deep learn-
ing model input. Then, the Probabilistic Localizer module
leverages the different deep models output in a probabilistic
framework to estimate the most probable user location using
the Deep Model Estimator and Probabilistic Estimator sub-
modules respectively.
B. Mathematical Model
Without a loss of generality, we assume a 2D physical
area of interest Lcontaining Maccess points. Ndiscrete
fingerprint locations are spread over the entire area, where
training data is collected. During the online localization phase,
a user holding a mobile device at an unknown location l∈L
scans for the nearby APs. Let a vector xiof Mdimensions
represents a single WiFi scan. Each entry iin this vector
is the received signal strength reading from access point i.
The problem then becomes: given a signal strength vector

x=(x1, ..., xM), we seek to find the fingerprint location li
that maximizes the probability P(li|x). In the next section,
we discuss the details of how WiDeep combines deep learning
and a probabilistic framework to achieve high accuracy and
robust localization in the continuous space.
III. THE WiDeep SYSTEM
We present the details of the offline models construction
phase and the online localization phase. We start by the
preprocessor module as it is common to both phases.
A. The Preprocessor
This module is responsible for mapping the recorded WiFi
RSS readings, xi, to the corresponding feature vectors. Note
that since not all MAPs that are installed in the area of interest
can be heard in every scan, this module assigns the weakest
RSS, i.e. −100dBm, to the APs that are not heard in a given
scan. This allows us to fix the feature vector size that is input to
the machine learning model. After that, the input RSS values
ranges are normalized to be in the range between [0,1] for
each AP. Features normalization is known to speed up model
training and increase the model robustness [24].
B. Offline Models Construction Phase
During this phase, the system builds Ndeep models
corresponding to Nfingerprint training points (i.e., a deep
model for each point). The system also addresses a number of
challenges including handling the noise and fluctuation of APs
signals as well as reducing the model over-fitting to training
data, allowing for better generalization and robustness.
We choose stacked denoising autoencoders as our model
as they are able to extract the latent features from noisy
data. Autoencoders are unsupervised learning models, where
their goal is to learn a concise mapping that can regenerate
the input to the autoencoder [25]. Denoising autoencoders
extends traditional autoencoders to handle noisy data in a
better way. Specifically, instead of feeding the original input
to the denoising autoencoder, we feed it a noisy version. This
allows the hidden layer of the autoencoder to learn important
features (Fig. 2). Specifically, the denoising autoencoder is
trained by first corrupting the input RSS vector xto obtain
vector ˜x. The goal is to learn the parameters of the hidden
layer hso that the output (ˆx) of the autoencoder matches the
original uncorrupted vector x. By using a noisy version of
the input, the autoencoder is forced to learn the latent features
of the input data. Note that the weights between the hidden
and output layers are the transpose of the weight between the
input and hidden layer, reflecting the decoding process of the
autoencoder. Training is performed using the gradient descent
algorithm, where the least square error between the original
input data xand the reconstructed data ˆxis used as the loss
function to adjust the weights.
In the balance of this section, we first describe how to
introduce noise to our input data. Then we provide the details
of our deep model.
0
0
WWT
Corruption
x̃
x̂
x
h
Fig. 2: Denoising Autoencoder: A sample xis stochastically
corrupted to produce ˜x. Next, the the hidden layer (h) maps ˜x
to the output ˆxto produce a reconstructed version of x. Error
is measured by the L(x, ˆx)loss function that compares the
reconstructed output with the original noise-free input.
1) Noise Injector: This module aims to enhance the model
ability to handle the noisy input data. This is achieved by
generating corrupted variations of the collected WiFi scans.
This also have the added advantage of reducing over-fitting
and coping better with devices heterogeneity as we quantify
in Section IV. To do that, the module adds stochastic noise
to the input data through two different techniques: Masking
corruption and Additive Gaussian corruption.
Masking corruption method: The intuition behind this
method is that the number of access points detected at a
fixed location varies with time due to multipath and fading
effects [26]. The masking method leverages this fact to emulate
fluctuating access points [27]. The idea is to generate a random
binary vector with specific probability of its elements to be
zeros determined by the corruption fraction parameter (f).
This generated binary vector is then multiplied by the original
input to get a noisy input signal, where the entries of the APs
corresponding to the zero random bits are dropped, emulating
not hearing them (Fig. 3a).
Additive Gaussian corruption method: Due to the noise in
the wireless channels and the diversity of the WiFi chips in
different devices, the magnitude of the RSS may be shifted
with some variance. Therefore, to emulate this behavior, this
technique adds white Gaussian noise with a specific standard
deviation sto the different entries of the RSS vector (Fig. 3b).
The synthesized vector is finally re-normalized so that all
entries are between 0 and 1.
2) The Model Trainer: Fig. 5 shows the WiDeep deep
model. It consists of a number of stacked denoising au-
toencoders, one stack for each fingerprint location. The used
activation function is the Sigmoid function formulated as [25]:
sigmoid(x)= 1
1+e−x(1)
The Sigmoid function ensures that its output will range be-
tween 0 and 1.
To train WiDeep model end-to-end, we use two training
stages: (1) greedy layer-wise pre-training stage and (2) a fine-

0.2 0.0
0.5 0.3
0.0 1.0
1.0 1.0
0.0 0.0 0.5 0.3
Input RSS vector
Mask
Masked vector
(a) Masking corruption method.
0.2 0.0
0.5
0.3
0.05 -0.01 0.07 0.06
0.25 0.0
0.57 0.36
Input RSS vector
Noise vector
Noisy vector
(b) Additive Gaussian corruption
method.
Fig. 3: Examples of different corruption techniques applied to
the normalized RSS signal.
tuning stage. Now, we discuss the details of each of these
stages.
1) Pre-training Stage: Fig. 4 shows the steps of the pre-
training deep neural network of stacked denoising autoen-
coders. The goal of this stage is to find good initial weights
for the different layers of the network, instead of using
initial random weights [28]. This is known to speed up the
convergence of the training, reduce the possibility of falling in
a local minima, and avoid the vanishing gradient problem [29].
In this stage, each autoencoder in the stack is trained
independently using the output of the previous autoencoder as
its input (input data for the first autoencoder). Each denoising
autoencoder consists of an encoder and a decoder. The encoder
tries to generate a latent representation (hidden layer param-
eters) of the input data while the decoder tries to reconstruct
the input data based on the generated (latent) code from the
encoder.
2) Fine-tuning Stage In this stage, we train the model
(Fig. 5) end-to-end. The weights are initialized to those
obtained from pre-training stage. Then, each input training
sample (WiFi scan) is passed through network to obtain the
reconstructed data of the input scan using forward propagation.
The sum of squared difference between the original input data
and the reconstructed data (i.e., output of the deep network) is
used as the loss function to adjust all the weights in different
layers with the gradient descent algorithm [30].
C. Reducing Model Over-training
To further reduce the possibility of model over-fitting, we
also use dropout regularization [27] during the fine-tuning
phase. To do that, some hidden neurons in the deep neural
network are temporarily dropped out stochastically (as illus-
trated by the crossed circles in Fig. 5). This dynamic change of
the network structure allows the network to generalize better
and hence become more robust to changes.
Finally, the learned weights of the Ndeep models for the
Nfingerprint locations in the area of interest are saved. Later
on, during the online localization phase, these weights are used
to estimate the unknown location of the mobile device as we
explain in the next section.
D. Online Localization Phase
During this phase, we harness the learned models to esti-
mate the unknown locations of the test inputs. Specifically, we
input the current WiFi scan at the unknown user location to
the different deep models at each fingerprinting location and
leverage the reconstruction similarity in a probabilistic frame-
work to estimate the most probable location. The intuition is
that the reconstructed scan will be closer to the input scan
in the fingerprinting locations near the actual user location.
hence, these locations should have a higher score/probability
compared to far-away locations.
More formally, the user is standing at an unknown location
lreceiving WiFi information that is preprocessed to obtain
a signal strength vector x=(x1, ..., xM), where Mis the
total number of APs in the environment. We want to find the
probability of being at a fingerprint location liin the area of
interest given the received signal strength vector x. That is, we
want to find P(li|x). Using the Bayes theorem, the posterior
probability P(li|x)is given as:
p(li|x)=p(x|li)p(li)
p(x)=p(x|li)p(li)
N
i=1 p(x|li)p(li)(2)
Where p(li)is the prior probability that the phone is located
at a given fingerprint location liand Nis the number of
locations in the fingerprint database (i.e, the training locations).
Assuming that all locations are equally probable1, Equation 2
can be rewritten as:
p(li|x)= p(x|li)
N
i=1 p(x|li)(3)
In traditional fingerprinting systems, e.g. [6], p(x|li)is
usually obtained by assuming the independence of the APs
using the RSS histograms, which does not capture the rich
and correlated relation between the different APs. On the
contrary, to calculate p(x|li)WiDeep leverages the constructed
offline deep learning models. Specifically, reconstructed ver-
sions of the input scan xiare obtained from each deep model
along with the associated similarity score to the input signal.
To obtain this similarity score, we use a radial basis kernel
as a similarity function since its output is bounded within 0 and
1 and can therefore be interpreted probabilistically. Denoting
the output of the similarity function as p(x|li)for the ith
model, we have that:
p(x|li)= 1
n
n

j=1
e−
xij −ˆxij 
λσ (4)
Where xij and ˆxij are the original and the reconstructed input
data of the jth scan respectively, σis the variance of the input
scans, λis selected to be a scaled version of the coefficient of
variation (CV) of the input scans, and nis the total number of
scans used in location determination. We quantify the effect
of these parameters on performance in Section IV.
1If the user location profile is known, it can be used directly in Equation 2.

Autoencoder 1
Original
RSS vector
Autoencoder 2
Autoencoder 3 Autoencoder 4
Corruption
Fig. 4: Greedy layer-wise pre-training process. The latent vector of every trained layer is used to train the subsequent layer.
1
2
N
Model
el
1
M
d
l
Model
l
l
dl
l
el
2
Mo
de
el
Model
del
el
el
N
2
1
Corruption
N
Fig. 5: Deep network architecture. Crossed nodes are example
of dropped-out neurons. Note that the output of this network
is connected to the probabilistic location inference module.
Till now, we can assign different probabilities to the discrete
fingerprinting locations based on the input scan(s). To enable
tracking the user in the continuous space, WiDeep estimates
the user location as the center of mass of all fingerprinting
points [31], taking the probability of each reference point
P(li|x)as its weight. Hence, the user location lis estimated
as
l=
N

i=1
p(li|x)li(5)
IV. EVALUATION
In this section, we evaluate the performance of WiDeep in
two typical indoor environments: a university building floor
and an apartment. We start by describing the data collec-
tion methodology. Next, we analyze the effect of different
parameters on the WiDeep system performance. Finally, we
compare our system to the state-of-the-art WiFi fingerprinting
techniques that are Horus [6] and DeepFi [8].
A. Data Collection
To collect the necessary data for evaluation, we deployed
our system in two buildings with different layouts and APs
densities (Table I). The first one is a floor of our university
building with a 37m×17marea containing offices, labs,
meeting rooms as well as corridors (Fig. 6). The second one,
shown in Fig. 7, is an L-shaped private studio apartment with
a14.5m×4.5marea. In both datasets, we leverage the RSSs
of pre-installed WiFi APs in the building or overheard from
TABLE I: Summary of used testbeds parameters.
Testbed University Apartment
Area 37m×17m14.5m×4.5m
Number of APs 122 59
Density of APs (AP/m2)0.19 1.05
Training points 29 81
Testing points 19 58
nearby floors/buildings (university floor has an overall of 122
APs whereas the apartment has 59 APs). 7200 samples in total
are collected at 48 different locations in the university dataset.
For the apartment dataset,2000 samples are collected at each
point of 139 different locations. The data is collected by five
participants using different Android phones (e.g., Samsung
Galaxy Note 3, Samsung Galaxy S4, Huawei P9 lite, among
others) over different days. This captures the time-variant
nature of the WiFi fingerprints as well as the heterogeneity
of users and devices.
We implemented a WiFi collector App using the Android
SDK to scan APs. The program records the (MAC address,
RSS, timestamp) for each heard WiFi AP. The scanning rate
was set to one per second. We implemented our deep learning
based training using the Google TensorFlow [32] framework
on the Google Collaboratory Cloud2. 40% of the data points
are held out for testing. We experimented with different deep
learning architectures and the one of 200×300×400×500
obtains the best performance.
B. Effect of Changing WiDeep Parameters
In this section, we study the effect of the different parame-
ters on the system performance including different techniques
used to add noise to the input, dropout regularization, different
values of radial basis function parameters used for probabilistic
online localization, number of input scans used in estimation,
and number of deep learning model layers. Table II shows
the default parameters values used throughout the evaluation
section.
1) Effect of different input noise corruption techniques:
We experimented with two different input noising methods
(Section III-B1): the masking method and additive Gaus-
sian noise method. Fig. 8 shows the mean location error
2https://colab.research.google.com

Fig. 6: University floorplan.
Fig. 7: Appartment floorplan.
of WiDeep when trained using different masking corruption
fraction values. Similarly, Fig. 9 shows the mean error of using
different standard deviation values to train the models using
the additive Gaussian method. The figures show that adding
more variations, by injecting artificial noise, of the data input
to the model enhances performance compared to the case when
when the noise level is zero. However, adding too much noise
distort the signal and increase the ambiguity between adjacent
locations, hence increases the localization error. An optimal
noise level for both techniques can be achieved at masking
probability f=0.1and s=0.1.
2) Effect of dropout regularization: Fig. 10 shows the effect
of the dropout regularization on the mean localization accu-
racy. The figure shows that as the dropout rate is increased, the
mean localization accuracy is improved. This is aligned with
our expectation because the regularization reduces the over-
fitting of the neural network to the input data. However, using
a large dropout rate leads to a decrease in the localization
accuracy as the network does not fit the data well (i.e., under-
fitting). Therefore, this trade-off is balanced at 0.5 dropout
rate.
3) Effect of the number of scans per estimate: Fig. 11
shows the effect of increasing the number of the scans (n) used
to estimate a location in the online phase. The figure shows
that as nincreases, the accuracy improves until it reaches an
optimal value at n=7beyond which it begins to deteriorate.
This is due to two opposing factors: (1) By increasing n,we
get more information which is useful for location estimate,
(2) However, as nincreases, more time is spent to collect
TABLE II: Default parameters.
Parameter Range Default value
Testbed University,
Apartment University
Masking fraction (f) 0 - 0.5 0.1
Additive Gaussian Stdev (s) 0 - 0.5 0.1
Learning rate 0.001 - 0.1 0.1
Batch size 16 - 2048 128
Dropout rate (r) 0 - 0.9 0.5
Radial Basis Function pa-
rameter (λ)0.25 - 6 4
Number of scans per esti-
mate (n)1-30 7
Number of epochs 1 - 20000 10000
Network architecture 200×300×400×500
these samples which may involve crossing reference points
boundaries. This has a negative effect on performance.
4) Effect of the λparameter in the Radial Basis Function:
Fig. 12 shows the effect of using different values of λin the
radial basis function on the localization error. The figure shows
that there is an optimal value for λat 4.
5) Effect of number of layers in the network: Fig. 13
shows the effect of changing the number of layers (stacked
autoencoders). The figure shows that increasing the layer,
increases the accuracy until reaching an optimal value at
four layers. After that, the accuracy starts to decrease as the
network begins to overfit the training data.
C. Robustness Experiments
In this section, we assess the robustness of WiDeep under
different challenging scenarios including reducing the density
of access points and reduced number of fingerprint locations.
1) Density of access points: Fig. 14 shows the effect of
reducing the density of the access points on accuracy. For
this, we uniformly removed access points from the total access
points detected in the area. The figure shows that even with
a density as low as 5% of the access points, WiDeep can
achieve high accuracy of less than 2.8m mean error. This is
due to the different noise-handling techniques as well as the
used regularization techniques. This highlights the robustness
of WiDeep.
2) Density of training points: Fig. 15 shows the per-
formance WiDeep when the number of training fingerprint
locations is reduced. The figure shows that, even though
reducing the number of training points/percentage linearly
leads to increasing the area associated with each training point
quadratically, the decrease in the accuracy does not grow as
fast. WiDeep can achieve a mean accuracy of 3.16m even with
50% reduction in training points.
D. Comparative Evaluation
In this section, we compare the location accuracy, robustness
to heterogeneous devices, and runtime of the WiDeep system
against two baseline systems. The first is a popular probabilis-
tic fingerprinting based indoor localization technique (Horus

!t
2.4
2.7
3
3.3
3.6
3.9
0 0.1 0.3 0.5
Mean location accuracy (m)
Corruption fraction (f)
Fig. 8: Effect of the masking probability
on accuracy.
3
3.1
3.2
3.3
3.4
3.5
0 0.04 0.08 0.12 0.16 0.2
Mean location accuracy (m)
Noise standard deviation (s)
Fig. 9: Effect of the noise standard devi-
ation on accuracy.
2.6
2.7
2.8
2.9
3
3.1
0 0.2 0.4 0.6 0.8
Mean location accuracy (m)
Dropout rate (r)
Fig. 10: Effect of the dropout regulariza-
tion rate on accuracy.
2.6
2.65
2.7
2.75
2.8
2.85
2.9
3 4 5 6 7 8 9 10
Mean location accuracy (m)
Number of scans per estimate (n)
Fig. 11: Effect of the number of scans
per estimate on the accuracy.
2.5
2.6
2.7
2.8
2.9
3
1 2 3 4 5 6
Mean location accuracy (m)
Lambda
Fig. 12: Localization error vs. RBF
parameterλ.
2.6
2.7
2.8
2.9
3
3.1
3 4 5 6
Mean location accuracy (m)
Number of layers
Fig. 13: Effect of the number of layers
on accuracy.
[6]) that assumes the independence of the APs and the second
is a recent deep-learning based indoor localization technique
(DeepFi [8]) that does not perform noise handling or model
over-fitting avoidance.
1) Localization accuracy: Fig. 16 and Fig. 17 show the
CDF of distance error for all systems in the university and
apartment testbeds. Table III and IV summarize the results.
The figures illustrate that our WiDeep system can achieve
significantly better mean localization accuracy than the other
systems by at least 29.8% and up to 169% in university and
apartment testbeds, respectively. Moreover, WiDeep enhances
all the other quantiles. This can be explained by noting
that traditional probabilistic fingerprinting techniques such as
Horus [6] cannot capture the correlation between the different
APs and the fingerprinting locations. Similarly, traditional
deep learning techniques such as DeepFi [8] do not take
the inherent noise of the wireless signals into consideration
nor avoid over-training. Therefore, their performance drops
noticeably when trained with noisy data. This can be seen in
figures 16 and 17, where the accuracy of DeepFi degrades in
such scenarios while WiDeep maintains its accuracy.
Note also that WiDeep performance is consistent in the
two testbeds, contrary to the other two techniques: DeepFi
performs better in the testbed with more available data (i.e
higher density of APs and training locations) while Horus can
tolerate better the lower APs density and lower in the other
testbed.
2) Device heterogeneity: Here, we evaluate the differ-
ent techniques robustness to devices heterogeneity. Initially,
WiDeep is trained and tested with the same device (i.e.
Samsung Galaxy Note 3). We then carry out experiments by
training the different systems with the Samsung Galaxy Note
3 tablet and testing with a Samsung S4 mini smartphone. The
two devices have completely different form factors and WiFi
chips.
Fig. 18 shows that WiDeep provides approximately the same
accuracy when testing with different device as when testing
with the same device. It can also be seen from the figure
that WiDeep has the best performance in handling device
heterogeneity compared to the other two systems across all
percentiles. This is due to the combination of additive noise
in the training data and the adoption of denoising autoen-
coders which gives WiDeep greater flexibility than the other
systems. In particular, this is true since device heterogeneity
can be considered to be a form of noise, which the WiDeep
network and training process are designed specifically to
combat. Horus also shows better adaptability than DeepFi.
This can be attributed to the fact that it utilizes probabilistic
techniques, which are known to perform well in the presence
of uncertainty or noise. On the other hand, DeepFi shows
poor performance to noisy data because of the lack of specific
provisions to handle such phenomena in its design.
3) Time per location estimate: Fig. 19 compares the run-
ning time per location estimate for the three techniques. The

2.6
2.65
2.7
2.75
2.8
5% 10% 20% 30% 40%
Mean location accuracy (m)
Percentage of used APs
Fig. 14: Effect of density of APs on
accuracy.
2.6
2.7
2.8
2.9
3
3.1
3.2
0% 15% 25% 30% 50%
Mean location accuracy (m)
Reduction percentage of trining points
Fig. 15: Effect of reducing the number
of training locations on accuracy.
0
0.2
0.4
0.6
0.8
1
0 5 10 15 20 25 30
CDF
Location error (m)
WiDeep
Horus [6]
DeepFi [8]
DeepFi with noisy data[8]
Fig. 16: Comparison of CDFs in the
University testbed.
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12
CDF
Location error (m)
WiDeep
DeepFi [8]
Horus [6]
DeepFi with noisy data [8]
Fig. 17: Comparison of CDFs in the
Apartment testbed.
0
5
10
15
20
25
30
35
40
Location accuracy (m)
DeepFi [8]
DeepFi with noisy data [8]
Horus [6]
WiDeep
WiDeep tested with same device
Fig. 18: Comparison based on device
heterogeneity.
0
100
200
300
400
500
WiDeep DeepFi[8] Horus[6]
Time (ms)
Fig. 19: Time per location estimate.
machine used for running the algorithms is an HP Omen
laptop with an i7 2.6 GHz processor, 16 GB RAM, and a
Nividia GTX965 GPU. The figure shows that the running time
of WiDeep and DeepFi is comparable. Horus has the lowest
running time per location estimate as its prediction is based
on only Bayesian inference. Deep learning techniques, on the
other hand, need to pass the data through all the layers of
the network. Nonetheless, all techniques can estimate the user
location in less than 412ms, which allows realtime tracking
of the user. This can be further enhanced if needed through
parallelization.
E. Discussion
WiDeep is designed to operate with heterogenous devices
without compromising on localization accuracy. The use of
a deep model alone (e.g as in DeepFi) cannot lead to this
design goal. WiDeep is able to achieve this as a combination
of the particular choice of deep network used and the associ-
ated design considerations. Specifically, the stacked denoising
autoencoder network used in WiDeep is, by definition, capable
of reconstructing the underlying input in the presence of noise
or distortion. Therefore for the best results, the training process
of this network necessitates the use of noisy data so that the
network truly learns to extract the underlying information from
the data as obtained from users’ heterogeneous devices [25].
This enhances the generalization ability of the network in
challenging scenarios, e.g. in the presence of device hetero-
geneity. At the same time, the use of noisy data alone with
DeepFi (which does not use this type of network model)
leads to a degradation of the obtained localization accuracy.
Additionally, the dropout regularization of WiDeep ensures
the quality of the final model by eliminating co-dependencies
between the constituent neurons [27], [33].
It can be seen from Fig. 9 that training WiDeep without
considering the noise injection process leads to a significant
drop of the localization performance to 3.47m. Similarly, Fig.
10 shows that the accuracy degrades to 2.69m without pe-
nalizing the training process with such dropout regularization.
Therefore, the combination of the network used and the regu-
larization techniques are able to yield significant improvements
over traditional deep learning models.
V. R ELATED WORK
In this section, we discuss the most relevant literature to
our WiDeep system. In particular, we cover two categories:
fingerprinting systems and crowd-sourcing systems.
A. Fingerprinting Systems
Fingerprinting systems present the most popular WiFi-based
indoor localization technique due to their high accuracy. Those
can be categorized into traditional and deep learning-based
fingerprinting systems.

1) Traditional Fingerprinting Systems:Radar [12] finger-
print captures the average RSS of the heard APs at the differ-
ent fingerprint locations. During the online phase, matching
is based on using the k-nearest neighbors algorithm [34].
Deterministic approaches though cannot deal well with he
noise and variations of the RF signal. To tackle the noisy
nature of RSS, probabilistic techniques have been proposed,
e.g. Horus [6], [35], [36]. In this case, the fingerprint reflects
the RSS histogram for each AP at each reference location,
assuming APs are independent. The most probable location
is estimated based on Bayesian inference. Many variants of
probabilistic techniques have been proposed over the years
to further enhance the localization performance [5], [7], [37].
For instance, [7] uses radial basis networks to predict the
unknown location. Despite probabilistic techniques being able
to handle the inherently noisy wireless signals in a better way
than deterministic techniques, they usually assume that the
signals from different APs are independent to avoid the curse
of dimensionality problem [19]. This leads to coarse-grained
accuracy.
WiDeep, on the contrary, harnesses a deep neural network
that is able to learn dependencies between signals from dif-
ferent APs. In addition, it is designed to address the inherent
noise in the RF signals. Moreover, it has provisions to handle
over-fitting, leading to better robustness.
2) Deep Learning Systems:Recently, different deep learn-
ing techniques have been proposed in order to train models to
provide a localization service. In DeepFi [8], [9], Restricted
Boltzman Machines are used to pre-train a deep learning
system. The localization service of DeepFi depends on the
magnitudes of the channel state information (CSI) data, as
compared to the standard received signal strength. Later,
deep convolution networks based on CSI data also have been
proposed to estimate the unknown locations [38], [39]. All
these techniques use CSI data, which needs special hardware
for collection, reducing the system ubiquity. In addition, they
do not have provisions to reduce over-fitting or handle the
inherent noise in the input data, reducing their robustness.
In contrast, the operation of WiDeep depends on standard
RSS readings, which can be received by the common on-board
WiFi radio present in all mobile devices using standard APIs
in the operating system. In addition,WiDeep is designed to
deal with noisy data and have provisions to avoid model over-
fitting, both leading to higher accuracy and more robustness.
B. Crowdsourcing Systems
To reduce the fingerprint construction overhead, a num-
ber of systems have been introduced in which the users
collaborate to improve the localization system by crowd-
sourcing the fingerprint. [40] uses crowd-sourcing to improve
the particle filter performance overtime and hence improve the
localization accuracy of the system. Other systems, e.g. [15],
[16], [20], [41]–[43], use the smartphone inertial sensors to
calculate the user location using dead-reckoning and leverage
different sensor-based landmarks, including WiFi, to reset the
accumulated error. These system, however require additional
TABLE III: Accuracy percentiles of different systems in the
University floorplan
Technique Average 50th
Percentile
75th
Percentile
100th
Percentile
WiDeep 2.64m 2.38m 3.38m 7.12m
Horus [6] 4.04m
(-53.03%)
2.25m
(5.46%)
4.03m
(-19.23%)
17.50m
(-145.78%)
DeepFi [8] 7.10m
(-168.93%)
6.09m
(-155.88%)
9.54m
(-182.24%)
24.14m
(-239.04%)
TABLE IV: Accuracy percentiles of different systems in the
apartment floorplan
Technique Average 50th
Percentile
75th
Percentile
100th
Percentile
WiDeep 1.21m 1.07m 1.62m 3.74m
DeepFi [8] 1.57m
(-29.75%)
1.25m
(-16.82%)
2.04m
(-25.92%)
4.97
(-32.88%)
Horus [6] 2.57m
(-112.39%)
2.17m
(-102.08%)
3.35m
(-106.79%)
8.73
(-133.42%)
sensors, which may not be available on all mobile devices,
especially in development countries where low-end phones are
more common. WiDeep can benefit from crowd-sourcing to
construct its fingerprint in an automatic manner. In addition,
based on deep learning and its different noise and robustness
handling modules it can provide robust and high accuracy
localization without the need of any additional sensors.
VI. CONCLUSION
We presented WiDeep, an accurate and robust WiFi fin-
gerprinting indoor localization technique based on a deep
neural network. The system leverages stacked denoising auto-
encoders in a probabilistic framework to mitigate the noise
in the RSS measurements. Additionally, it employs model
regularization to enable the network to generalize and avoid
over-fitting, leading to a more robust and stable models.
We evaluated WiDeep in two different challenging envi-
ronments that represent a university building and a domestic
apartment using different Android devices. The results show
the WiDeep comes with a localization accuracy better than
the state-of-the-art systems by at least 53% and 29.8% in the
large and small environments respectively. Moreover, its per-
formance is robust to different devices and different densities
of APs in different environments.
REFERENCES
[1] P. Davidson and R. Pich´
e, “A survey of selected indoor positioning
methods for smartphones,” IEEE Communications Surveys & Tutorials,
vol. 19, no. 2, pp. 1347–1370, 2017.
[2] D. Lymberopoulos and J. Liu, “The microsoft indoor localization
competition: Experiences and lessons learned,” IEEE Signal Processing
Magazine, vol. 34, no. 5, pp. 125–140, 2017.
[3] J.-P. Sheu, P.-C. Chen, and C.-S. Hsu, “A distributed localization scheme
for wireless sensor networks with improved grid-scan and vector-based
refinement,” IEEE transactions on mobile computing, vol. 7, no. 9, pp.
1110–1123, 2008.

[4] N. Lasla, M. F. Younis, A. Ouadjaout, and N. Badache, “An effective
area-based localization algorithm for wireless networks,” IEEE Trans-
actions on Computers, vol. 64, no. 8, pp. 2103–2118, 2015.
[5] M. Youssef and A. Agrawala, “Handling samples correlation in the Ho-
rus system,” in INFOCOM 2004. Twenty-third AnnualJoint Conference
of the IEEE Computer and Communications Societies, vol. 2. IEEE,
2004, pp. 1023–1031.
[6] ——, “The Horus WLAN location determination system,” in Proceed-
ings of the 3rd international conference on Mobile systems, applications,
and services. ACM, 2005, pp. 205–218.
[7] C. Laoudias, P. Kemppi, and C. G. Panayiotou, “Localization using
radial basis function networks and signal strength fingerprints in wlan,”
in Global telecommunications conference, 2009. GLOBECOM 2009.
IEEE. IEEE, 2009, pp. 1–6.
[8] X. Wang, L. Gao, S. Mao, and S. Pandey, “DeepFi: Deep learning
for indoor fingerprinting using channel state information,” in Wireless
Communications and Networking Conference (WCNC), 2015 IEEE.
IEEE, 2015, pp. 1666–1671.
[9] ——, “CSI-based fingerprinting for indoor localization: A deep learning
approach,” IEEE Transactions on Vehicular Technology, vol. 66, no. 1,
pp. 763–776, 2017.
[10] W. Zhang, K. Liu, W. Zhang, Y. Zhang, and J. Gu, “Deep neural
networks for wireless localization in indoor and outdoor environments,”
Neurocomputing, vol. 194, pp. 279–287, 2016.
[11] R. Elbakly and M. Youssef, “A calibration-free rf localization system,”
in Proceedings of the 23rd SIGSPATIAL International Conference on
Advances in Geographic Information Systems. ACM, 2015, p. 63.
[12] P. Bahl and V. N. Padmanabhan, “RADAR: An in-building RF-based
user location and tracking system,” in INFOCOM 2000. Nineteenth
Annual Joint Conference of the IEEE Computer and Communications
Societies. Proceedings. IEEE, vol. 2. Ieee, 2000, pp. 775–784.
[13] H. Abdel-Nasser, R. Samir, I. Sabek, and M. Youssef, “MonoPHY:
Mono-stream-based device-free WLAN localization via physical layer
information,” in Wireless communications and networking conference
(WCNC), 2013 IEEE. IEEE, 2013, pp. 4546–4551.
[14] K. Habak, K. A. Harras, and M. Youssef, “Bandwidth aggregation
techniques in heterogeneous multi-homed devices: A survey,” Computer
Networks, vol. 92, pp. 168–188, 2015.
[15] H. Wang, S. Sen, A. Elgohary, M. Farid, M. Youssef, and R. R.
Choudhury, “No need to war-drive: Unsupervised indoor localization,”
in Proceedings of the 10th international conference on Mobile systems,
applications, and services. ACM, 2012, pp. 197–210.
[16] H. Abdelnasser, R. Mohamed, A. Elgohary, M. F. Alzantot, H. Wang,
S. Sen, R. R. Choudhury, and M. Youssef, “SemanticSLAM: Using
environment landmarks for unsupervised indoor localization,” IEEE
Transactions on Mobile Computing, vol. 15, no. 7, pp. 1770–1782, 2016.
[17] M. Ibrahim, M. Torki, and M. ElNainay, “Cnn based indoor localization
using rss time-series,” in 2018 IEEE Symposium on Computers and
Communications (ISCC). IEEE, 2018, pp. 01044–01 049.
[18] M. Youssef, M. Abdallah, and A. Agrawala, “Multivariate analysis for
probabilistic WLAN location determination systems,” in The Second
Annual International Conference on Mobile and Ubiquitous Systems:
Networking and Services. IEEE, 2005, pp. 353–362.
[19] N. M. Nasrabadi, “Pattern recognition and machine learning,” Journal
of electronic imaging, vol. 16, no. 4, p. 049901, 2007.
[20] H. Aly and M. Youssef, “Dejavu: an accurate energy-efficient outdoor
localization system,” in Proceedings of the 21st ACM SIGSPATIAL Inter-
national Conference on Advances in Geographic Information Systems.
ACM, 2013, pp. 154–163.
[21] M. Elhamshary and M. Youssef, “CheckInside: a fine-grained indoor
location-based social network,” in Proceedings of the 2014 ACM In-
ternational Joint Conference on Pervasive and Ubiquitous Computing.
ACM, 2014, pp. 607–618.
[22] M. Elhamshary, M. Youssef, A. Uchiyama, H. Yamaguchi, and T. Hi-
gashino, “Transitlabel: A crowd-sensing system for automatic labeling
of transit stations semantics,” in Proceedings of the 14th Annual In-
ternational Conference on Mobile Systems, Applications, and Services.
ACM, 2016, pp. 193–206.
[23] S. Sen, B. Radunovic, R. R. Choudhury, and T. Minka, “You are
facing the mona lisa: spot localization using phy layer information,”
in Proceedings of the 10th international conference on Mobile systems,
applications, and services. ACM, 2012, pp. 183–196.
[24] Y. Kang, K.-T. Lee, J. Eun, S. E. Park, and S. Choi, “Stacked denoising
autoencoders for face pose normalization,” in International Conference
on Neural Information Processing. Springer, 2013, pp. 241–248.
[25] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.-A. Manzagol,
“Stacked denoising autoencoders: Learning useful representations in a
deep network with a local denoising criterion,” Journal of Machine
Learning Research, vol. 11, no. Dec, pp. 3371–3408, 2010.
[26] M. Youssef and A. K. Agrawala, “Small-scale compensation for WLAN
location determination systems.” in WCNC, 2003, pp. 1974–1978.
[27] H. Rizk, M. Torki, and M. Youssef, “CellinDeep: Robust and Accurate
Cellular-based Indoor Localization via Deep Learning,” IEEE Sensors
Journal, 2018.
[28] Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle, “Greedy layer-
wise training of deep networks,” in Advances in neural information
processing systems, 2007, pp. 153–160.
[29] D. Erhan, Y. Bengio, A. Courville, P.-A. Manzagol, P. Vincent, and
S. Bengio, “Why does unsupervised pre-training help deep learning?”
Journal of Machine Learning Research, vol. 11, no. Feb, pp. 625–660,
2010.
[30] G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of
data with neural networks,” science, vol. 313, no. 5786, pp. 504–507,
2006.
[31] M. Youssef and A. Agrawala, “Continuous space estimation for WLAN
location determination systems,” in Computer Communications and Net-
works, 2004. ICCCN 2004. Proceedings. 13th International Conference
on. IEEE, 2004, pp. 161–166.
[32] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S.
Corrado, A. Davis, J. Dean, M. Devin et al., “Tensorflow: Large-scale
machine learning on heterogeneous distributed systems,” arXiv preprint
arXiv:1603.04467, 2016.
[33] A. Shokry, M. Torki, and M. Youssef, “DeepLoc: a ubiquitous accurate
and low-overhead outdoor cellular localization system,” in Proceedings
of the 26th ACM SIGSPATIAL International Conference on Advances in
Geographic Information Systems. ACM, 2018, pp. 339–348.
[34] H. Rizk, S. Elgokhy, and A. Sarhan, “A hybrid outlier detection
algorithm based on partitioning clustering and density measures,” in
Computer Engineering & Systems (ICCES), 2015 Tenth International
Conference on. IEEE, 2015, pp. 175–181.
[35] M. A. Youssef, A. Agrawala, and A. U. Shankar, “Wlan location
determination via clustering and probability distributions,” in Pervasive
Computing and Communications, 2003.(PerCom 2003). Proceedings of
the First IEEE International Conference on. IEEE, 2003, pp. 143–150.
[36] K. El-Kafrawy, M. Youssef, and A. El-Keyi, “Impact of the human
motion on the variance of the received signal strength of wireless links,”
in Personal Indoor and Mobile Radio Communications (PIMRC), 2011
IEEE 22nd International Symposium on. IEEE, 2011, pp. 1208–1212.
[37] M. Youssef and A. Agrawala, “Location-clustering techniques for wlan
location determination systems,” International Journal of Computers and
Applications, vol. 28, no. 3, pp. 278–284, 2006.
[38] H. Chen, Y. Zhang, W. Li, X. Tao, and P. Zhang, “ConFi: Convolutional
Neural Networks Based Indoor Wi-Fi Localization Using Channel State
Information,” IEEE Access, vol. 5, pp. 18066–18 074, 2017.
[39] X. Wang, X. Wang, and S. Mao, “CiFi: Deep convolutional neural
networks for indoor localization with 5 GHz Wi-Fi,” in Communications
(ICC), 2017 IEEE International Conference on. IEEE, 2017, pp. 1–6.
[40] Z. Liu, L. Zhang, Q. Liu, Y. Yin, L. Cheng, and R. Zimmer-
mann, “Fusion of magnetic and visual sensors for indoor localization:
Infrastructure-free and more effective,” IEEE Transactions on Multime-
dia, vol. 19, no. 4, pp. 874–888, 2017.
[41] H. Aly, A. Basalamah, and M. Youssef, “Map++: A crowd-sensing
system for automatic map semantics identification,” in 2014 Eleventh
Annual IEEE International Conference on Sensing, Communication, and
Networking (SECON). IEEE, 2014, pp. 546–554.
[42] N. Mohssen, R. Momtaz, H. Aly, and M. Youssef, “It’s the human
that matters: accurate user orientation estimation for mobile computing
applications,” in Proceedings of the 11th International Conference on
Mobile and Ubiquitous Systems: Computing, Networking and Services.
ICST (Institute for Computer Sciences, Social-Informatics and , 2014,
pp. 70–79.
[43] H. Aly, A. Basalamah, and M. Youssef, “Lanequest: An accurate and
energy-efficient lane detection system,” in Pervasive Computing and
Communications (PerCom), 2015 IEEE International Conference on.
IEEE, 2015, pp. 163–171.