Design and Validation of Powered Knee
Orthosis System embedded with Wearable Sensors
Paulo Félix, Joana Figueiredo, and Cristina P. Santos
Center for MicroElectroMechanical Systems (CMEMS)
University of Minho
Juan C. Moreno
Neural Rehabilitation Group
Cajal Institute, Spanish National Research Council
Abstract— The development of new architectures for orthotic
devices has been playing a major role in the rehabilitation of gait
disorders. This paper proposes a new electronic and control
architecture for a powered orthosis, particularly, a knee orthosis.
The system was designed to be modular, being composed of the
orthosis and biomedical wearable sensors, such as inertial
measurement units, force sensitive resistors, and
electromyography. For each component, robust hardware and
software interfaces were designed and validated, plus two tracking
control strategies, namely, position control, that imposes a
trajectory based on the angle measured in the joint, and torque
control to act mostly as a passive component, using the measured
user-orthosis interaction torque. The whole system was validated
with healthy subjects walking in level-ground on a treadmill at
different speeds. The main results show that the system is
functional. The interfaces created as well as the assistive control
techniques were successfully validated. Moreover, the system
allows an efficient inclusion of other devices, given the modularity
achieved in its design.
Keywords— Gait Rehabilitation; Powered Knee Orthosis;
Werable Sensors; Real-time Control
In the last six decades, powered lower-extremity
exoskeletons and active orthoses have been widely addressed in
the fields of rehabilitation, assistive, and empowering devices
. Such devices have been developed with different types of
mechanical structure, actuators, interfaces, and assistive
strategies to increase the physical performance of the wearer and
to provide assistance to the motion, in a wide range of
applications , . Regarding the active lower limb orthoses,
these devices act in parallel with the human limb, being mostly
designed to assist subjects with lower limb pathologies .
Several orthotic devices have been developed for specific
application to the knee (e.g., climbing stairs , squatting –
; stand-to-sit and sit-to-stand tasks , , , and running
). In general, they rely on electric actuators, such as DC ,
– and AC servo  motors, series elastic actuators ,
, and pneumatic actuators , . Additionally, these
systems are embedded with sensors of several kinds to measure
system variables fundamental in the assistive techniques and
evaluation of the user’s performance, e.g., encoders , , ,
, potentiometers , , Hall effect sensors , and
Inertial Measurements Units (IMUs) ,  for angle and
velocity determination and gait phase estimation; load cells ,
, foot-switches and Force Sensitive Resistors (FSRs) ,
 for ground contact detection; force sensors ,  and
motor current-measurement  for motor torque calculation;
and electromyography (EMG) surface electrodes ,  for
muscle activity measurements. Furthermore, distinct types of
assistive strategies have been applied (e.g. model-based control
, , , predefined gait trajectory control , , ,
and predefined action based on gait pattern , ) to assist the
users in a set of activities and therapies.
The primary aim of this work compromises the development
of a new electronic and real-time control architecture for a
Powered Knee Orthosis (PKO). The system is formed by an
orthotic device and external wearable sensors, such as IMUs,
FSRs and EMG modules, enabling the development of smart
rehabilitation tools and motion assistive techniques. This paper
discloses modular hardware and software interfaces, as well two
tracking control strategies (position- and torque-based trajectory
control) for assistance in ground-level walking. Moreover, an
adaptive gait event detector is also included for gait pattern
analysis, using the data recorded from one gyroscope mounted
in the instep of the foot. The detector stands out by identifying
six human gait events, i.e. Heel Strike (HS), Flat Foot (FF),
Middle Mid-Stance (MMST), Heel-Off (HO), Toe-Off (TO),
and Middle Mid-Swing (MMSW) with high detection rate.
Concerning the validation of the whole system, trials with
healthy subjects in level-ground walking were conducted, for
both assistive strategies at different speeds. Overall, the major
contribution outlined in this papers focuses on the technical
description, design and validation of a novel modular
architecture for a PKO and its embedded sensory system, which
stands out by the ability of monitor the assisted human gait in
terms of kinematic, kinetic and physiologic parameters.
II. SYSTEM DESIGN
A. System Overview
The presented hard real-time system is centered on a
microcontroller (MCU) connected to an actuation device (PKO)
and wearable sensors (IMUs, FSRs, and EMGs). As presented
in Fig. 1, the PKO has embedded an electronic actuator (DC
brushless motor) and sensors (e.g., potentiometer and strain
gauges), all directly used in control techniques. Moreover, the
system is equipped with external sensors to supplement the
information provided by the PKO. Their combination with the
orthotic device provides a way for a more complex set-up that
goes behind the orthosis domain, and extends to the analysis
This work was supported by Fundação para a Ciência e Tecnologia (FCT)
with the scholarship references SFRH/BD/108309/2015 and
SFRH/BD/102659/2014, and on th e scope of project LIACC with reference
PEstC/EEI/UI0027/2015; by Fundo Europeu de Desenvolvimento Regional
(FEDER); by Programa Operacional Factores de Competitividade (POFC) –
COMPETE. Also, this work was partially funded by FCT with the reference
project UID/EEA/04436/2013, and by FEDER funds through the COMPETE
2020 with the reference project POCI-01-0145-FEDER-006941, and by grant
RYC-2014-16613 by Spanish Ministry of Economy and Competitiveness.
978-1-5090-6234-8/17/$31.00 ©2017 European Union
and monitoring of the gait pattern (e.g., foot plantar pressure
analysis, gait segmentation, and intention recognition), and
evaluation of the user performing distinct tasks (e.g., evaluation
of disability level, effort and progression among therapies).
The proposed system has been designed under a modular
approach, to allow for further inclusion of other wearable
sensors and orthoses. Although a functional version is only
implemented to the right leg, its design intends to target both
legs similarly. Heretofore, one PKO is mounted in the right
knee joint, one IMU was mounted on the instep of the right foot,
two FSR were placed in the right heel and toe, and EMG
electrodes (six channels) were attached to the surface of the
main lower limb muscles involved in the joint movements.
Moreover, these devices present distinct interfaces (digital and
analog) with the MCU, as illustrated in Fig. 1. The PKO and the
IMU are prepared to communicate with the same digital
protocol, i.e. a Control Area Network (CAN), while the FSRs
and EMGs have standard analog interfaces for the acquisition
of their output signals, i.e. Analog-to-Digital Converters
(ADCs) available in the MCU. The system stands for using
CAN, given its strict determinism, data collision avoidance,
optimized data transfer, and multiple-access points that allow
new devices to be easily connected to the physical layer .
B. Components and Interfaces
1) Processing Unit and Communication Interfaces
The choice of the processing unit was made regarding the
main requirements for the implementation of the system in real-
time. In general, the system must be fast and resourceful enough
to run advanced assistive motion algorithms and gait analysis
methods, flexible enough to allow easy inclusion of sensors and
actuation systems through Input/Output (I/O) analog interfaces
and/or digital protocols, and portable to provide easy mounting
on users. Thus, we selected the STM32F4-Discovery board
(STMicroelectronics), which is centralized on the
STM2F407VGT high performance MCU with an AMR®
Cortex® -M4 32-bit core, running at 168 MHz. This MCU
meets the proposed requirements, given its key features, such
as the high-speed embedded memories (flash memory up to 1
Mbyte and SRAM up to 192 Kbytes), and the extensive range
of enhanced I/Os and peripherals, with standard and advanced
communication interfaces. The present application takes
advantage of two 12-bit ADCs with 16 channels (ADC1 for
FSRs, and ADC2 for EMG channels), two CAN controllers
(CAN1 for orthoses, and CAN2 for IMUs), and three general-
purposed 32-bit timers (to trigger the acquisition of the sensors
and run the real-time control loop).
For this application, the ADC peripherals were configured
in scan mode (automatic conversion performed simultaneously
on a group) triggered by the overflow of timers. The acquisition
circuits that proceed the sensors were tuned to give an output in
the range 0-3.3V (voltage references of ADC). On the other
hand, two CAN buses were separately created, to establish
communication with the orthosis (CAN1) and IMUs (CAN2),
as shown in Fig. 1. This configuration allows future expansion
to include more active orthoses (up to six) and inertial sensors
(up to sixteen). The CAN controllers incorporated in the
SMT32 MCU are implemented in hardware, therefore, their use
does not bring any additional cost at the software level. Also,
this strategy minimizes scenarios where the bus might be busy,
providing more determinism in the control of both devices.
2) Powered Right Knee Orthosis
The orthosis consists in a modular joint (H2-Joint) from the
lower limb robotic H2-exoskeleton (Technaid S.L., Spain),
developed for gait rehabilitation in stroke survivors . A
technical description of the device is presented in Table 1.
TABLE I. H2-JOINT COMPONENTS
- 10 kΩ and linearity of ± 0.25%.
Coupled to a toothed pulley and belt to
transmit joint’s motion.
- From 0 to 100 degrees.
- Four strain gauges connected in a full
Wheatstone bridge (enhances accuracy
and sensitivity to temperature).
- From -50 to +50 Nm.
Fig. 1. System’s overview, illustrating the main components and interfaces between them.
- Measurement of motor’s angular speed
Nominal voltage of 24V, torque of 221
mNm and current of 4.23 A.
Gear ration of 160:1.
Continues net torque of 34 Nm and peak
torque of 180 Nm.
Coupled to motor.
- 64 MHz MCU (DsPIC0F4011,
Power management module.
Measurements of motor’s current (A).
MOSFET drive module.
CAN communication transceiver.
- Sensor’s data acquisition.
H2-Joint contains an embedded electronic board
responsible for the communication with external devices,
through CAN. To establish the communication between the
MCU and the board (as well with the IMUs), we created a
circuit based on the CAN transceivers SN65HVD251 (Texas
Instruments, USA). This circuit offers the capability of
transmission and reception between the CAN controller and
CAN bus. The H2-Joint board sends CAN packages with data
collected from the sensors at 1 KHz, and receives packages with
the output command of the controller, in the format presented
in Fig. 2. Also, each joint has an identifier (CAN ID), that is
included in the beginning of the CAN package.
Fig. 2. Representation of exchanged packages between H2-Joint and MCU. ID
is the identifier of the target. DLC counts the bytes of valid information. The
sensors information received in the MCU is organized by angular position (T),
angular velocity (Z), interaction torque (Winter) and motor torque (Wm).
3) IMU device
For measurements of foot kinematic data, we chose the
wearable Tech IMU v4 (Technaid S.L., Spain). This unit
integrates three different tri-dimensional MEMS (micro-
electromechanical systems) sensors, including an
accelerometer (range: ±16 g), a gyroscope (range: ±34.9 rad/s)
and a magnetometer (range: ±8.1 G), and a built-in calibration
which eliminates axes misalignment, sensibility and
compensates the measurements due to temperature variations.
Overall all, this device constitutes an optimal solution given its
small dimensions (11x26x36), weight (10 g) admissible power
consumption (70 mA) and built-in calibration.
The digital acquisition is performed following the protocol
defined by proprietary. Fig. 3 shows a sequential diagram
describing the protocol. As exemplified, the MCU starts the
communication by sending a ‘P’ command (one byte of
information), asking for physical data, ensuring a new
calibration of the device. After, the MCU sends ‘polling’
commands (zero bytes of information) at a given sampling
frequency to collect the data from all sensors. Each response to
the polling commands is composed of five packages of eight
bytes (sent sequentially to the MCU), containing the
information of the axis, represented as a 32-bit float (four
bytes). The chosen sampling frequency was 100 Hz.
Fig. 3. Digital protocol between the MCU and IMU, in a sequential diagram.
4) EMG Module
The EMG module aims the acquisition of electrical activity
of lower limb muscles, according to their effect on the human
joints movement (e.g., hamstrings and quadriceps femoris for
the knee, and tibialis anterior and gastrocnemius for the ankle).
Thus, we selected the MA-420 EMG preamplifier (Motion Lab
Systems, USA). As key features, this device incorporates radio-
frequency interference filters, electrostatic discharge protection
circuitry, a low-impedance output to eliminate cable noise and
cable motion artifacts, and an integral ground reference that
provides immunity to electromagnetic environmental noise,
constituting a reliable solution.
The EMG module is composed by hardware and software
interfaces, represented in Fig. 4. The hardware EMG interface
consists of one board with six channels. For each channel, the
same circuit was designed for proper signal conditioning.
Fig. 4. Hardware and software interfaces for one EMG channel, to process and
convert the analog signal to the respective digital format.
The first stage compromises the signal pre-amplification.
The device is designed to be used with disposable electrodes,
placed on the surface of the skin. The electrical signal is then
amplified with the gain of 20 ±1 (at 1kHz), being the output
voltage in a range of ±0.4 mV to ±40 mV. Additionally, the pre-
amplifier circuit has a high Common-Mode Rejection Ration
(CMRR) of 100 dB, meaning that a great percentage of
common mode voltage is eliminated, maximizing the Signal-
Noise Ratio (SNR). The next state represents an Anti-Aliasing
Low-Pass filter, implemented to avoid aliasing. Since the full
bandwidth of an EMG signal is up to 500 Hz, the cut-off
frequency must be set to this value, eliminating frequencies
outside this band. The last hardware stage has three main
functions, concerning the posterior acquisition by the ADC:
amplification; level-shifting; and voltage limitation of the
signal. The two first stages are performed by means of a
summing amplifier circuit. The same circuit amplifies the signal
with gain in a range of 90 to 260 (selectable gain with high
precision potentiometer) summing a constant voltage
representing half of the high voltage reference of the ADC,
particularly, +1.65 V (VREF+ = +3.3V). The selectable gain
feature was added to allow measurement of the electric activity
of different target muscles, since they can present distinct
amplitudes. At last, a limiter voltage circuit (buffer) is used to
prevent the output signal to exceed the VREF minimum (VREF-
= 0V) and maximum (VREF+ = +3.3V).
With respect to the software created for the acquisition of
the signal, the ADC was programmed to collect the data at 2
kHz (respecting the Nyquist theorem), with 12-bits, and with a
dynamic range of 0.806 mV/bit.
5) FSR modules
The FSR sensors used to measure ground reaction forces in
the foot correspond to the model 406 FSR (Interlink
Electronics). They consist of robust polymer thick film sensors
that exhibit a decrease in resistance when the force applied to
the surface increases. Also, this sensor stands for its high
repeatability (± 2%), cost-effectiveness, and simplicity of use.
Regarding the developed hardware, a simple voltage divider
circuit was designed, having low voltage (near 0V) at the output
if no force is applied and high voltage (around 3.3V) when more
than 10 N (sensitivity range) are measured. Concerning the
software developed for the signal acquisition, it was used the
same strategy implemented in the EMG module, at a sampling
frequency of 100 Hz.
C. Orthosis’ Use
Fig.5 shows one user wearing the proposed system. The
orthotic device was fixed on the right lower limb in four points
with straps: two in the upper limb and two in the lower limb.
Each time the user wears the system, a careful procedure is
made to align the mechanical joint with the human knee joint,
to minimize the loss of mechanical power. Also, the location of
the braces can be adjusted according to the user’s lower limb
length, allowing the device to be wearable and functional for
other users. Also, Fig. 5 discloses how the IMU and FSRs were
mounted in the foot.
Fig. 5. Proposed system (PKO and wearable sensors) mounted in one subject.
D. Tracking Control Strategies
Two control strategies were developed for the first set-up of
the PKO system: position-based trajectory control, which
corresponds to the classic position control and is based on the
difference between the desired angular position (θref) and real
angular position (θm); and torque-based trajectory control,
which is based on the difference between the desired torque
(τref) and the interaction torque between the limb and the
orthosis (τinter). Position control can be used in therapies that
ensure repetitive movements of the user’s limbs, suitable to
improve muscular strength and movement coordination in
neurological patients, such as patients with hemiparesis .
Torque control, in a scenario where the torque reference torque
is zero, composes a strategy that minimizes the mechanical
impedance of the joint, allowing the orthosis to behave as a
passive actuator. This approach allows the controller to actuate
at the joint in a way that the user should feel more freedom to
move accordingly with his/her intentions. For instance, this
controller can be used in learning mode, where the trajectories
and interaction torque are recorded, to be posteriorly applied
actively in other strategies.
The real-time control runs on the MCU, at a frequency of 1
KHz. The PKO sensor’s data are read asynchronously, and PID
commands are sent through the CAN bus, at this frequency.
Both control diagrams are presented in Fig. 6.
Fig. 6. Position (above) and torque (below) control schemes.
The controllers implemented are based on a Proportional-
Integral-Derivative (PID) control. The equation that describes
the digital controller generated is presented in Eq. 1.
To find the gains of the controller (Kp, Ki, and Kd), the
Ziegler-Nichols method was used. The correct tuning of these
values must result in a compliant motion, without oscillation in
the trajectory, overshoot response, and instability.
E. Gait Events Detection
As mentioned, the system also incorporates a gait event
detection tool, based on the information recorded from the
MEMS gyroscope, mounted on the instep of the foot (Fig. 7).
The angular velocity from the axis aligned with the sagittal
plane was recorded and computed, to detect six gait events: HS,
FF, MMST, HO, TO and MMSW.
Fig. 7. Gait human events (above) segmented throughout the angular velocity
recorded form the gyroscope over one gait cycle .
The proposed method for gait segmentation was based on a
finite state machine with decision rules and adaptive thresholds.
A detailed description of the algorithm and its validation on
healthy subjects are presented by Félix et. al. in .
As preliminary safety measures, some features were added
to the orthotic system to prevent damage or unreliable
movements in the users. Firstly, the range of motion of the PKO
was limited in software to 3 – 98 degrees. This prevents the
device to damage the human legs by applying overextension or
over flexion movements. Additionally, this strategy avoids
stress on the mechanical limits of the joint. Moreover, unstable
and abrupt movements of the joint are avoided by the correct
tune of the controller’s parameters and by limiting the PID
commands. Finally, another safety concern is the alignment of
the joints, that prevents undesired movements.
To validate the whole system, simple trials were conducted
with 5 healthy subjects (3 males and 2 females), with age of
26.80 ± 2.78 years old, height of 1.68 ± 0.07 m, and weight of
64.60 ± 8.5 Kg. The participants were asked to walk in level-
ground (in a treadmill), for different speed (1.0 km/h to 1.8
km/h), with the two assistive strategies proposed: position and
torque control. Simultaneously, data from the IMU and FSRs
were recorded from the foot. Furthermore, the EMG signal from
the tibialis anterior muscle was recorded, for the minimum (90),
medium (175) and maximum (260) gains.
III. RESULTS AND DISCUSSION
One of the goals of this work compromises the validation of
the designed electronic and control architecture. Its
achievement goes through the validation of the system modules
and the tracking control strategies.
A. EMG board
The EMG board proposed was projected with selectable
gains, tunable for the lower limb muscles of each user. During
the trials, the subjects were asked to walk in a treadmill while
the EMG of the tibialis anterior muscle was recorded for
different gains (Fig. 8).
Fig. 8. EMG signal recorded from three trials at 1km/h, with distinct gains
(minimum, medium, and maximum).
Fig. 8 shows that the EMG signal recorded (1 km/h) when
the gain is set to maximum presents better quality in the stance
phase (regions with maximum amplitude) when compared with
the medium and minimum. This can also be inferred with the
SNR obtained for each plot, -22.38 dB, -20.90 dB, and -19.49
dB, respectively, that shows a lower value (less noise) when the
gain is higher. Also, the signal never saturates for this walking
speed, such that there is no loss of information. A similar
procedure with the same muscle was made to validate the other
channels of the EMG board.
B. FSRs and IMU
The FSRs and IMU were both mounted on the foot (see Fig.
5). Fig 8 shows the data collected in one trial (1.5 km/h).
Fig. 8. FSRs and gyroscope signals recorded from foot, walking at 1.5km/h.
Through the data acquired, Fig. 8 shows the well-
functioning of the interfaces for the FSRs and IMU. As
illustrated, the output of the FSRs (heel in green and toe)
follows the detection of the human gait events (in black)
throughout the gyroscope signal (in blue).
C. Tracking Control Strategies
The two tracking control strategies were also validated
during the trials. Regarding the position control (Fig. 9), a
reference trajectory (blue line) was imposed to the user.
Although a delay between the control variables (reference and
real position) is observed, this approach shows good results
considering that the movement of the user’s limb is performed
smoothly. During the PID tuning, it was noticed that higher
values of the PID gains provoke abrupt movements of the joint,
which can cause discomfort or instability to the user.
Fig. 9. Output signal from position control, walking at 1 km/h.
In the second approach (Fig. 10), the reference torque (black
line) was set to zero. As expected, this allowed the user to move
with freedom in the direction of the interaction force (red line)
measured, with low resistance offered by the orthosis. Thus, the
participants were able to perform distinct trajectories (cyan
line) during walking. Comparing Fig. 9 and Fig. 10, the real
knee angle measured in the joint have a similar shape, although
the position control shows a constant pattern in all steps.
Fig. 10. Output signal from torque control, walking at 1.8 km/h.
The development of a new electronic and control
architecture for a powered knee orthosis was presented and
validated in this paper. Each component of the system and
interfaces were tested, with healthy subjects walking in a
treadmill and wearing the orthosis and sensors. Overall, the first
set-up of the system is functional, and ready to be used in
motion assistance therapies. Future work compromises the
inclusion of another powered orthosis, i.e., an ankle orthosis
(another module of H2-exoskeleton) and the replication and
tuning of the components, interfaces, and same control
strategies for this joint. Also, new assistive strategies will be
explored, aiming the development of a compliant actuation for
application in gait rehabilitation interventions of neurologically
 S. Viteckova, P. Kutilek, and M. Jirina, “Wearable lower limb robotics: A
review,” Biocybern. Biomed. Eng., vol. 33, no. 2, pp. 96–105, 2013.
 T. Yan, M. Cempini, C. M. Oddo, and N. Vitiello, “Review of assistive
strategies in powered lower-limb orthoses and exoskeletons,” Rob. Auton.
Syst., vol. 64, pp. 120–136, 2015.
 H. Herr, “Exoskeletons and orthoses: classification, design challenges and
future directions.,” J. Neuroeng. Rehabil., vol. 6, p. 21, 2009.
 J. E. Pratt, B. T. Krupp, C. J. Morse, and S. H. Collins, “The RoboKnee:
an exoskeleton for enhancing strength and endurance during walking,”
IEEE Int. Conf. Robot. Autom. 2004. Proceedings. ICRA ’04. 2004, vol.
3, no. April, pp. 2430–2435, 2004.
 K. Kim, C. H. Yu, G. Y. Jeong, M. Heo, and T. K. Kwon, “Analysis of
the assistance characteristics for the knee extension motion of knee
orthosis using muscular stiffness force feedback,” J. Mech. Sci. Technol.,
vol. 27, no. 10, pp. 3161–3169, 2013.
 A. Gams, T. Petric, T. Debevec, and J. Babic, “Effects of robotic knee-
exoskeleton on human energy expenditure.,” IEEE Trans. Biomed. Eng.,
vol. 60, no. c, pp. 1–9, 2013.
 N. Karavas, A. Ajoudani, N. Tsagarakis, J. Saglia, A. B icchi, and D.
Caldwell, “Tele-Impedance based stiffness and motion augmentation for
a knee exoskeleton device,” Proc. - IEEE Int. Conf. Robot. Autom., pp.
 A. N. Spring, J. Kofman, and E. D. Lemaire, “Design and Evaluation of
an orthotic knee extension assist,” J. Spacecr. Rockets, vol. 41, no. 6, pp.
 A. M. Dollar and H. Herr, “Design of a quasi-passive knee exoskeleton to
assist running,” 2008 IEEE/RSJ Int. Conf. Intell. Robot. Syst. IROS, pp.
 C. Fleischer and G. Hommel, “A Human – Exoskeleton Interface Utilizing
Electromyography,” vol. 24, no. 4, pp. 872–882, 2008.
 W. Y. Lai, H. Ma, W. H. Liao, D. T. P. Fong, and K. M. Chan, “HIP-
KNEE control for gait assistance with Powered Knee Orthosis,” 2013
IEEE Int. Conf. Robot. Biomimetics, ROBIO 2013, no. December, pp.
 M. Arazpour, a Chitsazan, M. a Bani, G. Rouhi, F. T. Ghomshe, and S.
W. Hutchins, “The effect of a knee ankle foot orthosis incorporating an
active knee mechanism on gait of a person with poliomyelitis,” Prosthet
Orthot Int, vol. 37, no. 5, pp. 411–414, 2013.
 G. Aguirre-Ollinger, J. E . Colgate, M. A. Peshki n, and A. Goswami,
“Inertia compensation control of a one-degree-of-freedom exoskeleton for
lower-limb assistance: Initial experiments,” IEEE Trans. Neural Syst.
Rehabil. Eng., vol. 20, no. 1, pp. 68–77, 2012.
 S. Corrigan, “Introduction to the controller area network (CAN),” 2002.
 M. Bortole et al., “The H2 robotic exoskeleton for gait rehabilitation after
stroke: early findings from a clinical study,” J. Neuroeng. Rehabil., vol.
12, no. 1, p. 54, 2015.
 P. Félix, J. Figueiredo, C. P. Santos, and J. C. Moreno, “Adaptive real-
time tool for human gait event detection using a wearable gyroscope,”
CLAWAR, 2017, pp. 1–8, unpublished.