Applying low cost WiFi-based localization to in-campus autonomous vehicles

Full text

Applying Low Cost WiFi-based Localization to
In-Campus Autonomous Vehicles
Noelia Hern´andez∗, Ahmed Hussein†, Daniel Cruzado†, Ignacio Parra∗and Jos ´e Mar´ıa Armingol†
∗Intelligent Vehicles and Traffic Technologies Group, Universidad de Alcal´a, Alcal´a de Henares, Spain
Email: noelia.hernandez@edu.uah.es, ignacio.parra@aut.uah.es
†Intelligent Systems Laboratory, Universidad Carlos III de Madrid, Legan ´es, Spain
Email: ahussein@ing.uc3m.es, dcruzado@pa.uc3m.es, armingol@ing.uc3m.es
Abstract—In this paper a new approach to provide low cost
localization to autonomous vehicles is proposed. It is based on
fingerprint WiFi localization improved by using Support Vector
Regression to increase localization resolution without the need of
increasing the number of positions to site-survey. Results shown
that the proposed method can emerge as a powerful tool to
provide localization at situations where the use of GPS is not
suitable.
I. INTRODUCTI ON
Autonomous driving has become a blooming topic among
car makers and research centres all across the globe in the past
years since the announcement of Google’s self-driving car in
2010. Since then, the interest of car makers in self-driving
has not ceased to grow and, as a matter of fact, autonomous
driving developments and publications have soared worldwide.
Despite rapid technological development, a number of is-
sues, not only legal, have still to be seriously addressed before
autonomous cars can robustly, safely, and efficiently circulate
and mix with manually-driven vehicles in real traffic. Experts
in the field agree that autonomous vehicles will become more
robust as they develop further cooperation capabilities. In other
words, cooperation with traffic infrastructure, as well as with
other agents, will make autonomous vehicles more robust
and reliable, given that it is widely accepted that standalone
self-driving is by far less robust than cooperative automated
driving.
The scientific community is clearly moving in this direction.
After the first solid demonstrations of self-driving cars in urban
scenarios carried out by Google [1], the University of Parma
[2], Daimler and KIT on the Bertha Route [3], and a number
of other car makers, such as Tesla and Nissan, much research
on automated driving has begun to be developed.
But the interest on autonomous navigation is not restricted
to road vehicles. Many other applications have emerged fol-
lowing this trend such as warehouse autonomous robots (Ama-
zon’s Kiva), Unmanned Aerial Vehicles (UAV) [4], or small
autonomous vehicles for transportation on airports, parks,
museums, etc [5].
In this line of work, the iCab (Intelligent Campus Automo-
bile) project [6] transports people in the Universidad Carlos
III de Madrid Campus at Legan´es (Figure 1). This platform
consists in an electric golf cart equipped with optical encoders,
a binocular camera, a laser-range-finder, a compass and a
GPS module. The autonomous navigation is performed using
Robotic Operating System (ROS) architecture implemented
in the on-board embedded computer. This computer is in
charge of performing all the necessary tasks (vehicle control,
navigation, route planning, perception and obstacle avoidance)
so the available computational resources are limited.
Fig. 1. iCab platform.
To provide a robust and safe autonomous navigation, there
are some tasks that must be accomplished: route planning so
the vehicle is able to reach the desired destination, navigation
so the vehicle is able to move through the environment and
obstacle avoidance so the vehicle is able to to navigate through
the environment safely. In order to accomplish these tasks,
an accurate localization of the vehicle is essential. With no
information about the position of the vehicle in the world it is
impossible to reach a destination or move only over allowed
areas.
Traditionally, localization is provided trough the use of GPS,
odometry, laser, the most recent LIDAR or the fusion of some
of them. Unfortunately, some of these technologies does not
provide accurate localization (as is the case of GPS in covered
areas or city canyons) while some others does not provide it
in real time or their cost is high (as is the case of LIDAR).
The system proposed in this paper is designed to solve the
localization problem in the iCab platform. Previous tests have
shown that GPS does not provide accurate positioning under
the conditions required for iCab localization and that LIDAR-
based positioning can not be performed in real time.

As a consequence, a new approach is needed to improve
iCab localization. In this paper, we proposed the use of
WiFi-based localization to provide outdoor positioning. WiFi-
based localization has been widely used to provide indoor
localization and has been proved as a cheap accurate solution
for this task. To the best of our knowledge this is the first work
using WiFi to provide outdoor localization to autonomous
vehicles.
The rest of the paper is structured as follows, the next
section describes the related work. Section III describes the
method proposed to obtain the position of the autonomous
vehicle. Next, in Section IV, the experimental set-up and the
results are detailed. Finaly, Section V concludes the paper by
exposing the conclusions and future work.
II. REL ATED W ORK
As mentioned in the introduction, WiFi has been proved
a very useful technology to provide indoor localization. It
has some important advantages: WiFi-based localization is
free, even for private networks, so the only need to provide
WiFi localization is to have a common WiFi interface and
WiFi networks in the surroundings. Nowadays, it is almost
impossible to find a device (mobile phone, tablet or computer)
without a WiFi interface, and, in addition, almost every
building (specially public ones) has WiFi networks to provide
connectivity to their users. However, WiFi localization is still
an open problem. Traditionally, two different approaches have
been used to provide WiFi localization: Propagation model
based algorithms [7]–[12] and fingerprint based algorithms
[13]–[16].
The former carries out localization using a WiFi signal
propagation model to estimate the distance to nearby APs
(Access Points). These models are used to translate the WiFi
RSS (Received Signal Strength) from an AP into the distance
to that AP. In consequence, localization is usually performed
using lateration algorithms, being necessary to know the exact
location of the APs in the environment. The main advantage
of this kind of systems is that they are accurate if the
propagation model is well adjusted, providing localization over
a continuous space, at the cost of the need of knowing the exact
location of all the used APs which is not always possible, as
is our case.
The latter systems use a two-stage approach to provide lo-
calization. During the first step, the training stage, a fingerprint
database composed of RSS from the surrounding APs is col-
lected at different positions of the environment. Then, during
the localization stage, new samples are compared with the
stored ones to obtain the estimated position of the device (usu-
ally by means of different classification algorithms). The main
disadvantage of fingerprint based systems is that collecting
the measurements required to build the fingerprint database is
effort and time consuming, especially if the required resolution
is high (which means more cells to site-survey).
However, some systems are recently arising to reduce the
site-survey effort using different approaches. In the one hand,
some methods seek to automate the measuring task [17]–[20].
On the other hand, some systems are designing methods to
reduce the number of positions required to be measured by
estimating the expected RSS at positions with no available
data [21]–[24].
The proposed method lies in the second group, making
use of Support Vector Regression (SVR) to interpolate the
expected RSS at the positions with no available data [25].
III. MET HOD D ESCRIPTION
The main objective of the method is to estimate the RSS
in positions of the environment where no training data is
available. This way, the effort and time required to construct
the fingerprint database will be reduced and the resolution of
the localization system will be increased. A block diagram of
the training and localization stages of the method is shown
in Figure 2. Both stages will be explained in detail in the
following subsections.
A. Training stage
During the training stage, the system will convert the
training data available at discrete positions of the environment
into RSS reference surfaces for each AP. This conversion will
be performed using an SVR algorithm [26] to estimate the
missing data. As a result, the fingerprint database will be
composed of a set of continuous reference surfaces (one per
AP) containing the expected RSS at the different coordinates
in which the environment has been divided.
It is important to highlight that the localization resolution
will depend on the size of the reference surfaces, where
the pixels (or coordinates) are the squared areas in which
the environment is divided. As an example, if the target
environment is a squared area with side 100 metres and the
reference surfaces are divided in 100x100 coordinates (100
pixels wide 100 pixels high), each coordinate will cover 1m2.
In consequence, the maximum resolution will be 1 metre.
For the experimentation shown in this paper, a resolution of
15 cm has been chosen as it has been proved to be the most
effective. Higher resolutions will reduce the effectiveness of
the method while resolutions under the wavelength (for the
802.11 b/g networks, working at 2.4 GHz, the wavelength is
12.5 cm) the small scale effect [27] will highly affect the WiFi
signal increasing the error in localization [25].
B. Localization stage
During the localization stage, new WiFi measurements are
collected in real time. As in a traditional fingerprint localiza-
tion method, the objective is to compare the new measurements
with the stored ones to obtain the most probable location of
the vehicle.
To do so, the measured RSS from each AP will be searched
in the corresponding reference surface. As a result, a surface
of possible coordinates will be created for each AP. These new
surfaces will be created by scoring the coordinates depending
on the similarity between the on-line measurements and the
data contained in the reference surfaces. If the measured RSS
is exactly the same as the stored RSS for a coordinate, that

Fig. 2. General architecture of the system [25].
coordinate will get the maximum score (score equal to one).
Then, the score will decrease as the difference between the
measured and stored RSS increase. This way, if the scores are
defined in steps of 0.1, a coordinate with a difference of ±1dB
will score 0.9, a coordinate with a difference of ±2dB will
score 0.8 and so on, while all the coordinates with differences
of ±10dB or more will score 0. In our experiments, the scores
are defined in steps of 0.1 (corresponding to a maximum
difference of 10dB) [25]. This is the highest variation in the
RSS due to temporal variations [27] which are caused by
signal noise and non-controllable changes in the environment.
Once the scored surfaces for all the APs are created, they
are summed up obtaining a new surface where the coordi-
nates with the higher values are the coordinates with higher
probability of being the location of the vehicle. However, the
complete surface cover areas where it is impossible for the
vehicle to be (inside buildings, park areas, fountain areas, etc.).
As a consequence, a mask adapted to cover only allowed areas
is applied to prevent the system to provide impossible locations
as result.
Finally, the location of the vehicle is estimated as the
coordinate with the highest score in the masked resulting
surface.
IV. EXP ERIMENTAL ANALYSIS
This section describes the experimental results reported
using the method described in the previous section. First, the
experimental set-up used to perform the tests is described.
Then, an analysis of the validity of the reference surfaces
created to provide outdoor localization is provided. Finally,
localization results of the iCab platform obtained under real
conditions are described.
A. Experimental set-up
The proposed WiFi localization system has been tested in a
complex real-world environment with a surface of 12000m2.
The experiments have been performed on the campus of the
Universidad Carlos III de Madrid (UC3M) located in Legan´es
(Madrid, Spain) (Figure 3). The area selected to perform the
experiments is located in a green area surrounded by the
University buildings. In consequence, the GPS localization is
not accurate in this area. In addition, it is a crowded area since
lots of pedestrians use it to walk or rest. Since it is an open
area, the WiFi signal attenuation is really low, making more
difficult to differentiate between close positions since the RSS
at near positions is very similar.
Fig. 3. UC3M test-bed environment (30 training positions).
In the experiments, 175 APs, deployed over the environment
with the aim of providing Internet access to the students but
disregarding localization purposes, have been detected. There
is no information available about the APs, so their exact
location and configuration (variable emission power, channel,
etc) is unknown. All the APs have been used for localization.
In consequence, the localization system is composed of 175
reference surfaces.

The discrete training data is composed of measurements
collected at 30 positions uniformly distributed over the envi-
ronment covering the whole area (Figure 3). At each position,
20 fingerprints have been collected and averaged. In conse-
quence, one fingerprint per topological position has been used
as training data.
The tests have been carried out using data collected with
the iCab embedded computer while it was following different
trajectories over the environment. Test data was collected on
a different day and time than the training data and it was
acquired at the WiFi interface’s maximum allowed measuring
rate (1 sample per second).
Finally, the groundtruth has been obtained using a local-
ization algorithm using LIDAR information. This algorithm
has been proved to be accurate, but it has to be executed
off-line due to its high computational cost. This groundtruth
is used to calculate the mean distance error of the trajectory
tests. The distance error is computed as the distance between
the estimated coordinate and the target coordinate in the
groundtruth.
B. Reference surfaces validation
This section shows how the estimated RSS at coordinates
without measurements in the training dataset adjust to new
measurements collected in a posterior stage. To make the
validation results easier to understand, figures have been
simplified by using data collected in straight line trajectories
so they can be represented in a 2D graph. Namely, the RSS
from two different APs at positions 8 to 13 is shown in Figures
4 and 5. In these figures, the RSS values used to create the
reference surfaces (training data) are depicted as magenta stars,
the estimated RSS is represented using a red line (the dotted
red lines are the representation of the margin used to score the
new RSS values as explained in Section III-B) and, finally, the
RSS collected for validation purposes are represented as blue
circles.
As it can be seen, test data adjust to the estimated RSS
curve. However, there are still some samples falling outside
the scoring margin. This is due to the noisy nature of the WiFi
signal, being the main source of error in the localization.
Fig. 4. Reference surfaces validation: AP 1.
Fig. 5. Reference surfaces validation: AP 2.
C. Experimental results
The system has been tested using the data collected with
the iCab’s embedded computer during two repetitions of two
different trajectories (Figure 6).
(a) (b)
Fig. 6. Test trajectories. (a) Trajectory 1. (b) Trajectory 2.
On the first trajectory, the iCab moved in a path of approx-
imately 145 m at a mean speed of 4,86 km/h. Figure 6(a)
shows the trajectory followed during the first experiment. In
this figure, each blue circle represents the groundtruth location
where online WiFi samples were collected.
On the second trajectory, the iCab moved in a path of
approximately 135 m at a mean speed of 4,77 km/h. Figure
6(b) shows the second trajectory with the same format as in
Figure 6(a).
Table I summarises the results obtained for both trajectories
using the proposed WiFi localization method. In addition,
results applying a simple Kalman filter to the WiFi localization
results (without any other information) are shown to remove
some evident big jumps in the positioning. As can be seen, the
localization mean distance error is highly affected by jumps in
the position that can be decreased by applying a simple filter.
After filtering the results, error is reduced more than a 30%
for trajectory 1 and 40% for trajectory 2.
Figure 7 shows the CDF for both trajectories with and
without the filtering. CDFs show that some positioning errors
are very high (more than 40 metres during trajectory 1 and
more than 50 m during trajectory 2 for some samples) but

TABLE I
EXP ERIM ENTAL R ESU LTS: TEST U SIN G THE TR AJE CTOR Y DATAS ETS.
Mean error
WiFi Localization Filtered WiFi Localization
Trajectory 1 9.34 m 6.18 m
Trajectory 2 17.70 m 10.56 m
also that this error can be easily reduced (more than 50%) by
applying a filter over the results.
0 5 10 15 20 25 30 35 40 45
Error (meters)
0
10
20
30
40
50
60
70
80
90
100
CDF
Trajectory 1
Trajectory 1 filtered
Trajectory 2
Trajectory 2 filtered
Fig. 7. CDF using the trajectories dataset.
V. CO NCLUS ION S AND F UTURE WORK
In this paper, we presented a new method to provide low
cost positioning to the iCab project. This method improves
fingerprint WiFi-based localization methods by estimating the
RSS at non site-surveyed positions of the environment by us-
ing an SVR algorithm. Thanks to this method the localization
resolution can be improved while the effort in constructing
the fingerprint database is reduced. Experimentation shown
promising results, but we have found the need of fuse the
proposed method with the information provided by other
sensors.
In the future, we are planning on fusing the proposed
WiFi localization method with the compass and odometry
information. This way, the error will be reduced by filtering
locations far away from the previous ones and only allowing
locations in the heading direction of the vehicle.
ACK NOWLED GME NT
This work was supported by the projects TRA2016-78886-
C3-1-R from the Comisi´on Interministerial de Ciencia y
Tecnolog´ıa, S2013/MIT-2713 from the Comunidad Autonoma
de Madrid and Research Grant DPI2014-59276-R from the
Spanish Ministry of Economy).
REF ERE NCE S
[1] “Google self-driving car project,” accessed on April 2017. [Online].
Available: https://www.google.com/selfdrivingcar/
[2] A. Broggi, P. Cerri, S. Debattisti, M. C. Laghi, P. Medici, D. Molinari,
M. Panciroli, and A. Prioletti, “PROUD-Public Road Urban Driverless-
Car Test,” IEEE Transactions on Intelligent Transportation Systems,
vol. 16, no. 6, pp. 3508–3519, 2017.
[3] T. Dang, J. Ziegler, U. Franke, H. Lategahn, P. Bender, M. Schreiber,
T. Strauss, N. Appenrodt, C. Keller, E. Kaus, C. Stiller, and R. Herrtwich,
“Making Bertha drive - An autonomous journey on a historic route,”
IEEE Intelligent Transportation Systems Magazine, vol. 6, no. 2, pp.
8–20, 2014.
[4] A. Hussein, A. Al-Kaff, A. de la Escalera, and J. M. Armingol, “Au-
tonomous indoor navigation of low-cost quadcopters,” in Proceedings
of the 2015 IEEE International Conference on Service Operations And
Logistics, And Informatics (SOLI), 2015, pp. 133–138.
[5] C. Fern ´andez, R. Dom´ınguez, D. Fern´andez-Llorca, J. Alonso, and M. A.
Sotelo, “Autonomous navigation and obstacle avoidance of a micro-bus,”
International Journal of Advanced Robotic Systems, vol. 10, no. 4, p.
212, 2013.
[6] D. Gomez, P. Marin-Plaza, A. Hussein, A. Escalera, and J. M. Armingol,
“ROS-based architecture for autonomous intelligent campus automo-
bile (iCab),” UNED Plasencia Revista de Investigacion Universitaria,
vol. 12, pp. 257–272, 2016.
[7] A. Kotanen, M. H¨annik¨ainen, H. Lepp¨akoski, and T. D. H¨am¨al¨ainen,
“Experiments on local positioning with Bluetooth,” in Proceedings of
the International Conference on Information Technology: Coding and
Computing, 2003, pp. 297–303.
[8] A. Bose and C. H. Foh, “A practical path loss model for indoor
WiFi positioning enhancement,” in Proceedings of the International
Conference on Information, Communications Signal Processing, 2007,
pp. 1–5.
[9] S. Mazuelas, A. Bahillo, R. M. Lorenzo, P. Fernandez, F. A. Lago,
E. Garcia, J. Blas, and E. J. Abril, “Robust indoor positioning provided
by real-time RSSI values in unmodified WLAN networks,” IEEE Journal
of Selected Topics in Signal Processing, vol. 3, no. 5, pp. 821–831, 2009.
[10] J. Yang and Y. Chen, “Indoor localization using improved RSS-based
lateration methods,” in Proceedings of the IEEE conference on Global
telecommunications, 2009, pp. 4506–4511.
[11] F. Herranz, “Simultaneous localization and mapping using range only
sensors,” Ph.D. dissertation, University of Alcal´a, 2013.
[12] J. Yang, H. Lee, and K. Moessner, “Multilateration localization based on
singular value decomposition for 3d indoor positioning,” in Proceedings
of the 2016 International Conference on Indoor Positioning and Indoor
Navigation, 2016, pp. 1–8.
[13] P. Bahl and V. N. Padmanabhan, “RADAR: An in-building RF-based
user location and tracking system,” in Proceedings of the Annual Joint
Conference of the IEEE Computer and Communications Societies, 2000,
pp. 775–784.
[14] J. Torres-Sospedra, G. M. Mendoza-Silva, R. Montoliu, O. Belmonte,
F. Benitez, and J. Huerta, “Ensembles of indoor positioning systems
based on fingerprinting: Simplifying parameter selection and obtaining
robust systems,” in Proceedings of the 2016 International Conference
on Indoor Positioning and Indoor Navigation, 2016, pp. 1–8.
[15] W. Li, D. Wei, H. Yuan, and G. Ouyang, “A novel method of wifi fin-
gerprint positioning using spatial multi-points matching,” in Proceedings
of the 2016 International Conference on Indoor Positioning and Indoor
Navigation, 2016, pp. 1–8.
[16] T. Garc´ıa-Valverde, A. Garc´ıa-Sola, H. Hagras, J. Dooley, V. Callaghan,
and J. A. Bot´ıa, “A fuzzy logic-based system for indoor localization
using WiFi in ambient intelligent environments,” IEEE Transactions on
Fuzzy Systems, vol. 21, no. 4, pp. 702–718, 2013.
[17] C. Wu, Z. Yang, Y. Liu, and W. Xi, “WILL: Wireless indoor local-
ization without site survey,” in Proceedings of the IEEE International
Conference on Computer Communications, 2012, pp. 64–72.
[18] 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 International Conference on Mobile Systems,
Applications, and Services, 2012, pp. 197–210.
[19] A. Rai, K. K. Chintalapudi, V. N. Padmanabhan, and R. Sen, “Zee:
Zero-effort crowdsourcing for indoor localization,” in Proceedings of the
Annual International Conference on Mobile Computing and Networking,
2012, pp. 293–304.
[20] Z. Yang, C. Wu, and Y. Liu, “Locating in fingerprint space: Wireless
indoor localization with little human intervention,” in Proceedings of the
Annual International Conference on Mobile Computing and Networking,
2012, pp. 269–280.

[21] M. Sch ¨ussel and F. Pregizer, “Coverage gaps in fingerprinting based in-
door positioning: The use of hybrid gaussian processes,” in Proceedings
of the 2015 International Conference on Indoor Positioning and Indoor
Navigation, 2015, pp. 1–9.
[22] P. Richter, A. Pe ˜na-Torres, and M. Toledano-Ayala, “A rigorous evalua-
tion of gaussian process models for WLAN fingerprinting,” in Proceed-
ings of the 2015 International Conference on Indoor Positioning and
Indoor Navigation, 2015, pp. 1–10.
[23] K. Chintalapudi, A. Padmanabha Iyer, and V. N. Padmanabhan, “Indoor
localization without the pain,” in Proceedings of the Annual Interna-
tional Conference on Mobile Computing and Networking, 2010, pp.
173–184.
[24] G. Caso and L. D. Nardis, “On the applicability of multi-wall multi-floor
propagation models to wifi fingerprinting indoor positioning,” in Future
Access Enablers for Ubiquitous and Intelligent Infrastructures, ser.
Lecture Notes of the Institute for Computer Sciences, Social-Informatics
and Telecommunications Engineering, 2015, vol. 159, pp. 166–172.
[25] N. Hern ´andez, M. Oca ˜na, J. M. Alonso, and E. Kim, “Continuous space
estimation: Increasing WiFi-based indoor localization resolution without
increasing the site-survey effort,” Sensors, vol. 17 (1), no. 147, pp. 1–23,
2017.
[26] H. Drucker, C. J. C. Burges, L. Kaufman, A. Smola, and V. Vapnik,
“Support vector regression machines,” Advances in Neural Information
Processing Systems, vol. 9, no. 9, pp. 155–161, 1997.
[27] M. Youssef and A. Agrawala, “Small-scale compensation for WLAN
location determination systems,” in Proceedings of the IEEE Wireless
Communications and Networking, vol. 3, 2003, pp. 1974–1978.