A Telerehabilitation System based on Wireless Motion Capture

Sensors

Pedro Macedo

1,2

, José A. Afonso

, Luis A. Rocha

and Ricardo Simoes

2,3,4

Department of Industrial Electronics, University of Minho, Guimarães, Portugal

Institute for Polymers and Composites IPC/I3N, University of Minho, Guimarães, Portugal

Life and Health Sciences Research Institute (ICVS/3Bs), University of Minho, Braga, Portugal

School of Technology, Polytechnic Institute of Cávado and Ave, Barcelos, Portugal

Keywords: Wireless Sensor Networks, Motion Capture, Wearable Sensors, Home-based Rehabilitation.

Abstract: The constant growth of the elderly population in the world creates new challenges and opportunities in

health care systems. New technological solutions have to be found in order to meet the needs and demands

of our aging society. The welfare and quality of life of the elderly population must be a priority. Continuous

physical activity will play an important role, due to the increase of the retirement age. However,

physiotherapy can be expensive, even when the desire movements are autonomous and simple, also requires

people to move to rehabilitation centres. Within this context, this paper describes the development and

preliminary tests of a wireless sensor network, based on wearable inertial and magnetic sensors, applied to

the capture of human motion. This will enable a personalized home-based rehabilitation system for the

elderly or people in remote physical locations.

1 INTRODUCTION

Developed societies nowadays face an emergent and

possibly problematic situation, if not addressed

properly: the growth ratio of elderly population

compared with the total population. This continuous

growth is changing the demographic structure of

societies, with lower birth rates and increasing life

expectancy (Stula, 2012). This demographic change

(Linz and Stula, 2012) imposes new challenges in

order to create services and products for ambient

assisted living (AAL). These products will enable

the creation of better life conditions for the older

generation in their environment by increasing their

self-confidence, autonomy and mobility (Sun et al.,

2009) (Fuchsberger, 2008) (Kleinberger et al.,

2007).

The monitoring of human body movements,

body kinematics, is a growing research field in areas

such as health care, entertainment and sports.

Nevertheless when creating a wireless sensor

network (WSN) to monitor human body motion,

several factors must be taken into account in order to

assure a reliable operation. The authors in (Hadjidj

et al., 2012) describe the technical challenges faced

when creating a WSN for rehabilitation applications

and review existing projects. Several challenges

need to be addressed on these systems, such as the

fixation of the sensor node, energy efficiency

(normally the sensors are energy constrained), the

usually high amount of generated data per sensor

(high sampling rate) and the impact of the human

body on the signal propagation (the wireless signal

propagation suffers from diffraction around the body

and reflection from the body). Human motion

tracking systems are usually classified as either

visual, by the combination of data from several

cameras to achieve a 3D location of the patient’s

body and limbs (Moeslund and Granum, 2001);

(Wang et al., 2003), or non-visual, such as the use of

wearable sensors.

This paper focuses on the non-visual method,

and describes the development of a wireless sensor

network, based on wearable inertial and magnetic

sensors. When compared to a camera strategy, this

method has the advantages of higher flexibility and

mobility, as it can be used in uncontrolled

environments, without lighting and line of sight

concerns (Aminian and Najafi, 2004).

The capabilities and limitations of this kind of

electronic sensors, when applied to the analysis of

the human posture and movements, are highlighted

Macedo P., Afonso J., Rocha L. and Simoes R..

A Telerehabilitation System based on Wireless Motion Capture Sensors.

DOI: 10.5220/0004873800550062

In Proceedings of the International Conference on Physiological Computing Systems (PhyCS-2014), pages 55-62

ISBN: 978-989-758-006-2

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

in (Wong et al., 2007). Several authors have also

underlined the high performance of inertial and

magnetic sensors, when applied to body kinematics

measurement. Accelerometers in (Farella et al.,

2008) are used to exploit their capabilities in terms

of body posture recognition in dynamic everyday

life activities. A protocol to measure the upper limb

kinematics using the Xsens MT9B technology was

developed in (Cutti et al., 2007), with the purpose of

measuring its accuracy (through the protocol) during

movements of clinical relevance. The benefits of

physical activity are exposed in (Steffen et al.,

2011), where the authors propose a personalized,

home-based exercise trainer for elderly people. The

captured movements are evaluated and compared to

a prescribed exercise; the purpose is to assist the

user to correctly perform the proposed movements.

The purpose of this paper is to present the

current status of a wireless posture monitoring

system, based on wearable inertial and magnetic

sensors, applied to rehabilitation.

This paper is organized as follows. The next

section presents an overview of all components of

the implemented system. In the third section, the

method of calculation of the orientation angles is

presented. Experimental results concerning a

rehabilitation session are presented in section 4.

Section 5 tackles future work and presents some

preliminary conclusions.

2 SYSTEM OVERVIEW

The electronic system of this posture monitoring

WSN is composed by a base station connected to a

PC and multiple nodes that collect sensory

information from the user. Each sensor module is

firmly attached to one body segment which we want

to monitor. The readings from the inertial and

magnetic sensors are continuously collected and sent

wirelessly to the base station, which forwards them

to the PC through a serial connection. The

information processing in the PC allows the

calculation, in real-time, of the 3D orientation of the

module, expressed by the pitch, roll and yaw angles.

The structure of this system is represented in Figure

Communication between the base station and the

sensor nodes is made using the CC2530, from Texas

Instruments (Texas Instruments, 2009). The CC2530

is a true system-on-chip (SoC) solution for

IEEE 802.15.4 applications (IEEE Std 802.15.4,

2006) which integrates an 8051 based

microcontroller and an 802.15.4 transceiver working

in the licence-free 2.4 GHz frequency band.

Figure 1: System components.

The Enhanced Low Power Real Time (eLPRT)

protocol (Afonso et al., 2011), a MAC protocol that

was designed to optimize the quality of service

(QoS) provisioning and the bandwidth utilization

efficiency, is used to control the communication

through the wireless medium.

2.1 Base Station Architecture

The base station is responsible for coordinating

operation of the wireless sensor network. Its

responsibilities include: associate new sensor nodes

to the network; allocate time slots on the eLPRT

superframe for nodes to transmit; and keep the

synchronization in the network, with the help of the

periodic transmitted beacons.

The base station is powered by the PC, so its

energy consumption is not a concern. The

architecture of the base station is shown in Figure 2.

The USART (Universal Synchronous Asynchronous

Receiver Transmitter) and Radio peripherals are

controlled by the CC2530 with the assistance of the

DMA (Direct Memory Access) subsystem.

Figure 2: Base station architecture.

Figure 3 presents the hardware of the base

station, which is composed by a CC2530EM

(Evaluation Module) attached to a SmartRF05EB

(Evaluation Board), both from Texas Instruments.

PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems

Figure 3: Base station.

2.2 Sensor Node Architecture

The data messages between each sensor node and

the base station are transmitted periodically

according to the superframe period set by the user’s

application. The number of sensor readings in each

period is also set by the user. These multiple

readings are aggregated into a single message and

sent to the base station. Currently a superframe

period of 100 milliseconds and 3 sensor readings is

in use. This gives us a rate of 30 frames per second,

a typical value for motion capture systems (Claypool

et al., 2006).

The sensor board contains 3-axis inertial and

magnetic sensors, which measure variations in the

gravitational and magnetic field forces, respectively.

It is due to these variations that it is possible to

calculate the orientation of the sensor node. The

sensor node architecture is shown in Figure 4.

Figure 4: Sensor node architecture.

The sensory information is collected through the

SPI bus (serial peripheral interface) and sent to the

base station using the eLPRT protocol. Detailed

information about the sensors present in the next

section.

2.2.1 Sensor Node Prototype

The sensor board prototype is illustrated in Figure 5,

with the bottom layer on the left side and the top

layer on the right side. The sensor board is

connected to a CC2530EM module through two

20-pin header connectors.

Figure 5: Sensor board.

There are two sensors present in the sensor

board. The MPU-6000 Motion Processing Unit from

InvenSense, which has an embedded 3-axis MEMS

gyroscope (not used at the moment), a 3-axis MEMS

accelerometer and a digital motion processor (DMP)

hardware accelerator engine with an auxiliary I

port that interfaces to a third party digital sensor,

such as a magnetometer. The embedded 16-bit

accelerometer has a digital-output triple-axis user-

programmable full-scale range of ±2g, ±4g, ±8g and

16g, a selectable output data rate from 4 Hz to 1 kHz

and a sensitivity of 16.384 LSB/g at ±2g.

The auxiliary I

C bus allows the MPU-6000 to

directly access the data registers of an external

digital sensor. In the case of this prototype, the

Honeywell 3-axis Digital Compass IC HMC5883L

is used. Designed for low-field magnetic sensing, it

contains a 12-bit ADC that enables 1 to 2 degrees

compass heading accuracy; with a resolution of

5 milliGauss in ±8 gauss fields.

2.3 PC Software

In the proposed system, the component with the

greatest relevance to the user is the PC application.

This application mediates between the user and the

base station; it can send commands and receive

information sent by the sensor nodes attached to the

body segments. After receiving the data from the

sensors, it calculates the angles of rotation and then

ATelerehabilitationSystembasedonWirelessMotionCaptureSensors

presents the movement of the user’s body in

real-time on a 3D model of the human body. The

software provides means to create rehabilitation

session files, where user information and motion

angles are stored. These session files can later be

employed to evaluate the user’s progress. The

referred application, Figure 6, was created in the

object-oriented programming language Java, due to

its portability between operating systems.

Figure 6: Application interface.

One sensor node per monitored segment is

necessary to measure the movement, with three

degrees of freedom (DOF), expressed by the Euler

angles pitch, roll and yaw. The complete model

consists on fifteen rigid body segments: pelvis,

torso, head, upper arms, forearms, hands, upper legs,

lower legs and foots; connected by virtual joints.

The implementation of this model was made

using the Java OpenGL (JOGL), which is a wrapper

library that allows OpenGL to be used within the

Java environment.

The system was designed to be as intuitive and

easy to handle by the user as possible. There are two

main components to take into account, in order to

have the system in full operation mode: the

configuration of the network (configuration of base

station and sensor nodes), started automatically by

the program when the Start button is pressed; and

the calibration of the sensor nodes, that is, the

process of finding orientation and placement of the

multiple sensor nodes attached to body segments.

These components are described in the next section.

2.3.1 Configuration and Calibration

Within the application folder, there are two

configuration files of utmost importance:

 Configuration File: Herein lay the configuration

parameters of sensor registers present in the

sensor board. Although there is a default

configuration stored in the CC2530

microcontroller, which is then loaded to the

sensors, the configuration parameters can be

altered, maybe not by the end user, but by a

medical specialist (physiotherapist) at the

beginning of the treatment.

 Calibration File: This calibration file

corresponds to maximum and minimum values

of the accelerometer and magnetometer readings.

These values are compared to the actual readings

of the sensors, for normalization, producing

outputs in the range of -1 to 1, as in ‘g-forces (in

the case of the accelerometer, this happens when

no external force besides gravity is present). One

calibration file for each sensor node in the

network is required, since not all the sensors give

the same readings. This file is generated by a

different application. For a better functionality of

the system, this calibration should be done at the

operation location.

2.3.2 Program Execution Steps

In the application’s main window, shown in

Figure 6, there is a Start button that, when pressed,

initiates the configuration of the system. The

configuration is divided in the following steps:

 Serial Port Detection: The connection between

the PC and the base station is made through a

serial port interface, so at this point is

fundamental to determine which COM port is

associated to the base station. To achieve this,

the application starts by discovering all the COM

ports available in the PC, and then sends a

command through them. The COM port which

answers according to what is expected is the port

connected to the base station. This process

enables the user to use any available port in the

PC, which is important with equipment such as a

laptop, where USB to serial converters are used.

 Open Configuration File: Opening the

configuration file and saving its parameter values

(sensors registers) on the application. This file is

common to all the sensor nodes in the network.

The superframe period and number of samples

per period are also configured here. These two

parameters combined give us the sampling rate

of the sensor readings.

 Configure Superframe Period in the Base

Station: A command containing the superframe

period is sent to the base station. Immediately

after receiving this value, it starts sending

beacons at that interval to the network. These

beacons allow the synchronization between the

PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems

sensor nodes and the base station.

 Turn on the Sensor Nodes: A notification

message appears at the application’s main

window, alerting that the sensor modules can be

turned on by the user. When activated, the sensor

node will become associated to the base station.

After that, the base station allocates a time slot in

the superframe, according to the eLPRT

protocol, for the sensor to transmit the data.

 Acquire Associated Nodes Information: A

command message is sent to the base station,

enquiring for the number of associated sensor

nodes and their respective physical address.

 Configure Sensor Nodes: After the discovery of

the physical addresses of all the nodes in the

network, the application sends the configuration

of the sensors registers, superframe period and

number of samples to each sensor node (through

the base station).

 Data Acquisition: Since the sensor modules are

now configured, the application sends a start

command message to the base station, for the

sensor nodes to begin their readings at the

defined sampling rate. The base station

broadcasts the message to the network.

The Stop button sends a command to stop data

acquisition at the sensor nodes. After that, the link

between base station and application is closed.

After these steps are executed, it is necessary to

discover the orientation of the sensor nodes (Section

2.3.3) and their body placement (Section 2.3.4). This

is achieved by pressing the Calibration button. This

action leads to the appearance of the Calibration

window. Note that after a calibration is performed,

these orientation and body placement values are

stored in a file. The purpose of this setting is to

facilitate the modules configuration by the user, who

can get the stored configuration by pressing the Last

Know Configuration button (Calibration window).

2.3.3 Node Orientation Discovery

The wireless sensor nodes are attached to the body

segments with Velcro straps, always with the

antenna opposed to the body, in order to minimize

the body effects in the wireless propagation signal,

and placed sideways when attached to legs or arms.

Firstly, the user is asked to stand up straight and

be still, as depicted in the 3D model, for a period of

time. During that amount of time, the API of the

application is collecting raw data from the sensor

nodes and through them calculating the Euler angles.

The estimation of the pitch and roll angles is made

using the gravity component obtained from the

measurement of the accelerometer, while the yaw is

determined from the measures from the electronic

compass unit (magnetometer).

The three coordinate rotations in a determined

sequence describe the orientation of the module; the

rotations of the Euler angles are discussed in

Section 3. Once the final rotation matrix is

calculated, it is multiplied by the gravitational vector

(0,0,1) in x, y and z coordinates, respectively. This

multiplication gives the weight of the module

orientation on each axis. With this weight and the

normalized raw values of the sensors, the orientation

of the sensor node in the body segment can be

determined.

Since there is no offset calculation on the axis, it

implies that the module will always be considered to

have the calculated axis with more weight facing the

gravity, when considering its maximum value (+1g).

This can be problematic if the sensor node is not

attached correctly to the chosen body segment; to

tackle this issue, offset calculations will be

introduced in the future.

2.3.4 Body Segment Discovery

The process of finding the location of the sensor

nodes on the body starts immediately after the

application concludes the orientation discovery

process. A set of instructions (info messages) are

given to the user, with the intent of guiding him

through the procedure of determining the placement

of the sensor nodes attached to the body. These

guidance instructions consist of leading the user to

make specific movements that allow the application

to recognize the corresponding movements for each

sensor node installed. Take for instance the case

where only two sensor nodes are installed. When

rehabilitating the arm, with the modules in the upper

or lower member (forearm), the user is asked to

move the forearm. Then the application detects

which has the largest variation of movement. Even

with a small movement, this is valid. The segment

detection for the leg uses the same principle.

3 ANGLE MEASUREMENT

The accuracy of the sensor nodes, after on-site

calibration, is limited by the sensor’s noise, which

adds to the system a measure of 1 degree error.

Adding to this, it is known that accelerometer

readings are influenced, not only by gravity, but also

from acceleration during the motion of the body.

ATelerehabilitationSystembasedonWirelessMotionCaptureSensors

This type of acceleration is not desirable, because it

induces errors at the time of converting raw data

from the sensors into angles. To tackle this problem,

a compensation algorithm is applied to the

accelerometer. Each accelerometer reading is

composed by three values (one for each reading

axis). If no linear acceleration is present (only the

gravitational component), the vector magnitude must

always equal to 1g. If this magnitude is not unitary

under a certain tolerance, meaning that the measure

contains linear acceleration besides gravity, a

compensation algorithm is initiated. This algorithm

uses the values from the current and previous

magnetic readings to compensate for the linear

acceleration errors. This method consists on using

two consecutive magnetic readings (previous and

current) to evaluate their angle () and axis of

rotation (axis), as shown in equations 1 and 2. The

calculation of the angle and the axis of rotation,

allows the rotation of the previous gravity vector to

its new orientation (Dunn and Parberry, 2011).









⋅









(1)



















(2)

The smallest angle between the two vectors

(normalized previous and current magnetic readings)

is provided by the dot product seen in equation 1.

This equation is valid since the magnitude of the

magnetic vector is unitary. The cross product

between those two vectors, equation 2, results on the

rotation vector, around which the previous

acceleration vector (gravity vector) is rotated.

The three angles that describe the orientation of

the sensor nodes are obtained using the equations 3,

4 and 5 respectively. The calculus of Pitch and Roll

use the variations of the normalized accelerometer

vector on each axis (a

). Yaw angle is obtained

using only the horizontal components (



,



) of

earth’s magnetic field, previously calculated.























(3)













(4)













(5)

The right-handed coordinate system was used to

calculate the three orientation angles, as shown in

Figure 7. The forward orientation of the module is

set to be along the y axis, with the pitch defining the

rotation on the x axis, the roll on the y axis and the

yaw on the z axis.

Figure 7: Right-handed coordinate system axis

representation.

4 EXPERIMENTAL RESULTS

Experiments were carried out to test the developed

system. One of them had two sensor nodes attached

to the right arm (upper arm and forearm). The goal

was to verify the replication of the body movement

in the 3D model present in the application. Figure 8

and Figure 9 exemplify these experiments by means

of photographs taken while the movement was being

executed.

Figure 8: Right arm movement replication, first position.

Figure 9: Right arm movement replication, second

position.

PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems

In order to simulate a rehabilitation session, a

second test was performed. The purpose was to

measure the angle between the upper and forearm.

The user was requested to do a simple movement,

flex and extend the arm, multiple times. The angle

() between two joint connected body segments, in

this case upper arm and forearm, is calculated using

equation 6.











⋅



















(6)

At this point, the application only deals with

rehabilitation in both arms and legs, with two sensor

nodes in the network. Figure 10 is a draft of what is

desired in a session file. The wave variations seen on

the right side reflect the angle variation (in degrees)

between the body segments over time.

Figure 10: Rehabilitation application draft interface.

5 CONCLUSIONS AND FUTURE

WORK

This paper describes early studies of the design and

development of a wireless sensor network, based on

wearable and easily placed inertial and magnetic

sensors, when applied to physical activities. A Java

application capable of collecting data from the

sensors and present them in a real-time 3D model of

a human body is under development, with the aim of

assisting elderlies in the practice of physical activity.

At this stage, only the motion capture system

capabilities can be evaluated. The displayed Java

application does not transmit any kind of movement

feedback; merely reproduces the executed motion.

The main goal of this project consists in developing

a home-based rehabilitation system that, through

constant monitoring of the movement, is capable of

interacting in real-time with the user. The objective

of this interaction is to assist the user to perform the

prescribed exercise correctly. Poorly executed

exercises can delay the rehabilitation process, or

even cause more damage. Thus, medical specialists

(physiotherapists) should take into consideration

physical limitation of the clinical subject in the

preparation of physiotherapy sessions, and set well-

defined objectives with regard to patient outcomes.

As future work it is intended that the patient,

while doing the prescribed exercises at home,

receives feedback from the application about the

correctness of those movements. Also, future work

includes inserting more sensor nodes in the network,

in order to obtain (simultaneously) the angle

between several joint connected segments, thus

allowing better assessing the user’s performance.

The application graphical interface will be

improved, simplifying it and giving more emphasis

to the 3D model of the human body and

rehabilitation graph.

Studies are being carried out in order to facilitate

the automatic body segment recognition for more

than two sensor nodes. As the number of nodes

increases, the number of required movements also

increases, and recognition will have to occur in

separate steps. In certain physical activities, it is

advisable to have one sensor node attached to the

torso, as it can provide a more accurate perspective

of the entire motion sequence.

Studies are also being conducted to introduce the

3-axis MEMS gyroscope present in the MPU-6000

in the module’s orientation calculation. The use of

the gyroscope allows compensation (not elimination)

of the linear acceleration detected by the

accelerometer and compensates false angular

movements.

The natural evolution of this system includes

also the development of a single, smaller board

containing both the CC2530 and the sensors, as well

as a printed circuit antenna to replace the existing

external antenna.

ACKNOWLEDGEMENTS

Project “AAL4ALL”, co-financed by the European

Community Fund FEDER through COMPETE –

Programa Operacional Factores de Competitividade

(POFC). FCT – Foundation for Science and

Technology – Lisbon, Portugal, through project

PEst-C/CTM/LA0025/2013.

ATelerehabilitationSystembasedonWirelessMotionCaptureSensors

REFERENCES

Afonso, J., Silva, H., Macedo, P., Rocha, L., 2011. An

Enhanced Reservation-Based MAC Protocol for IEEE

802.15.4 Networks. Sensors, Vol. 11, Issue 4, April

2011, pp. 3852-3873.

Aminian, K., Najafi, B., 2004. Capturing human motion

using body-fixed sensors: outdoor measurement and

clinical applications. In Computer Animation and

Virtual Worlds, Volume 15, Issue 2, pp. 79-94.

Cutti, A., Giovanardi, A., Rocchi, L., Davalli, A.,

Sacchetti, R., 2007. Ambulatory measurements of

shoulder and elbow kinematics through inertial and

magnetic sensors. In Medical & Biological

Engineering & Computing, Volume 46, Issue 2, pp.

169-178.

Claypool, M., Claypool, K., Damaa, F., 2006. The Effects

of FrameRate and Resolution on Users Playing First

Person Shooter Games. In Proceedings of a

ACM/SPIE Multimedia Computing and Networking

(MMCN).

Dunn, F., Parberry, I., 2011. 3D Math Primer for Graphics

and Game Development. Second Edition pp.141-143,

ISBN-13: 978-1-4398-6981-9.

Farella, E., Pieracci, A., Benini, L., Rocchi, L.,

Acquaviva, A., 2008. Interfacing human and computer

with wireless body area sensor networks: the

WiMoCA solution. In Multimedia Tools and

Applications, Volume 38, Issue 3, pp. 337-363.

Fuchsberger, M., 2008. Ambient Assisted Living: Elderly

People’s Needs and How to face Them. In

Proceedings of the 1

ACM international workshop on

Semantic ambient media experiences.

Hadjidj, A., Souil, M., Bouadballah, A., Challal, Y.,

Owen, H., 2012. Wireless sensor networks for

rehabilitation applications: Challenges and

opportunities. In Journal of Network and Computer

Applications.

IEEE Std 802.15.4-2006, Part 15.4: Wireless Medium

Access Control (MAC) and Physical Layer (PHY)

Specifications for Low-Rate Wireless Personal Area

Networks (WPANs), September 2006.

Kleinberger, T., Becker, M., Ras, E., Holzinger, A., 2007.

Ambient Intelligence in Assisted Living: Enable

Elderly People to Handle Future Interfaces. In 4

International Conference on Universal Access in

Huma-Computer Interaction, UAHCI.

Linz, K., Stula, S., 2012. Demographic change in Europe –

An Overview. Working Paper No. 4 of the

Observatory for Sociopolitical Developments in

Europe.

Moeslund, T. B., Granum, E., 2001. A survey of computer

vision-based human motion capture. In Computer

Vision and Image Understanding.

Steffen, D., Bleser, G., Weber, M., Stricker, D., Fradet, L.,

Marin, F., 2011. A personalized Exercise Trainer for

Elderly. In 5

International Conference on Pervasive

Computing Technologies for Healthcare

(PervasiveHealth) and Workshops.

Stula, S., 2012. Living in Old Age in Europe – Current

Developments and Challenges. Working Paper No. 7

of the Observatory for Sociopolitical Developments in

Europe.

Sun, H., De Florio, V., Gui, N., Blondia, C., 2009.

Promises and Challenges of Ambient Assisted Living

Systems. In Information Technology: New

Generations. ITNG’09.

Texas Instruments, “CC2530 data Sheet” April 2009.

Retrieved from http://www.ti.com

Wang, L., Hu, W., Tan, T., 2003. Recent developments in

human motion analysis. In Pattern Recognition,

Volume 36, Issue 3, pp. 585-601.

Wong, W., Wong, M., Lo, K., 2007. Clinical Applications

of Sensors for Human Posture and Movement

Analysis: A Review. In International Society for

Prosthetics and Orthotics.

PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems