See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/330410380

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

Deep Learning

Conference Paper · January 2019

CITATIONS READS

6 1,074

5 authors, including:

Moustafa Abbas Moustafa Elhamshary

Alexandria University Tanta University
1 PUBLICATION   6 CITATIONS    17 PUBLICATIONS   177 CITATIONS   

SEE PROFILE SEE PROFILE

Hamada Rizk Marwan Torki

Tanta University Alexandria University
6 PUBLICATIONS   24 CITATIONS    44 PUBLICATIONS   614 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

RGBD Pose Recognition View project

Cognitive Radio Cloud View project

All content following this page was uploaded by Marwan Torki on 16 January 2019.

The user has requested enhancement of the downloaded file.

 WiDeep: WiFi-based Accurate and Robust Indoor
Localization System using Deep Learning
Moustafa Abbas Moustafa Elhamshary Hamada Rizk
Dept. of Comp. and Sys. Eng., Dept. of Comp. and Cont. Eng., Dept. of Comp. Sci. and Eng.,
Alexandria University Tanta University EJUST, Alexandria, Egypt
Alexandria, Egypt Tanta, Egypt & Tanta University, Tanta, Egypt
[email protected] [email protected] [email protected]

Marwan Torki Moustafa Youssef

Dept. of Comp. and Sys. Eng., Dept. of Comp. and Sys. Eng.,
Alexandria University Alexandria University
Alexandria, Egypt Alexandria, Egypt
[email protected] [email protected]

Abstract—Robust and accurate indoor localization has been the techniques do not work well with heterogeneous phones
goal of several research efforts over the past decade. Due to the and their accuracy is usually less than fingerprinting-based
ubiquitous availability of WiFi indoors, many indoor localization techniques.
systems have been proposed relying on WiFi fingerprinting.
However, due to the inherent noise and instability of the wireless On the other hand, fingerprinting techniques leverage the
signals, the localization accuracy usually degrades and is not recorded WiFi APs signatures (i.e. fingerprints) to estimate
robust to dynamic changes in the environment. the device location. Typical fingerprint-based WiFi localiza-
We present WiDeep, a deep learning-based indoor localization tion techniques work in two phases: The first one is the
system that achieves a fine-grained and robust accuracy in offline phase (i.e., calibration) during which the received
the presence of noise. Specifically, WiDeep combines a stacked
denoising autoencoders deep learning model and a probabilistic signal strength (RSS) readings from the multiple access points
framework to handle the noise in the received WiFi signal and (APs) installed in the area of interest are recorded at known
capture the complex relationship between the WiFi APs signals locations. Then, in the tracking phase, RSS measurements
heard by the mobile phone and its location. WiDeep also intro- from the detected APs at an unknown location are matched
duces a number of modules to address practical challenges such against the stored fingerprints to estimate the best location
as avoiding over-training and handling heterogeneous devices.
We evaluate WiDeep in two testbeds of different sizes and match either deterministically, e.g. [12], or probabilistically,
densities of access points. The results show that it can achieve e.g. [6]. Fingerprinting-based techniques are widely adopted
a mean localization accuracy of 2.64m and 1.21m for the larger due to their relatively good accuracy. Practically however,
and the smaller testbeds, respectively. This accuracy outperforms the deployment of such techniques faces major challenges
the state-of-the-art techniques in all test scenarios and is robust due to the inherent noise in the wireless signals that affects
to heterogeneous devices.
Index Terms—WiFi, Deep learning, indoor, localization, finger-
localization accuracy [13], [14]. Therefore, many systems have
printing been proposed to address these challenges over the years, e.g.
[6], [15]–[17]. Probabilistic techniques such as [6], [18] can
I. I NTRODUCTION counter the inherent wireless signal noise in a better way than
deterministic techniques [12]. However, they usually assume
As people spend most of their time indoors, academia and
that the signals from different access points are independent
industry have recognized the value of the indoor localization
to avoid the curse of dimensionality problem [19]. This leads
problem and have devoted much effort and resources into
to coarse-grained accuracy. Hybrid techniques, e.g. [15], [16],
solving it [1], [2]. Due to the wide-spread coverage of WiFi
[20]–[22], leverage the sensors that are available on high-
and the support of the IEEE 802.11 standard by the majority of
end smartphones to combat the wireless channel noise. Other
mobile devices, most proposed indoor localization systems are
techniques, e.g. [8], [23], leverage the detailed channel state
WiFi-based including propagation- and fingerprinting-based
information obtained from specialized WiFi chips to combat
techniques [3]–[10]. Propagation-based techniques, e.g. [3],
the noise. However, both of the last two categories are not
[4], [11], aim to model the relation between the received
supported by the vast majority of mobile devices, limiting their
signal and distance without site surveying. Despite their ease
ubiquitous deployment.
of deployment without the need for prior calibration, these
In this paper, we propose WiDeep: a WiFi-based indoor
This work has been supported in part by a grant from the Egyptian National fingerprinting localization system that can achieve robust and
Telecommunication Regulatory Authority(NTRA). 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 Offline Online
assuming access points’ independence as in current probabilis-
tic techniques. However, leveraging a deep network alone as in Signature Collector
[8] may not lead to the required performance in the presence RSS Collector
of device heterogeneity which can be considered as a form of Locc 1 Locc 2 Locc N
noise. To ensure the robustness and the generalization ability Preprocessor
of the system in this challenging scenario, we adopt a deep
network model utilizing stacked denoising autoencoders to
Probabilistic Localizer
robustly extract a good representation of the relation between Preprocessor
the noisy WiFi scans and the different fingerprint locations. Deep
p Model Estimator
Furthermore, WiDeep also employs a regularization technique Noise Injector
Probabilistic Estimator
to avoid model over-fitting and boost the robustness of the
system. Model Trainer
During tracking, the output of the deep learning models is
fused using a probabilistic framework to further handle the Model 1 Model 2 Model N

noise in the input signal.

We implemented and deployed WiDeep on different Android
phones and evaluated its performance in two different testbeds:
a 629m2 university building and a 65m2 residential apartment.
The two buildings have different layouts and WiFi APs den- Fig. 1: WiDeep system architecture.
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 measurements by injecting artificial noise to the collected WiFi
accuracy than traditional fingerprinting techniques and other scans. This helps not only in simulating the distorted wireless
basic deep learning techniques by at least 29.8% and up to channel but also in reducing the model over-fitting by forcing
168%. This accuracy is maintained under significant decrease the model to learn the inherent feature of the data. For each
in the access points density as well as under heterogeneous fingerprint point, the original and corrupted noisy data are
devices, highlighting WiDeep promise as a robust and accurate forwarded as input to the Model Trainer module. This module
indoor localization technique. is responsible for creating and training a stacked denoising
The rest of the paper is organized as follows: Section II auto-encoders deep model corresponding to each fingerprint
presents an overview on how WiDeep works and its mathe- point. Moreover, it enhances the robustness of the model by
matical model. Section III presents the details of the WiDeep avoiding over-training. Finally, all the trained models for the
system. We evaluate the system performance in Section IV. different fingerprint locations are stored to be used during the
Finally, sections V and VI discuss related work and conclude online localization phase.
the paper respectively. During the online localization phase, the user is tracked
in realtime. The process starts by scanning for the APs and
II. S YSTEM OVERVIEW AND M ATHEMATICAL M ODEL
their RSSs at the unknown user location. This data is first
A. System Overview preprocessed (shaped and normalized) to fit with deep learn-
Fig. 1 shows the system architecture. The system runs in two ing model input. Then, the Probabilistic Localizer module
phases: offline training phase and online localization phase. leverages the different deep models output in a probabilistic
During the offline phase, the system builds N deep neural framework to estimate the most probable user location using
networks corresponding to N fingerprint training points (i.e., the Deep Model Estimator and Probabilistic Estimator sub-
a deep neural network for each reference point). This helps in modules respectively.
scaling the system to large areas as well as keeping the model
size small and easier to train. B. Mathematical Model
To collect the training data, the Signature Collector module Without a loss of generality, we assume a 2D physical
is used to scan for the APs and their associated signal strengths area of interest L containing M access points. N discrete
at the different fingerprint locations in the area of interest. fingerprint locations are spread over the entire area, where
These measurements are opportunistically transfered to the training data is collected. During the online localization phase,
WiDeep server in the cloud. The Preprocessor module is a user holding a mobile device at an unknown location l ∈ L
used to transform the WiFi measurements to fit the format scans for the nearby APs. Let a vector xi of M dimensions
required in the deep network training model as well as perform represents a single WiFi scan. Each entry i in this vector
the required normalization. The preprocessed data is then is the received signal strength reading from access point i.
fed to the Noise Injector module to corrupt the original The problem then becomes: given a signal strength vector
x = (x1 , ..., xM ), we seek to find the fingerprint location li x x̃ x̂
that maximizes the probability P (li |x). In the next section,
0 h
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. Corruption W WT
III. T HE WiDeep S YSTEM
0
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.
Fig. 2: Denoising Autoencoder: A sample x is stochastically
A. The Preprocessor corrupted to produce x̃. Next, the the hidden layer (h) maps x̃
to the output x̂ to produce a reconstructed version of x. Error
This module is responsible for mapping the recorded WiFi is measured by the L(x, x̂) loss function that compares the
RSS readings, xi , to the corresponding feature vectors. Note reconstructed output with the original noise-free input.
that since not all M APs 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 1) Noise Injector: This module aims to enhance the model
the machine learning model. After that, the input RSS values ability to handle the noisy input data. This is achieved by
ranges are normalized to be in the range between [0,1] for generating corrupted variations of the collected WiFi scans.
each AP. Features normalization is known to speed up model This also have the added advantage of reducing over-fitting
training and increase the model robustness [24]. and coping better with devices heterogeneity as we quantify
in Section IV. To do that, the module adds stochastic noise
B. Offline Models Construction Phase to the input data through two different techniques: Masking
corruption and Additive Gaussian corruption.
During this phase, the system builds N deep models
corresponding to N fingerprint training points (i.e., a deep Masking corruption method: The intuition behind this
model for each point). The system also addresses a number of method is that the number of access points detected at a
challenges including handling the noise and fluctuation of APs fixed location varies with time due to multipath and fading
signals as well as reducing the model over-fitting to training effects [26]. The masking method leverages this fact to emulate
data, allowing for better generalization and robustness. fluctuating access points [27]. The idea is to generate a random
We choose stacked denoising autoencoders as our model binary vector with specific probability of its elements to be
as they are able to extract the latent features from noisy zeros determined by the corruption fraction parameter (f ).
data. Autoencoders are unsupervised learning models, where This generated binary vector is then multiplied by the original
their goal is to learn a concise mapping that can regenerate input to get a noisy input signal, where the entries of the APs
the input to the autoencoder [25]. Denoising autoencoders corresponding to the zero random bits are dropped, emulating
extends traditional autoencoders to handle noisy data in a not hearing them (Fig. 3a).
better way. Specifically, instead of feeding the original input Additive Gaussian corruption method: Due to the noise in
to the denoising autoencoder, we feed it a noisy version. This the wireless channels and the diversity of the WiFi chips in
allows the hidden layer of the autoencoder to learn important different devices, the magnitude of the RSS may be shifted
features (Fig. 2). Specifically, the denoising autoencoder is with some variance. Therefore, to emulate this behavior, this
trained by first corrupting the input RSS vector x to obtain technique adds white Gaussian noise with a specific standard
vector x̃. The goal is to learn the parameters of the hidden deviation s to the different entries of the RSS vector (Fig. 3b).
layer h so that the output (x̂) of the autoencoder matches the The synthesized vector is finally re-normalized so that all
original uncorrupted vector x. By using a noisy version of entries are between 0 and 1.
the input, the autoencoder is forced to learn the latent features 2) The Model Trainer: Fig. 5 shows the WiDeep deep
of the input data. Note that the weights between the hidden model. It consists of a number of stacked denoising au-
and output layers are the transpose of the weight between the toencoders, one stack for each fingerprint location. The used
input and hidden layer, reflecting the decoding process of the activation function is the Sigmoid function formulated as [25]:
autoencoder. Training is performed using the gradient descent 1
algorithm, where the least square error between the original sigmoid(x) = (1)
1 + e−x
input data x and the reconstructed data x̂ is used as the loss
function to adjust the weights. The Sigmoid function ensures that its output will range be-
In the balance of this section, we first describe how to tween 0 and 1.
introduce noise to our input data. Then we provide the details To train WiDeep model end-to-end, we use two training
of our deep model. stages: (1) greedy layer-wise pre-training stage and (2) a fine-
 Input RSS vector Input RSS vector D. Online Localization Phase
0.2 0.0 0.5 0.3 0.2 0.0 0.5 0.3 During this phase, we harness the learned models to esti-
mate the unknown locations of the test inputs. Specifically, we
Mask Noise vector input the current WiFi scan at the unknown user location to
0.0 1.0 1.0 1.0 0.05 -0.01 0.07 0.06 the different deep models at each fingerprinting location and
leverage the reconstruction similarity in a probabilistic frame-
Masked vector Noisy vector work to estimate the most probable location. The intuition is
0.0 0.0 0.5 0.3 0.25 0.0 0.57 0.36 that the reconstructed scan will be closer to the input scan
in the fingerprinting locations near the actual user location.
(a) Masking corruption method. (b) Additive Gaussian corruption hence, these locations should have a higher score/probability
method. compared to far-away locations.
Fig. 3: Examples of different corruption techniques applied to More formally, the user is standing at an unknown location
the normalized RSS signal. l receiving WiFi information that is preprocessed to obtain
a signal strength vector x = (x1 , ..., xM ), where M is the
total number of APs in the environment. We want to find the
probability of being at a fingerprint location li in the area of
tuning stage. Now, we discuss the details of each of these interest given the received signal strength vector x. That is, we
stages. want to find P (li |x). Using the Bayes theorem, the posterior
1) Pre-training Stage: Fig. 4 shows the steps of the pre- probability P (li |x) is given as:
training deep neural network of stacked denoising autoen-
coders. The goal of this stage is to find good initial weights p(x|li )p(li ) p(x|li )p(li )
p(li |x) = = N (2)
i=1 p(x|li )p(li )
for the different layers of the network, instead of using p(x)
initial random weights [28]. This is known to speed up the
convergence of the training, reduce the possibility of falling in Where p(li ) is the prior probability that the phone is located
a local minima, and avoid the vanishing gradient problem [29]. at a given fingerprint location li and N is the number of
In this stage, each autoencoder in the stack is trained locations in the fingerprint database (i.e, the training locations).
independently using the output of the previous autoencoder as Assuming that all locations are equally probable1 , Equation 2
its input (input data for the first autoencoder). Each denoising can be rewritten as:
autoencoder consists of an encoder and a decoder. The encoder p(x|li )
tries to generate a latent representation (hidden layer param- p(li |x) = N (3)
i=1 p(x|li )
eters) of the input data while the decoder tries to reconstruct
the input data based on the generated (latent) code from the In traditional fingerprinting systems, e.g. [6], p(x|li ) is
encoder. usually obtained by assuming the independence of the APs
2) Fine-tuning Stage In this stage, we train the model using the RSS histograms, which does not capture the rich
(Fig. 5) end-to-end. The weights are initialized to those and correlated relation between the different APs. On the
obtained from pre-training stage. Then, each input training contrary, to calculate p(x|li ) WiDeep leverages the constructed
sample (WiFi scan) is passed through network to obtain the offline deep learning models. Specifically, reconstructed ver-
reconstructed data of the input scan using forward propagation. sions of the input scan xi are obtained from each deep model
The sum of squared difference between the original input data along with the associated similarity score to the input signal.
and the reconstructed data (i.e., output of the deep network) is To obtain this similarity score, we use a radial basis kernel
used as the loss function to adjust all the weights in different as a similarity function since its output is bounded within 0 and
layers with the gradient descent algorithm [30]. 1 and can therefore be interpreted probabilistically. Denoting
the output of the similarity function as p(x|li ) for the ith
C. Reducing Model Over-training
model, we have that:
To further reduce the possibility of model over-fitting, we
n
also use dropout regularization [27] during the fine-tuning 1  − xij −xˆij 
phase. To do that, some hidden neurons in the deep neural p(x|li ) = e λσ (4)
n j=1
network are temporarily dropped out stochastically (as illus-
trated by the crossed circles in Fig. 5). This dynamic change of Where xij and xˆij are the original and the reconstructed input
the network structure allows the network to generalize better data of the j th scan respectively, σ is the variance of the input
and hence become more robust to changes. scans, λ is selected to be a scaled version of the coefficient of
Finally, the learned weights of the N deep models for the variation (CV) of the input scans, and n is the total number of
N fingerprint locations in the area of interest are saved. Later scans used in location determination. We quantify the effect
on, during the online localization phase, these weights are used of these parameters on performance in Section IV.
to estimate the unknown location of the mobile device as we
explain in the next section. 1 If the user location profile is known, it can be used directly in Equation 2.
 Autoencoder 1 Autoencoder 2 Autoencoder 3 Autoencoder 4

Corruption

Original
RSS vector
Fig. 4: Greedy layer-wise pre-training process. The latent vector of every trained layer is used to train the subsequent layer.

Model
elN
TABLE I: Summary of used testbeds parameters.
N
Model
Mode
del
el2
el 2 Testbed University Apartment
Model
M dell1 1
Area 37m × 17m 14.5m × 4.5m
Number of APs 122 59
Density of APs (AP/m2 ) 0.19 1.05
Corruption

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
Fig. 5: Deep network architecture. Crossed nodes are example
are collected at 48 different locations in the university dataset.
of dropped-out neurons. Note that the output of this network
For the apartment dataset,2000 samples are collected at each
is connected to the probabilistic location inference module.
point of 139 different locations. The data is collected by five
participants using different Android phones (e.g., Samsung
Till now, we can assign different probabilities to the discrete Galaxy Note 3, Samsung Galaxy S4, Huawei P9 lite, among
fingerprinting locations based on the input scan(s). To enable others) over different days. This captures the time-variant
tracking the user in the continuous space, WiDeep estimates nature of the WiFi fingerprints as well as the heterogeneity
the user location as the center of mass of all fingerprinting of users and devices.
points [31], taking the probability of each reference point We implemented a WiFi collector App using the Android
P (li |x) as its weight. Hence, the user location l is estimated SDK to scan APs. The program records the (MAC address,
as RSS, timestamp) for each heard WiFi AP. The scanning rate
N was set to one per second. We implemented our deep learning
l= p(li |x)li (5) based training using the Google TensorFlow [32] framework
i=1 on the Google Collaboratory Cloud2 . 40% of the data points
IV. E VALUATION are held out for testing. We experimented with different deep
learning architectures and the one of 200×300×400×500
In this section, we evaluate the performance of WiDeep in obtains the best performance.
two typical indoor environments: a university building floor
and an apartment. We start by describing the data collec- B. Effect of Changing WiDeep Parameters
tion methodology. Next, we analyze the effect of different In this section, we study the effect of the different parame-
parameters on the WiDeep system performance. Finally, we ters on the system performance including different techniques
compare our system to the state-of-the-art WiFi fingerprinting used to add noise to the input, dropout regularization, different
techniques that are Horus [6] and DeepFi [8]. values of radial basis function parameters used for probabilistic
A. Data Collection online localization, number of input scans used in estimation,
and number of deep learning model layers. Table II shows
To collect the necessary data for evaluation, we deployed the default parameters values used throughout the evaluation
our system in two buildings with different layouts and APs section.
densities (Table I). The first one is a floor of our university 1) Effect of different input noise corruption techniques:
building with a 37m × 17m area containing offices, labs, We experimented with two different input noising methods
meeting rooms as well as corridors (Fig. 6). The second one, (Section III-B1): the masking method and additive Gaus-
shown in Fig. 7, is an L-shaped private studio apartment with sian noise method. Fig. 8 shows the mean location error
a 14.5m × 4.5m area. In both datasets, we leverage the RSSs
of pre-installed WiFi APs in the building or overheard from 2 https://colab.research.google.com
 TABLE II: Default parameters.
Parameter Range Default value
University,
Testbed University
Apartment
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-
0.25 - 6 4
rameter (λ)
Number of scans per esti-
1 - 30 7
mate (n)
Number of epochs 1 - 20000 10000
Fig. 6: University floorplan. 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 λ parameter in the Radial Basis Function:
Fig. 12 shows the effect of using different values of λ in the
radial basis function on the localization error. The figure shows
that there is an optimal value for λ at 4.
5) Effect of number of layers in the network: Fig. 13
shows the effect of changing the number of layers (stacked
Fig. 7: Appartment floorplan. 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
of WiDeep when trained using different masking corruption network begins to overfit the training data.
fraction values. Similarly, Fig. 9 shows the mean error of using
different standard deviation values to train the models using C. Robustness Experiments
the additive Gaussian method. The figures show that adding In this section, we assess the robustness of WiDeep under
more variations, by injecting artificial noise, of the data input different challenging scenarios including reducing the density
to the model enhances performance compared to the case when of access points and reduced number of fingerprint locations.
when the noise level is zero. However, adding too much noise 1) Density of access points: Fig. 14 shows the effect of
distort the signal and increase the ambiguity between adjacent reducing the density of the access points on accuracy. For
locations, hence increases the localization error. An optimal this, we uniformly removed access points from the total access
noise level for both techniques can be achieved at masking points detected in the area. The figure shows that even with
probability f = 0.1 and s = 0.1. a density as low as 5% of the access points, WiDeep can
2) Effect of dropout regularization: Fig. 10 shows the effect achieve high accuracy of less than 2.8m mean error. This is
of the dropout regularization on the mean localization accu- due to the different noise-handling techniques as well as the
racy. The figure shows that as the dropout rate is increased, the used regularization techniques. This highlights the robustness
mean localization accuracy is improved. This is aligned with of WiDeep.
our expectation because the regularization reduces the over- 2) Density of training points: Fig. 15 shows the per-
fitting of the neural network to the input data. However, using formance WiDeep when the number of training fingerprint
a large dropout rate leads to a decrease in the localization locations is reduced. The figure shows that, even though
accuracy as the network does not fit the data well (i.e., under- reducing the number of training points/percentage linearly
fitting). Therefore, this trade-off is balanced at 0.5 dropout leads to increasing the area associated with each training point
rate. quadratically, the decrease in the accuracy does not grow as
3) Effect of the number of scans per estimate: Fig. 11 fast. WiDeep can achieve a mean accuracy of 3.16m even with
shows the effect of increasing the number of the scans (n) used 50% reduction in training points.
to estimate a location in the online phase. The figure shows
that as n increases, the accuracy improves until it reaches an D. Comparative Evaluation
optimal value at n = 7 beyond which it begins to deteriorate. In this section, we compare the location accuracy, robustness
This is due to two opposing factors: (1) By increasing n, we to heterogeneous devices, and runtime of the WiDeep system
get more information which is useful for location estimate, against two baseline systems. The first is a popular probabilis-
(2) However, as n increases, more time is spent to collect tic fingerprinting based indoor localization technique (Horus
 3.9 3.5 3.1
Mean location accuracy (m)

Mean location accuracy (m)

Mean location accuracy (m)

3.6 3.4 3

3.3 3.3 2.9

3 3.2 2.8

2.7 3.1 2.7

2.4 3 2.6
0 0.1 0.3 0.5 0 0.04 0.08 0.12 0.16 0.2 0 0.2 0.4 0.6 0.8
Corruption fraction (f) Noise standard deviation (s) Dropout rate (r)

Fig. 8: Effect of the masking probability Fig. 9: Effect of the noise standard devi- Fig. 10: Effect of the dropout regulariza-
!t on accuracy. ation on accuracy. tion rate on accuracy.

2.9 3 3.1
Mean location accuracy (m)

Mean location accuracy (m)

Mean location accuracy (m)

2.85 2.9 3
2.8
2.8 2.9
2.75
2.7 2.8
2.7

2.65 2.6 2.7

2.6 2.5 2.6

3 4 5 6 7 8 9 10 1 2 3 4 5 6 3 4 5 6
Number of scans per estimate (n) Lambda Number of layers

Fig. 11: Effect of the number of scans Fig. 12: Localization error vs. RBF Fig. 13: Effect of the number of layers
per estimate on the accuracy. parameterλ. on accuracy.

[6]) that assumes the independence of the APs and the second 2) Device heterogeneity: Here, we evaluate the differ-
is a recent deep-learning based indoor localization technique ent techniques robustness to devices heterogeneity. Initially,
(DeepFi [8]) that does not perform noise handling or model WiDeep is trained and tested with the same device (i.e.
over-fitting avoidance. Samsung Galaxy Note 3). We then carry out experiments by
1) Localization accuracy: Fig. 16 and Fig. 17 show the training the different systems with the Samsung Galaxy Note
CDF of distance error for all systems in the university and 3 tablet and testing with a Samsung S4 mini smartphone. The
apartment testbeds. Table III and IV summarize the results. two devices have completely different form factors and WiFi
The figures illustrate that our WiDeep system can achieve chips.
significantly better mean localization accuracy than the other Fig. 18 shows that WiDeep provides approximately the same
systems by at least 29.8% and up to 169% in university and accuracy when testing with different device as when testing
apartment testbeds, respectively. Moreover, WiDeep enhances with the same device. It can also be seen from the figure
all the other quantiles. This can be explained by noting that WiDeep has the best performance in handling device
that traditional probabilistic fingerprinting techniques such as heterogeneity compared to the other two systems across all
Horus [6] cannot capture the correlation between the different percentiles. This is due to the combination of additive noise
APs and the fingerprinting locations. Similarly, traditional in the training data and the adoption of denoising autoen-
deep learning techniques such as DeepFi [8] do not take coders which gives WiDeep greater flexibility than the other
the inherent noise of the wireless signals into consideration systems. In particular, this is true since device heterogeneity
nor avoid over-training. Therefore, their performance drops can be considered to be a form of noise, which the WiDeep
noticeably when trained with noisy data. This can be seen in network and training process are designed specifically to
figures 16 and 17, where the accuracy of DeepFi degrades in combat. Horus also shows better adaptability than DeepFi.
such scenarios while WiDeep maintains its accuracy. This can be attributed to the fact that it utilizes probabilistic
Note also that WiDeep performance is consistent in the techniques, which are known to perform well in the presence
two testbeds, contrary to the other two techniques: DeepFi of uncertainty or noise. On the other hand, DeepFi shows
performs better in the testbed with more available data (i.e poor performance to noisy data because of the lack of specific
higher density of APs and training locations) while Horus can provisions to handle such phenomena in its design.
tolerate better the lower APs density and lower in the other 3) Time per location estimate: Fig. 19 compares the run-
testbed. ning time per location estimate for the three techniques. The
 3.2 1
2.8
Mean location accuracy (m)

Mean location accuracy (m)

3.1 0.8
2.75 3
0.6

CDF
2.9
2.7
0.4
2.8
WiDeep
2.65 0.2 Horus [6]
2.7
DeepFi [8]
DeepFi with noisy data[8]
2.6 2.6 0
5% 10% 20% 30% 40% 0% 15% 25% 30% 50% 0 5 10 15 20 25 30
Percentage of used APs Reduction percentage of trining points Location error (m)

Fig. 14: Effect of density of APs on Fig. 15: Effect of reducing the number Fig. 16: Comparison of CDFs in the
accuracy. of training locations on accuracy. University testbed.

1 40 500
DeepFi [8]
35 DeepFi with noisy data [8]
0.8 Horus [6] 400
Location accuracy (m)

30 WiDeep
WiDeep tested with same device
0.6 25

Time (ms)
300
CDF

20
0.4
15 200
WiDeep
0.2 DeepFi [8] 10
Horus [6] 100
DeepFi with noisy data [8] 5
0
0 2 4 6 8 10 12 0 0
Location error (m) WiDeep DeepFi[8] Horus[6]

Fig. 17: Comparison of CDFs in the Fig. 18: Comparison based on device Fig. 19: Time per location estimate.
Apartment testbed. heterogeneity.

machine used for running the algorithms is an HP Omen challenging scenarios, e.g. in the presence of device hetero-
laptop with an i7 2.6 GHz processor, 16 GB RAM, and a geneity. At the same time, the use of noisy data alone with
Nividia GTX965 GPU. The figure shows that the running time DeepFi (which does not use this type of network model)
of WiDeep and DeepFi is comparable. Horus has the lowest leads to a degradation of the obtained localization accuracy.
running time per location estimate as its prediction is based Additionally, the dropout regularization of WiDeep ensures
on only Bayesian inference. Deep learning techniques, on the the quality of the final model by eliminating co-dependencies
other hand, need to pass the data through all the layers of between the constituent neurons [27], [33].
the network. Nonetheless, all techniques can estimate the user It can be seen from Fig. 9 that training WiDeep without
location in less than 412ms, which allows realtime tracking considering the noise injection process leads to a significant
of the user. This can be further enhanced if needed through drop of the localization performance to 3.47m. Similarly, Fig.
parallelization. 10 shows that the accuracy degrades to 2.69m without pe-
nalizing the training process with such dropout regularization.
E. Discussion Therefore, the combination of the network used and the regu-
larization techniques are able to yield significant improvements
WiDeep is designed to operate with heterogenous devices over traditional deep learning models.
without compromising on localization accuracy. The use of
a deep model alone (e.g as in DeepFi) cannot lead to this V. R ELATED W ORK
design goal. WiDeep is able to achieve this as a combination
of the particular choice of deep network used and the associ- In this section, we discuss the most relevant literature to
ated design considerations. Specifically, the stacked denoising our WiDeep system. In particular, we cover two categories:
autoencoder network used in WiDeep is, by definition, capable fingerprinting systems and crowd-sourcing systems.
of reconstructing the underlying input in the presence of noise
or distortion. Therefore for the best results, the training process A. Fingerprinting Systems
of this network necessitates the use of noisy data so that the Fingerprinting systems present the most popular WiFi-based
network truly learns to extract the underlying information from indoor localization technique due to their high accuracy. Those
the data as obtained from users’ heterogeneous devices [25]. can be categorized into traditional and deep learning-based
This enhances the generalization ability of the network in fingerprinting systems.
 1) Traditional Fingerprinting Systems: Radar [12] finger- TABLE III: Accuracy percentiles of different systems in the
print captures the average RSS of the heard APs at the differ- University floorplan
ent fingerprint locations. During the online phase, matching Technique Average 50th 75th 100th
is based on using the k-nearest neighbors algorithm [34]. Percentile Percentile Percentile
Deterministic approaches though cannot deal well with he WiDeep 2.64m 2.38m 3.38m 7.12m
noise and variations of the RF signal. To tackle the noisy Horus [6] 4.04m 2.25m 4.03m 17.50m
nature of RSS, probabilistic techniques have been proposed, (-53.03%) (5.46%) (-19.23%) (-145.78%)
e.g. Horus [6], [35], [36]. In this case, the fingerprint reflects DeepFi [8] 7.10m 6.09m 9.54m 24.14m
the RSS histogram for each AP at each reference location, (-168.93%) (-155.88%) (-182.24%) (-239.04%)
assuming APs are independent. The most probable location
is estimated based on Bayesian inference. Many variants of TABLE IV: Accuracy percentiles of different systems in the
probabilistic techniques have been proposed over the years apartment floorplan
to further enhance the localization performance [5], [7], [37].
For instance, [7] uses radial basis networks to predict the Technique Average 50th 75th 100th
Percentile Percentile Percentile
unknown location. Despite probabilistic techniques being able
to handle the inherently noisy wireless signals in a better way WiDeep 1.21m 1.07m 1.62m 3.74m
than deterministic techniques, they usually assume that the DeepFi [8] 1.57m 1.25m 2.04m 4.97
signals from different APs are independent to avoid the curse (-29.75%) (-16.82%) (-25.92%) (-32.88%)
of dimensionality problem [19]. This leads to coarse-grained Horus [6] 2.57m 2.17m 3.35m 8.73
(-112.39%) (-102.08%) (-106.79%) (-133.42%)
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 sensors, which may not be available on all mobile devices,
noise in the RF signals. Moreover, it has provisions to handle especially in development countries where low-end phones are
over-fitting, leading to better robustness. more common. WiDeep can benefit from crowd-sourcing to
2) Deep Learning Systems: Recently, different deep learn- construct its fingerprint in an automatic manner. In addition,
ing techniques have been proposed in order to train models to based on deep learning and its different noise and robustness
provide a localization service. In DeepFi [8], [9], Restricted handling modules it can provide robust and high accuracy
Boltzman Machines are used to pre-train a deep learning localization without the need of any additional sensors.
system. The localization service of DeepFi depends on the
magnitudes of the channel state information (CSI) data, as VI. C ONCLUSION
compared to the standard received signal strength. Later, We presented WiDeep, an accurate and robust WiFi fin-
deep convolution networks based on CSI data also have been gerprinting indoor localization technique based on a deep
proposed to estimate the unknown locations [38], [39]. All neural network. The system leverages stacked denoising auto-
these techniques use CSI data, which needs special hardware encoders in a probabilistic framework to mitigate the noise
for collection, reducing the system ubiquity. In addition, they in the RSS measurements. Additionally, it employs model
do not have provisions to reduce over-fitting or handle the regularization to enable the network to generalize and avoid
inherent noise in the input data, reducing their robustness. over-fitting, leading to a more robust and stable models.
In contrast, the operation of WiDeep depends on standard We evaluated WiDeep in two different challenging envi-
RSS readings, which can be received by the common on-board ronments that represent a university building and a domestic
WiFi radio present in all mobile devices using standard APIs apartment using different Android devices. The results show
in the operating system. In addition,WiDeep is designed to the WiDeep comes with a localization accuracy better than
deal with noisy data and have provisions to avoid model over- the state-of-the-art systems by at least 53% and 29.8% in the
fitting, both leading to higher accuracy and more robustness. large and small environments respectively. Moreover, its per-
formance is robust to different devices and different densities
B. Crowdsourcing Systems of APs in different environments.
To reduce the fingerprint construction overhead, a num-
ber of systems have been introduced in which the users R EFERENCES
collaborate to improve the localization system by crowd- [1] P. Davidson and R. Piché, “A survey of selected indoor positioning
sourcing the fingerprint. [40] uses crowd-sourcing to improve methods for smartphones,” IEEE Communications Surveys & Tutorials,
the particle filter performance overtime and hence improve the vol. 19, no. 2, pp. 1347–1370, 2017.
[2] D. Lymberopoulos and J. Liu, “The microsoft indoor localization
localization accuracy of the system. Other systems, e.g. [15], competition: Experiences and lessons learned,” IEEE Signal Processing
[16], [20], [41]–[43], use the smartphone inertial sensors to Magazine, vol. 34, no. 5, pp. 125–140, 2017.
calculate the user location using dead-reckoning and leverage [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
different sensor-based landmarks, including WiFi, to reset the refinement,” IEEE transactions on mobile computing, vol. 7, no. 9, pp.
accumulated error. These system, however require additional 1110–1123, 2008.
 [4] N. Lasla, M. F. Younis, A. Ouadjaout, and N. Badache, “An effective [24] Y. Kang, K.-T. Lee, J. Eun, S. E. Park, and S. Choi, “Stacked denoising
area-based localization algorithm for wireless networks,” IEEE Trans- autoencoders for face pose normalization,” in International Conference
actions on Computers, vol. 64, no. 8, pp. 2103–2118, 2015. on Neural Information Processing. Springer, 2013, pp. 241–248.
[5] M. Youssef and A. Agrawala, “Handling samples correlation in the Ho- [25] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.-A. Manzagol,
rus system,” in INFOCOM 2004. Twenty-third AnnualJoint Conference “Stacked denoising autoencoders: Learning useful representations in a
of the IEEE Computer and Communications Societies, vol. 2. IEEE, deep network with a local denoising criterion,” Journal of Machine
2004, pp. 1023–1031. Learning Research, vol. 11, no. Dec, pp. 3371–3408, 2010.
[6] ——, “The Horus WLAN location determination system,” in Proceed- [26] M. Youssef and A. K. Agrawala, “Small-scale compensation for WLAN
ings of the 3rd international conference on Mobile systems, applications, location determination systems.” in WCNC, 2003, pp. 1974–1978.
and services. ACM, 2005, pp. 205–218. [27] H. Rizk, M. Torki, and M. Youssef, “CellinDeep: Robust and Accurate
[7] C. Laoudias, P. Kemppi, and C. G. Panayiotou, “Localization using Cellular-based Indoor Localization via Deep Learning,” IEEE Sensors
radial basis function networks and signal strength fingerprints in wlan,” Journal, 2018.
in Global telecommunications conference, 2009. GLOBECOM 2009. [28] Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle, “Greedy layer-
IEEE. IEEE, 2009, pp. 1–6. wise training of deep networks,” in Advances in neural information
processing systems, 2007, pp. 153–160.
[8] X. Wang, L. Gao, S. Mao, and S. Pandey, “DeepFi: Deep learning
[29] D. Erhan, Y. Bengio, A. Courville, P.-A. Manzagol, P. Vincent, and
for indoor fingerprinting using channel state information,” in Wireless
S. Bengio, “Why does unsupervised pre-training help deep learning?”
Communications and Networking Conference (WCNC), 2015 IEEE.
Journal of Machine Learning Research, vol. 11, no. Feb, pp. 625–660,
IEEE, 2015, pp. 1666–1671.
2010.
[9] ——, “CSI-based fingerprinting for indoor localization: A deep learning [30] G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of
approach,” IEEE Transactions on Vehicular Technology, vol. 66, no. 1, data with neural networks,” science, vol. 313, no. 5786, pp. 504–507,
pp. 763–776, 2017. 2006.
[10] W. Zhang, K. Liu, W. Zhang, Y. Zhang, and J. Gu, “Deep neural [31] M. Youssef and A. Agrawala, “Continuous space estimation for WLAN
networks for wireless localization in indoor and outdoor environments,” location determination systems,” in Computer Communications and Net-
Neurocomputing, vol. 194, pp. 279–287, 2016. works, 2004. ICCCN 2004. Proceedings. 13th International Conference
[11] R. Elbakly and M. Youssef, “A calibration-free rf localization system,” on. IEEE, 2004, pp. 161–166.
in Proceedings of the 23rd SIGSPATIAL International Conference on [32] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S.
Advances in Geographic Information Systems. ACM, 2015, p. 63. Corrado, A. Davis, J. Dean, M. Devin et al., “Tensorflow: Large-scale
[12] P. Bahl and V. N. Padmanabhan, “RADAR: An in-building RF-based machine learning on heterogeneous distributed systems,” arXiv preprint
user location and tracking system,” in INFOCOM 2000. Nineteenth arXiv:1603.04467, 2016.
Annual Joint Conference of the IEEE Computer and Communications [33] A. Shokry, M. Torki, and M. Youssef, “DeepLoc: a ubiquitous accurate
Societies. Proceedings. IEEE, vol. 2. Ieee, 2000, pp. 775–784. and low-overhead outdoor cellular localization system,” in Proceedings
[13] H. Abdel-Nasser, R. Samir, I. Sabek, and M. Youssef, “MonoPHY: of the 26th ACM SIGSPATIAL International Conference on Advances in
Mono-stream-based device-free WLAN localization via physical layer Geographic Information Systems. ACM, 2018, pp. 339–348.
information,” in Wireless communications and networking conference [34] H. Rizk, S. Elgokhy, and A. Sarhan, “A hybrid outlier detection
(WCNC), 2013 IEEE. IEEE, 2013, pp. 4546–4551. algorithm based on partitioning clustering and density measures,” in
[14] K. Habak, K. A. Harras, and M. Youssef, “Bandwidth aggregation Computer Engineering & Systems (ICCES), 2015 Tenth International
techniques in heterogeneous multi-homed devices: A survey,” Computer Conference on. IEEE, 2015, pp. 175–181.
Networks, vol. 92, pp. 168–188, 2015. [35] M. A. Youssef, A. Agrawala, and A. U. Shankar, “Wlan location
[15] H. Wang, S. Sen, A. Elgohary, M. Farid, M. Youssef, and R. R. determination via clustering and probability distributions,” in Pervasive
Choudhury, “No need to war-drive: Unsupervised indoor localization,” Computing and Communications, 2003.(PerCom 2003). Proceedings of
in Proceedings of the 10th international conference on Mobile systems, the First IEEE International Conference on. IEEE, 2003, pp. 143–150.
applications, and services. ACM, 2012, pp. 197–210. [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,”
[16] H. Abdelnasser, R. Mohamed, A. Elgohary, M. F. Alzantot, H. Wang,
in Personal Indoor and Mobile Radio Communications (PIMRC), 2011
S. Sen, R. R. Choudhury, and M. Youssef, “SemanticSLAM: Using
IEEE 22nd International Symposium on. IEEE, 2011, pp. 1208–1212.
environment landmarks for unsupervised indoor localization,” IEEE
[37] M. Youssef and A. Agrawala, “Location-clustering techniques for wlan
Transactions on Mobile Computing, vol. 15, no. 7, pp. 1770–1782, 2016.
location determination systems,” International Journal of Computers and
[17] M. Ibrahim, M. Torki, and M. ElNainay, “Cnn based indoor localization Applications, vol. 28, no. 3, pp. 278–284, 2006.
using rss time-series,” in 2018 IEEE Symposium on Computers and [38] H. Chen, Y. Zhang, W. Li, X. Tao, and P. Zhang, “ConFi: Convolutional
Communications (ISCC). IEEE, 2018, pp. 01 044–01 049. Neural Networks Based Indoor Wi-Fi Localization Using Channel State
[18] M. Youssef, M. Abdallah, and A. Agrawala, “Multivariate analysis for Information,” IEEE Access, vol. 5, pp. 18 066–18 074, 2017.
probabilistic WLAN location determination systems,” in The Second [39] X. Wang, X. Wang, and S. Mao, “CiFi: Deep convolutional neural
Annual International Conference on Mobile and Ubiquitous Systems: networks for indoor localization with 5 GHz Wi-Fi,” in Communications
Networking and Services. IEEE, 2005, pp. 353–362. (ICC), 2017 IEEE International Conference on. IEEE, 2017, pp. 1–6.
[19] N. M. Nasrabadi, “Pattern recognition and machine learning,” Journal [40] Z. Liu, L. Zhang, Q. Liu, Y. Yin, L. Cheng, and R. Zimmer-
of electronic imaging, vol. 16, no. 4, p. 049901, 2007. mann, “Fusion of magnetic and visual sensors for indoor localization:
[20] H. Aly and M. Youssef, “Dejavu: an accurate energy-efficient outdoor Infrastructure-free and more effective,” IEEE Transactions on Multime-
localization system,” in Proceedings of the 21st ACM SIGSPATIAL Inter- dia, vol. 19, no. 4, pp. 874–888, 2017.
national Conference on Advances in Geographic Information Systems. [41] H. Aly, A. Basalamah, and M. Youssef, “Map++: A crowd-sensing
ACM, 2013, pp. 154–163. system for automatic map semantics identification,” in 2014 Eleventh
[21] M. Elhamshary and M. Youssef, “CheckInside: a fine-grained indoor Annual IEEE International Conference on Sensing, Communication, and
location-based social network,” in Proceedings of the 2014 ACM In- Networking (SECON). IEEE, 2014, pp. 546–554.
ternational Joint Conference on Pervasive and Ubiquitous Computing. [42] N. Mohssen, R. Momtaz, H. Aly, and M. Youssef, “It’s the human
ACM, 2014, pp. 607–618. that matters: accurate user orientation estimation for mobile computing
[22] M. Elhamshary, M. Youssef, A. Uchiyama, H. Yamaguchi, and T. Hi- applications,” in Proceedings of the 11th International Conference on
gashino, “Transitlabel: A crowd-sensing system for automatic labeling Mobile and Ubiquitous Systems: Computing, Networking and Services.
of transit stations semantics,” in Proceedings of the 14th Annual In- ICST (Institute for Computer Sciences, Social-Informatics and , 2014,
ternational Conference on Mobile Systems, Applications, and Services. pp. 70–79.
ACM, 2016, pp. 193–206. [43] H. Aly, A. Basalamah, and M. Youssef, “Lanequest: An accurate and
[23] S. Sen, B. Radunovic, R. R. Choudhury, and T. Minka, “You are energy-efficient lane detection system,” in Pervasive Computing and
facing the mona lisa: spot localization using phy layer information,” Communications (PerCom), 2015 IEEE International Conference on.
in Proceedings of the 10th international conference on Mobile systems, IEEE, 2015, pp. 163–171.
applications, and services. ACM, 2012, pp. 183–196.

View publication stats