Intelligent Systems Conference 2017

7-8 September 2017 | London, UK

Development of a Control System for Electric

Wheelchairs based on Head Movements

Guilherme Marins André Marcato †

and Diego Carvalho* and Ivo Júnior +
School of Electrical Engineering School of Electrical Engineering
Federal University of Juiz de Fora Federal University of Juiz de Fora
Juiz de Fora, Minas Gerais, Brazil Juiz de Fora, Minas Gerais, Brazil
Email: [email protected] Email: [email protected] †
Email: [email protected] * Email: [email protected] +

Abstract—This paper presents an economic solution for in- and we intend to implement in the next step a real wheelchair
dividuals with disabilities in the legs and arms. The goal is to based on the results obtained. Quadriplegics individuals have
use an IMU (Inertial Measurement Unit) to capture movements no arms and legs movements, however, are able to do the
from the user’s head and through this movement be able head movements perfectly. So, a module able to capture the
to move an electric wheelchair. For the extraction and data individual head movements makes possible the use of a chair
processing an Arduino Uno is used for data classification and
the MatLab R for a quadriplegic person. Thus, in this work is used a module
software for simulating the movement. Euclidean
distance, Mahalanobis distance, and artificial neural network was placed on the head that is capable of analyzing the user
used for data classification. In this way, we seek an alternative to movements and actuate in the wheelchair.
wheelchair users with greater limitations and minimizing costs
for development.
To analyse the movements, we first create a standard, like
presented in Fig. 1, and from the standard we use methods
Keywords—Robotics; wheelchair; accessibility of pattern recognition,using three methods in this article:
Euclidian distance, Mahalanobis and artificial neural network
to define the standard movements for the user and assign
I. I NTRODUCTION
correctly each movement due to his position, or like will be
In the current Brazilian scenario 45.6 million of people presented, a simulated robot.
have some kind of disability, and approximately 4.5 million
This article is organized as follows: In Section II is
have a severe physical limitation. Moreover, about 20 million
presented Related Works, Section III a theoretical foundation
disabled Brazilians have incomes below the minimum wage
to generally explain the proposed system, Section IV present
[10], [11].
the project development with materials and ways to achieve
Individuals with injuries from the C4 Neurological level the objective, Section V presents discussions and results from
are able to breathe almost normally. Generally, they are capa- the project built and a simulation test using the device and
ble of making head and neck movements. Thus, an electric finally in Section VI is presented the conclusions of this work
wheelchair that requires neck movements is effective for and possible future works.
patients with this type of injury [9], [14].
Based on this, a significant number of Brazilians require II. R ELATED W ORKS
the use of wheelchairs for their mobility, however, conventional
wheelchairs do not meet all kinds of physical disabilities, as Given the need of equipments that make locomotion of
tetraplegic individuals, for example, are not met by conven- disabled people easier and safe, several researches has been
tional chairs. Alternatively, there is widespread on the market made with this intention. Alternative solutions seek to offer
for the motorized wheelchairs, however, help a new portion of options for diferent kinds of disabilities and create equipments
the population, but are still not accessible to quadriplegics and that meet their needs [1]–[3], [12], [13].
according to most of the disability poverty rate, many do not Some authors have automated chairs with options for
have financial conditions of purchasing motorized wheelchairs use, among them manual, shared and autonomous control
available in the market. [13], [18], [19]. The article [13] first presents the completely
The work presented seek an mobility alternative to manual control, the shared control in which it receives the
quadriplegic individuals or with some disabilities that do not information desired by the user and adjusts the destination
allow the use of conventional wheelchairs, in addition to in order to deflect obstacles and improve their experience and,
seeking an alternative with low cost to offer an option to the finally, autonomous system that presents better results than the
parcel population with financial difficulties. previous methods. In the article [18] the voice command is also
used, for shared control, as an option for the disabled people
Unfortunately for the proposed work, there is only the with more severe injuries. In addition, the article [18] uses two
implementation of a module to drive a robot in simulation level ultrasonic sensors for localization and movement completely

978-1-5090-6435-9/17/$31.00 2017
c IEEE 996 | P a g e
 Intelligent Systems Conference 2017
7-8 September 2017 | London, UK

Fig. 2. Accelerometer operating principle.

A. Inertial Sensors
Inertial sensors measure acceleration and angular velocity
and are used in applications of movement analysis using
accelerometers and gyroscopes. It is common to use magne-
tometers to get in a complete manner the orientation variables
analysis.
Accelerometers are transducers that produce an output
Fig. 1. Diagram of movements. signal related to the acceleration, the more common are the
piezoelectric.
The gyroscope is a sensor which measures angular velocity.
autonomous, however, it should be noted that this model can Its operation is based on two parallel oscillating masses that are
be uncomfortable for the user. always in opposite directions and at the same velocity. When
The article [12] presents a shared control approach for an angular velocity is applied, the masses move, varying an
automated chairs, which allows the user to activate the chair internal capacitance. The sensor output is proportional to the
using the Emotiv EPOC’s Expressiv [20] device that sends difference between the two capacitance variations. Thus, in the
the user signals to the system through an interface. The case presented are obtained any change of position by reading
objective of the shared control developed in the described the gyroscope, because it will present a variation when the
work is to allow the autonomy of the user, using a route-based user change his position.
intelligence based on vector fields that adjusts the path desired
As explained above is used in this work one IMU to acquire
by the user so that it travels safely in the way that is expected.
the individual drive data. For this work, information acquired
An IMU is an eletronic device that acquires and sends is used by the accelerometer and gyroscope of the IMU. Such
data of angular velocity and movimentation of a body, using information is able to provide the positioning of the individual
a combination of accelerometer and gyroscope. This device is head and his movement according to the directions rotations
responsable for sending movimentation data for an Arduino Roll and Pitch (Fig. 3). It is important to highligth that, in this
that will treat them. This equipment is largely used in robotic case, it is not necessary Yaw informations, this way, it is used
aplications [4], [5]. the accelerometer and the gyroscope only in the axes x and y.
To data treatment is used an Arduino UNO that is a
microcontroller capable of receiving data from the IMU and B. Euclidean Distance Classifier
send to a computer.
The Euclidean distance is a pattern recognition method that
uses the distance between two points in two different classes in
III. T HEORETICAL F OUNDATION a plane. Thus, using the Euclidean distance (1) for separating
In this section is presented the main idea of inertial sensors the classes with respect to their average, given a vector in the
and the three methods used to pattern recognition. Comparing space x = (x1 , x2 , ..., xn ) we calculate the Euclidean distance
the methods, as it will be presented in Section VI of this work, with respect to the mean vector μ of the class [6].
it can be seen that the Mahalanobis method is more efficient
than the Euclidean distance and artificial neural networks n

because it has greater potential to be embedded and has better DE = (xi − μi )2 (1)
results. i=1

978-1-5090-6435-9/17/$31.00 2017
c IEEE 997 | P a g e
 Intelligent Systems Conference 2017
7-8 September 2017 | London, UK

Fig. 4. ANN model.

A. Materials and Methods

For the development of the prototype we used a sensor GY-
87, an IMU containing accelerometer, gyroscope and magne-
tometer, with all measurements in three orthonormal axes (x, y
and z).
This sensor was fixed on a headphone, so that the plane xy
is on the top of the user’s head, and x axis pointing forward.
Fig. 3. Roll, pitch and yaw.
Fig. 5 presents the prototype and five situations that classify
the motion controller in wheelchair perspective, classified as:
C. Mahalanobis Distance Classifier neutral, right, left, front and back, are represented respectively
by C1, C2 C3, C4 and C5.
A technique widely used for groups classification is close
to Euclidean distance methodology, however, consider the
probabilistic correlations between data sets, which equating
is determined by the covariance matrix S. (2) describes the
Mahalanobis distance.

DM (x) = (x − μ)T S −1 (x − μ) (2)

D. Artificial Neural Networks

The artificial neural networks use as a basis for operat-
ing the neural structure of intelligent organisms and acquire
knowledge through experience. The networks are composed of
neurons that communicate through synapses, which is a region
where two neurons come into contact and nerve impulses are
transmitted between them [8].
Then, an artificial neural network compose a configuration
with neurons, divided into three layers types: the input layer
neurons receive the initial information, the output layer sends
the final response of the proposed problem and the intermediate
layers has hidden layers that treat the data realizing the
learning of the network. The knowledge network is in the
synapses between the neurons, providing that the network
learn every new iteration. Fig. 4 presents the exemplification
of an artificial neural network [7].
Fig. 5. Acting classes.
IV. P ROJECT D EVELOPMENT
This section presents the methods, materials, treatment and In this project was used an Arduino UNO microcontroller,
the classifier project developed in this work. whose objective is data acquisition from the GY-87 sensor
978-1-5090-6435-9/17/$31.00 2017
c IEEE 998 | P a g e
 Intelligent Systems Conference 2017
7-8 September 2017 | London, UK

through the I2C protocol, initial treatment of bias signals and

transmission through serial to MatLab R
.
$FFHOHURPHWHU&RPSRQHQWV
MatLab R
performs pattern recognition algorithms for clas- 
sifier training, which can later be embedded in their own
microcontroller to directly drive an motorized wheelchair. 

B. Acquisition and Conditioning of Data


As each user has a specific positioning of the head and

$\J
there is an inaccuracy in the time to wear the prototype, every 
time the user should starts the system in neutral position, so
a bias calibration routine is performed, which in addition to
compensate the electronic offsets, calibrates the user’s position 
as the inertial frame. After this the variables are available to
be interpreted by the program in real time, they are: axis x 
and y of the accelerometer and axis x and y of the gyroscope.
Initially we tried to use only head inclination readings 
through the accelerometer, but was noted a mixture of un-       
desired classes result of the transition moments when the $[J
user performs a movement, therefore, was used two more & & & & &
dimensions from the Gyroscope readings, making possible *\URVFRSHFRPSRQHQWV
another discrimination through the angular velocities. 

The data set was developed with about 75000 events

through acquisitions of supervised sensors and rated by the 
users, performing hundreds of pre-determined movements. Fig.
6 shows the projection of 1,000 randomly selected events on 
the accelerometer and gyroscope plane, becoming apparent
*\UDGV

mixing between classes and regions of concentration of each 

class C.
TABLE I. P ERFORMANCE OF D ISTANCE M ETHODS 
Method Average Error Rate Average time
Euclidean Distance 11.53% 3.9 μs
Mahalanobis Distance 8.87% 6.4 μs 


V. R ESULTS AND D ISCUSSIONS         
*[UDGV
A. Classifier Project
The classifier was chosen after studying three different
methods based on data collected standardized between -1
and 1: Euclidean distance, Mahalanobis Distance and artificial Fig. 6. Data distribution.
neural networks. Linear classifiers were discarded because
the number of classes involved in the problem, which would
require the use of several lines to restrict them. mapped as shown in Table II. The comparison of results
Initially, it is necessary to analyze which method will be obtained using ANN is compared with the others and analyzed
used for the project, so starting with the Euclidean distance in greater detail in Section V-B.
method mainly because is simple to implement and analyze,
then do the analysis by the Mahalanobis distance in order to TABLE II. ANN C LASSIFICATION
determine which technique will be more efficient. Class C1 C2 C3 C4 C5
S1 -1 1 -1 -1 1
The study involved data division in a group of development S2 -1 1 1 1 -1
and testing in the proportion of 80% and 20% respectively, S3 -1 -1 1 -1 -1
each method has been performed with several different data
arrangements to establish efficiency. In order to obtain better results than distance methods, we
Firstly, we tested methods based on distance, the result studied the use of artificial neural networks, the performance
is shown in Table I, with mean values of error rate and was evaluated based on the number of neurons, and the chosen
computational cost for operation of an event. architecture was a feedforward, with a hidden layer with
radial basis activation function, the Fig. 7 shows the mean
Furthermore we analyze the project using artificial neural performance related to the number of neurons in relation to
networks, where the output vector [S1, S2, S3]T has been error rate of test groups.

978-1-5090-6435-9/17/$31.00 2017
c IEEE 999 | P a g e
 Intelligent Systems Conference 2017
7-8 September 2017 | London, UK

Fig. 9 shows a second simulation for a different map,

enhancing their maneuvers ability in the simulated space.









\>PP@
Fig. 7. Artificial neural networks architecture assessment. 


From 40 neurons, the artificial neural network have a better
performance than the Mahalanobis distance method, but the 
error tended to stabilize with the addition of neurons, resulting
in insignificant improvements. Due to cost-effective in relation

to performance and simplicity for future real application in em-          
bedded software, we decided to use the Mahalanobis method
for implementation. So, the ANN method was implemented, Fig. 9. Testing with mobile robotics simulator 02.
however was not used in the final application because did not
show great improvement comparing with Mahalanobis method.
VI. C ONCLUSIONS AND F UTURE W ORK
B. Simulations Based on the paper presented we can first infer that the
artificial neural networks method requires a lot of neurons to
To prove the controller performance, tests were per- a satisfactory answer, and for the case presented would need
formed in a mobile robotics simulation platform developed in an amount between 40 and 50 neurons to get a more precise
MatLab R
, where you can assign linear and angular velocities answer than the Mahalanobis distance method. It important
in a computer model of a differential robot, which is a category to highlight that values greater than 50 neurons show no
where the wheelchair fits. significant gain for the system.
Fig. 8 presents a simple path in a closed environment, Thus, the Mahalanobis distance is indicated as the most sat-
showing the maneuvering capability of the developed con- isfactory technique for the proposed work because it presents
troller. more accurate results than the Euclidean distance with a
small increase in complexity. Furthermore, the artificial neural
networks method requires a higher complexity to make a per-
formance close to the Mahalanobis method, making interesting
to use the Mahalanobis distance to be a method with less
complexity and thus more feasible to implement future practice
in a wheelchair embedded system.
So, the main goal was achieved as presented in the results.
The methods studied led to a way of possible implementation.
In this way, the proposal of a motorized wheelchair for
quadriplegic and low cost was possible to be implemented
at the simulation level. It is now intended to implement it
physically in order to construct a first model.
This work started a series of projects involving robotic
assistive, among the possible improvements we can highlight a
more complex control design for the developed controller using
a fuzzy system, where speeds can take continuous values in
a range. Another study will involve using a laser scan around
the chair to make a shared control preventing collisions and
improving the maneuvers accurate.

ACKNOWLEDGEMENTS
We would like to thank the Graduation Program in Electri-
cial Engineering and to the Inteligent Robotic Group of Federal
Fig. 8. Testing with mobile robotics simulator. University of Juiz de Fora.
978-1-5090-6435-9/17/$31.00 2017
c IEEE 1000 | P a g e
 Intelligent Systems Conference 2017
7-8 September 2017 | London, UK

R EFERENCES http://www.deficienteciente.com.br/censo-ibge-46-das-pessoas-com-
deficiencia-recebem-ate-1-salario.html, acesso em 26 de agosto de
[1] Kim, Keun-Tae, Heung-Il Suk, and Seong-Whan Lee, Commanding a 2016
Brain-Controlled Wheelchair using Steady-State Somatosensory Evoked
Potentials., 2016. [11] Flávia Villela, IBGE: 6,2% da população têm algum tipo de
deficiência, disponı́vel em http://www.ebc.com.br/noticias/2015/08/ibge-
[2] Makwana, Soniya D., and Anuradha G. Tandon, Touch screen based 62-da-populacao-tem-algum-tipo-de-deficiencia, acesso em 26 de agosto
wireless multifunctional wheelchair using ARM and PIC microcontroller., de 2016
2016, International Conference on Microelectronics, Computing and
Communications (MicroCom). IEEE, 2016. [12] Rohmer, Eric, et al, A novel platform supporting multiple control strate-
gies or assistive robots., Robot and Human Interactive Communication
[3] Rebsamen, Brice, et al, A brain controlled wheelchair to navigate (RO-MAN), 2015 24th IEEE International Symposium on. IEEE, 2015.
in familiar environments., IEEE Transactions on Neural Systems and
Rehabilitation Engineering 18.6, 2010: 590-598. [13] Olivi, Leonardo, et al, Shared control for assistive mobile robots based
on vector fields., URAI. 2013.
[4] Wu, Jian, Lu Sun, and Roozbeh Jafari, A Wearable System for Recog-
nizing American Sign Language in Real-time Using IMU and Surface [14] de Lima, Ana Karolina Pontes, Fisioterapia na lesão medular.
EMG Sensors. [15] Honeywell 3-Axis Digital Compass IC HMC5883L datasheet
[5] IRC’SPHERE, E. P. S. R. C., a Sensor Platform for HEalthcare in a [16] INVENSENSE, MPU-6000 and MPU-6050: Product Specification, Re-
Residential Environment. vision 3.3 Disponı́vel em: ¡http://www.gotronic.fr/pj-956.pdf¿ Acesso
[6] National Institute of Standards and Technology, Euclidian distance, em: 12 de agosto de 2015.
disponı́vel em ¡https://xlinux.nist.gov/dads//HTML/euclidndstnc.html¿, [17] MACIEL, G. M. Compensação de Posicionamento Angular de um
acesso em 30 de agosto de 2016 Manipulador Robótico Sobre uma Base com Três Graus de Liberdade.
[7] Kovács, Zsolt L, Redes neurais artificiais, Editora Livraria da Fisica, 2015. Trabalho de Conclusão de Curso - Universidade Federal de Juiz
2002 de Fora.
[8] Carvalho, APLF, Redes neurais artificiais, disponı́vel em http://www. [18] Sheikh, S. A. M. S. and Rotake, D. R., An evolutionary approach for
icmc. usp. br/andre/research/neural, 29, 05–09, 2005 smart wheelchair syste, Communications and Signal Processing (ICCSP),
2015 International Conference on IEEE, 2015, pp. 1811-1815.
[9] Wise Young, Ph.D., M.D., Spinal Cord Injury Levels and Classification,
disponı́vel em http://www.sci-info-pages.com/levels.html, acesso em 30 [19] Kuno, Y. a=et al., Robotic wheelchair with three control modes, in
de agosto de 2016 Robotics and Automation, 1999, Proceedings. 1999 IEEE International
Conference on, vol. 4. IEEE, 1999, pp. 2590-2595.
[10] Vera Garcia, Censo IBGE: 46% das pessoas com
deficiência recebem até 1 salário mı́nimo, disponı́vel em [20] Emotiv. The emotiv epoc interface. [Online]. Available:
http://www.emotiv.com/

978-1-5090-6435-9/17/$31.00 2017
c IEEE 1001 | P a g e