Analysis of Robust Implementation of an EMG Pattern Recognition

based Control

Simone Benatti

, Elisabetta Farella

1,3

, Emanuele Gruppioni

and Luca Benini

1,4

DEI, University of Bologna, Bologna, Italy

Centro protesi, INAIL, Vigorso di Budrio, Italy

ICT Unit, Fondazione Bruno Kessler, Trento, Italy

Integrated System Laboratory, ETHZ, Zurich, Switzerland

Keywords:

EMG, Patter Recognition, Multisession, Active Prosthesis.

Abstract:

Control of active hand prostheses is an open challenge. In fact, the advances in mechatronics made available

prosthetic hands with multiple active degrees of freedom; however the predominant control strategies are

still not natural for the user, enabling only few gestures, thus not exploiting the prosthesis potential. Pattern

recognition and machine learning techniques can be of great help when applied to surface electromyography

signals to offer a natural control based on the contraction of muscles corresponding to the real movements.

The implementation of such approach for an active prosthetic system offers many challenges related to the

reliability of data collected to train the classiﬁcation algorithm. This paper focuses on these problems and

propose an implementation suitable for an embedded system.

1 INTRODUCTION

Assistive technologies in the last years are boosting

interesting research efforts to enhance quality of life.

In fact, the availability of accurate sensing technolo-

gies at a relatively low price and the possibility to ex-

ploit the power of low-cost low-power and fast mi-

crocontrollers enable a whole new branch of applica-

tions, where wearable smart sensors and embedded

systems can be used not only for monitoring but also

for the implementation of sensor fusion techniques or

complex machine learning algorithms and for provi-

sion of stimuli through actuation. An application ﬁeld

such as the control of prostheses can greatly beneﬁt

from this rapid technology evolution.

Recently, multi-ﬁnger active prostheses of the up-

per limb have appeared at commercial level (e.g.

Touch Bionics i-Limb, RSL Steeper BeBionics 3,

Otto Bocks SensorHand) enabling a larger set of ges-

tures w.r.t. previous prostheses, therefore asking for

an adequate strategy for their control. State of the

art technologies for the feed-forward control of ac-

tive hand prostheses are controlled via surface elec-

tromyography (EMG) in a way that forces the user to

learn to associate contraction of what remains of the

muscle to unrelated postures of the prosthesis, e.g.,

sequences of wrist ﬂexion and extension correspond

to various gestures. While at one side these tech-

niques grant a good reliability and short activation

time (within 30ms for the detection of the activity and

less than 300ms for the classiﬁcation), on the other

side the control strategy is non-natural, requiring fo-

cus and a non-trivial learning curve for the user. It

would be desirable instead to command the prosthe-

sis movement by activating the muscle as to move the

phantom limb.

Luckily, scientiﬁc literature recently proved that

machine learning applied to EMG signals could be

beneﬁcial in prosthetics to bring the control towards

more intuitive and natural strategies. In fact, convinc-

ing results have been shown both on healthy subjects

(Englehart et al., 2001) and amputees (Castellini et

al., 2009). This paper describes the preliminary anal-

ysis done in an on-going work towards a real-time

embedded implementation of an EMG based control

of an active hand prosthesis by use of machine learn-

ing techniques and in particular by use of a Support

Vector Machine. Starting from the lesson learnt by

literature, this work faces, as ﬁrst step, the variability

on classiﬁcation results due to the changes in place-

ments of the EMG sensors that occurs in real life. In

fact, the wearer limb is subject day by day to physical

differences due to swelling, fatigue, perspiration that

can cause misplacements of the sensors w.r.t. the de-

Benatti S., Farella E., Gruppioni E. and Benini L..

Analysis of Robust Implementation of an EMG Pattern Recognition based Control.

DOI: 10.5220/0004800300450054

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2014), pages 45-54

ISBN: 978-989-758-011-6

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

sired position. Furthermore the same gesture can be

performed in various positions and orientation of the

arm (along the body, lifted up, etc.). These changes

affects the classiﬁcation performance and must be ad-

dressed properly.

Starting from a data collection repeated in differ-

ent days on 10 healthy subjects, this paper proposes

the analysis of the reliability of the gesture recog-

nition in an EMG signal controlled hand prosthesis.

The analysis takes into account that our ﬁnal target is

an embedded implementation. The analysis is based

on three main elements: (i) the selection of a cor-

rect signal acquisition chain, (ii) the analysis of the

physiological best placement of the EMG sensors to

grant a robust classiﬁcation, and (iii) the analysis of

the performance of the system through days, evalu-

ating the difference of performance when the sensors

are placed and removed in different sessions, as in the

real use of the device for the classiﬁcation of natu-

ral movements. The activation of the same gesture is

considered in multiple combinations of the arm orien-

tation and position to better study their inﬂuence on

the classiﬁcation performance and consequently im-

prove the robustness of the classiﬁer.

The paper is organized as follows. Section 2 il-

lustrates background and related works,Section 3 in-

troduces the architecture of the system used to get the

EMG signals. Section 4 describes the tests made with

the collected data. In section 5 we show the results

and discuss the solution advantages. Finally in sec-

tion 6 we draw the conclusions.

2 BACKGROUND AND RELATED

WORKS

Three types of prostheses are widely available for

people with upper limb amputations: passive, body

powered, and electrically powered. Passive prosthe-

ses are often employed for cosmetic purposes and

have limited functionality. Body-powered prostheses

are used to restore basic tasks such as opening and

closing a terminal device. These devices are often

used because they are simple, robust, and relatively

inexpensive. The user can actuate electrically pow-

ered or active prostheses but they are advantageous

w.r.t. body-powered ones because they require less

user effort, as movement is actuated with DC mo-

tors. They can be controlled through a variety of

means such as force sensors, linear potentiometers,

and EMG signals. Electrically powered prostheses re-

store some functionality to amputees, but control of

these devices is typically limited to only one or two

degrees of freedom.

As mentioned, however, recently, multi-ﬁnger ac-

tive prostheses of the upper limb have appeared at

commercial level (Bebionics, 2012, TouchBionics,

2013, Ottobock, 2009). These prostheses, driven

by EMG sensors, can replicate most of the princi-

pal movements of the hand. To achieve robustness

the movement of the active prostheses are typically

driven by non-natural activation patterns, i.e. they

decode mainly sequence of ﬂexion and extension of

the wrist (Castellini and Smaag, 2009). The most

accurate EMG signal is taken directly on the spot

near the muscular ﬁbers by use of implantable sens-

ing electrodes; however, they are invasive and pose

safety issue, needing surgery. Our proposed applica-

tion prefers surface EMG sensors. They suffer lack of

performance, due to the noise of the skin surface and

the crosstalk of near muscle. Nevertheless we can use

improving signal methods (Reaz et al., 2006) for a

machine learning approach.

In literature, machine learning algorithms chosen

to extract muscular pattern and classify gestures vary

from Linear Discriminant Analysis (LDA) classiﬁer

(Young et al., 2013) to Neural Network (Matsumura

et a.l, 2002). Liarokapis and al. (2012) compare a

set of classiﬁers for EMG signal and conclude that

SVM (Boser et al., 1992) is the most accurate algo-

rithm for pattern recognition with these kind of sig-

nals. Other works, like (Oskoei and Hu, 2008) and

(Chen and Wang, 2013), made comparison for EMG

pattern recognition and conclude that SVM gives the

best result for these signals. The work of Englehart

and al. (2001) studies how to enhance the perfor-

mance of the SVM algorithm to reach the best accu-

racy of the classiﬁcation, by optimizing the classiﬁ-

cation parameters and the feature extraction. These

works give contributions in ﬁeld of signal processing

because they propose an optimized solution to a clas-

siﬁcation problem, but they do not keep in account the

issues related to the use in daily life use, like for ex-

ample the variation of classiﬁcation accuracy during

arm movements, or the sensor misplacing due to day

by day application of the prosthesis.

Some papers show experiments with a high num-

ber of subjects and features to test the accuracy of

their classiﬁcation algorithm. These results reach ac-

curacy ratio near 100% but they are not applicable

in real scenarios, because they do not take into ac-

count the variability of the signal caused by the place-

ment of the EMG sensors. The misplacement of the

EMG electrodes among different sessions and the po-

sitions of the arm during the use of the prosthesis af-

fect the classiﬁcation performance. This work tries to

address the problem analysing the best placement of

four EMG sensors to maximize the recognition per-

BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

formance. Furthermore, we analyse the variability of

training data along different days and caused by dif-

ferent positions of arm and forearm while performing

the same gesture.

3 SYSTEM SETUP

3.1 EMG Signal Acquisition

The EMG signal measures the electrical activation of

the muscular ﬁbres. Muscle tissue conducts electrical

potentials similar to the way nerves do and these elec-

trical potentials are named muscle action potentials.

When EMG electrodes are placed on the skin surface

the signal is composed by all the action potential of

the ﬁbres underlying the electrode. The surface EMG

sensors are made by two metal plates each one con-

nected to the inputs of a differential ampliﬁer that can

sense the action potential of muscular cells.

The amplitude of this signal is 20mV (-10 to +10),

with a bandwidth of 2000Hz. This kind of signals are

also very noisy and difﬁcult to manage. The main

causes of this noise are the motion artefacts, the elec-

trical equipment noise and the ﬂoating ground noise,

because the body is not referred to a solid ground po-

tential. To obtain information useful for the classi-

ﬁcation algorithms, it is required that sensors min-

imize the noise and provide signiﬁcant quantitative

pattern of muscular activation. For these reason we

used, instead of classical differential sensors, the Ot-

tobock 13E200 (Figure 1), a family of pre ampliﬁed

sensors with single ended output (Ottobock, 2009). In

Ottobock sensor the EMG signal is ampliﬁcated and

integrated to reach an output span of 0 - 3.3V, ideal

for the single ended stage of an embedded microcon-

troller ADC.

The bandwidth of the Ottobock sensor is 90-

450Hz with a further notch ﬁlter for the 50Hz. This

is because the sensors for the classiﬁcation of the ges-

tures do not need extensive frequency information but

a clear low noise signal. The analog signals are ac-

quired from an embedded board based on ARM COR-

TEX M4 microcontroller. The internal 16-bit ADC

samples the data and an external Bluetooth interface

sends it to a laptop. The embedded custom solution

is preferable with respect to a data logger solution in

this application because the ﬁnal goal is to implement

a complete embedded real time system. The board

used for this application is shown in Figure 2. The

data collected by the PC are managed in Matlab, for

the signal processing and the pattern recognition.

Figure 1: Ottobock sensor.

Figure 2: Embedded board.

3.2 Sensor Placement

The position of sensors is a critical task in using

a limb prosthesis. Muscles in healthy subjects are

placed in known position and it is possible to have lit-

tle differences in the amplitude of the activation signal

or in the position of EMG sensor. Amputees presents

speciﬁc characteristics in muscular structure and dif-

ference of amplitude of activation signal if compared

with healthy subjects and, for this reason, it is impor-

tant to evaluate subject by subject the possibility for

an amputee to use an active prosthesis. The position-

ing strategy is tested on healthy subject in this paper

and the test on amputees is the goal of future works.

Our strategy started from a previous work of

Castellini et al. (2009), that placed 5 sensors on an

elastic strip, with an equal distance each others. The

idea was to consider a sort of general pattern of the

forearm muscle activation, without focusing attention

on the anatomical structure of the arm. The risk of

such approach is to lose the contribution of one or

more sensors if misplaced. We propose an approach,

based on a four sensors conﬁguration, in two steps: a

AnalysisofRobustImplementationofanEMGPatternRecognitionbasedControl

theoretical one that considers the muscles involved in

hand movement and a practical one that veriﬁes the

good placement of the electrodes and the information

given from all sensors.

The theoretical approach starts from the analysis

of the muscular tissues of the limb. Signals have max-

imum quality where the muscle is wider, and the mus-

cular ﬁbres are spread. In the upper limb, the optimal

zone is the proximal third. In the forearm, the mus-

cles are divided in four groups, and the use of sur-

face EMG sensors requires that muscle near the skin

surface must be preferred for the classiﬁcation. We

can divide the forearm muscles in two groups: mus-

cles in the internal part of the forearm (ﬂexor radi-

alis carpi, palmares longus, ﬂexor carpi ulnaris, ﬂexor

superioris digitalis), involved in ﬂexion movements,

and muscles placed in the external part of the forearm

(externsor comunis digitorum, extensor digiti minimi,

externsor carpi ulnaris), involved mainly in extension

movements.

Nevertheless, the theoretical selection of muscles

is not sufﬁcient, because in the practical case the elec-

trode catches a zone of many muscular ﬁbres, intro-

ducing noise and requiring to verify the usability and

the integrity of the target signal. Forearm was divided

in different zones considering the dimension of the

selected Ottobock sensors. The zones are numbered

with progression, starting from the ﬂexor carpi radi-

alis, with counter clockwise sequence. Sensors are

placed on the muscle primarily involved in movement

selected. The muscles are found with tactile anal-

ysis and the corresponding number is used to place

the sensor on the elastic strip used in this experiment

(Figure 3).

By placing sensors on the ﬂexor carpi radialis,

ﬂexor carpi ulnaris, extensor digitorum communis

and extensor carpi ulnaris we obtained a good differ-

entiation in classiﬁcation pattern. The assumption for

a pattern recognition control system is that the set of

signals and features describing a given state of muscu-

lar activation are different from one state of activation

to another. Figure 4 shows the gesture of the hand

and the corresponding activation pattern acquired by

the four sensors. The amplitude of each of the four

sensors is a clear discriminant among patterns, even

if the presented solution is not intended as the best

Figure 3: Muscle selection.

Figure 4: EMG ampliﬁed signal pattern for selected ges-

tures.

placement, because the ﬁnal target of this application

is the use of the board on transradial amputees, and for

these patients the placement strategy must be tuned

subject by subject.

3.3 Classiﬁcation of the Data

A pattern recognition algorithm classiﬁes the differ-

ent hand gestures. The control is considered a super-

vised classiﬁcation problem. Training data are col-

lected during the controlled training session and a set

of labels are assigned to each pattern to train the clas-

siﬁer. SVM is a machine learning technique used for

pattern recognition (Burges, 1998). SVM goal is to

found the optimal separation hyperplane between two

classes. Support vectors are samples of the training

dataset used to build the separation hyperplane.

In the SVM algorithm, only the data of the train-

ing set are used to create the separation hyperplane

and maximize distance between plane and datasets.

Sometimes data are not linearly separable, and SVM

algorithm can map the predictor on a higher dimen-

sion space, where it is possible to separate data. If

BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

the system becomes too complex and the calculation

of the Euclidean distance between the point of the

dataset and the separation hyperplane becomes too

hard, SVM can introduce Kernel functions. These

kind of functions operate in a feature space and calcu-

late only the inner product between images of all pairs

in the feature space instead of the Euclidean distance

in a high dimensional space. This operation is com-

putationally simple and allows using SVM with high

dimensional classiﬁcation problems. Initially the al-

gorithm was developed only for a 2-class problem, but

it is possible to expand it with a One Versus All ap-

proach. The implementation of the SVM algorithm

used in this paper is taken from libSVM (libSVM,

2011) a freeware library implemented for Matlab and

C, compiled by GCC and usable in embedded sys-

tems.

The EMG signals present differences from one ac-

quisition to another. This behavior is due to many fac-

tors, clinical, electrical, and patient dependent. The

not fully repeatable placement of the electrodes that

can generate different traces and crosstalk, the differ-

ent muscular contraction strength due to fatigue or re-

peated muscular activity, the variable conductivity of

the skin surface, due to humidity, perspiration or other

physiological and environmental condition, can in-

duce signiﬁcant changes in the shape of the activation

pattern. Some training strategies can try to overcome

this problem. In the work of Hargrove et al. (2008),

authors used a sort of extended dataset method, based

on particular high density EMG system. The dataset

collected using these sensors can make more robust

the training phase but it is difﬁcult to use this system

in a real application. Furthermore, it is not consid-

ered that the activation patterns can change in differ-

ent days and for different position of the arm. For this

reason, an important part of this work is a collection

of a robust dataset of EMG signals.

3.4 Acquisition Protocol Description

The collection of the dataset follows an acquisition

protocol described in this paragraph. Selected move-

ments are the most common hand gesture used in

daily life, to ensure a good quality of social reintegra-

tion. Gesture selected are shown in Figure 4 (closed

hand, open hand 2-ﬁnger pinch, 3-ﬁnger pinch, point

index).

The classiﬁcation includes also the rest position of

the hand, recorded between two subsequent gestures.

There are 9 subjects involved in the dataset acquisi-

tion. Gestures are acquired in 10 sessions collected 1

for day in 10 days not necessarily consecutive.

The sequence of gestures is repeated and scram-

bled in four arm positions (Figure 5). Each acquisi-

tion session is divided in 4 steps, one for each position

of the forearm. The collected sequence is composed

by 10 repetitions of muscular contraction 3 second

long. Between each contraction there are 3 seconds

of rest. One ﬁle for each acquisition is captured. Sub-

jects wear an elastic strip with the 4 EMG sensors.

Figure 5: Different arm position.

Sensors positions are tuned with a simple proce-

dure described below. An arm as shown in Figure 6.

The operator ideally traces a line on the axial direc-

tion of the forearm (Figure 6-left). Then the opera-

tor places Sensors 1 and 2 at 30mm respectively on

the left and on the right side of the line at the prox-

imal third of the forearm. The operation is repeated

for sensors 3 and 4 with the arm ﬂex in the position

shown in Figure 6-right by considering the ideal line

and placing sensor 3 and 4 at 30mm of distance sym-

metrically at the two sides of the line, at the proxi-

mal third of the forearm. Once the sensor are placed,

a good positioning signals trace is acquired to avoid

misplacement of the strip. The test is made with the

gestures of hand open and closed,, which correspond

to a non-zero signal for all the sensors.

To avoid that the subject learns the sequence of

gestures and loses the naturalness in movements, the

sequence are scrambled and each acquisition in a dif-

ferent arm position has a different pattern of gesture.

Figure 6: Sensor placement.

AnalysisofRobustImplementationofanEMGPatternRecognitionbasedControl

Table 1: Confusion matrix for single training session.

ACCURACY % REST PREC OPEN POWER POINT

99,4 REST 77185 118 0 339 0

91,6 PREC 149 24992 0 1688 445

75,8 OPEN 125 416 10088 2689 0

95,7 POWER 186 162 6 14745 302

82,3 POINT 82 756 0 1820 12371

For the same reason, the positions of the arm are not

always in the same sequence: for example in ses-

sion 1 sequences are captured in distal position and

in proximal position, in session 2 these sequence is

inverted, and scrambled with different pattern of ges-

ture in other sessions.

4 EXPERIMENTAL RESULTS

The implementation of a pattern recognition system

in a real application for the gesture classiﬁcation has

three main issues: the correct placement of the EMG

sensors, the variability of the activation pattern due

to the various positions of the arm and the varia-

tion of muscular contraction among following acqui-

sition sessions. The importance of this aspect is ac-

knoledged in literature, where other works address

speciﬁcally the problem of placement and the dif-

ference between surface and intramuscular sensors

(Hargrove et al., 2007) and propose their placement

schemes. The training of the SVM is a critical phase

since the quality of the training and the resulting ma-

trix strictly affects the ﬁdelity of the recognition. In

turn, the segmentation phase performed on data col-

lected is therefore crucial to augment the quality of

the training. However data preparation, thresholding

and segmentation are subjective operations based on

the knowledge of the signal and on experience (Milo-

sevic et al., 2010).

Figure 7: Accuracy for a single training session.

Figure 8: Difference of accuracy between classiﬁcation in

the position of the training and in other arm position.

4.1 Single Arm Position Classiﬁcation

Our ﬁrst purpose for the experiments is to verify the

performance of the proposed placement scheme, to

guarantee that a good classiﬁcation is possible. For

this application, we considered only the three ﬁn-

gers precision grip, because many commercial pros-

thetic systems, in the operating conﬁguration, admit

to choose between two or three ﬁnger precision grips.

In the initial test, we considered only one session

dataset (i.e. 10 gestures per position) per each patient.

The training set is the 25% of the session dataset (i.e.

3 gestures per repetition) obtained with manual seg-

mentation. The training stage is performed ofﬂine on

MATLAB libSVM. After the creation of the model,

the classiﬁcation gives the mean accuracy for each

gesture. We selected the proximal position of the arm

because it is natural and tipically used in the EMG test

on pattern recognition accuracy. The data presented

are the result of the mean values of accuracy for all

subject involved in the experiment. Confusion ma-

trix (Table 1) shows that the rest position, the power

grip and the precision grip are the gestures recognized

with higher accuracy.

The open hand and the index point gestures are

recognized with lower accuracy because it is more

difﬁcult for the subject to repeat them exactly. Av-

erage classiﬁcation accuracy is between 75% and

99.4%, as shown in Figure 7. This data validates the

positioning strategy proposed and the setup used in

our approach, because results meet the performance

declared in literature on this topic.

BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

4.2 Different Arm Position

Classiﬁcation

The variability in arm position is another important is-

sue, when an EMG classiﬁcation system is proposed

for upper limb prostheses. Scheme et al. (2010) and

Fougner et al. (2011) start to analyse this problem and

propose a solution based on the placement of two in-

ertial sensors used to detect the exact position of the

limb. The position of sensors is on the arm and fore-

arm and the combination of their information is used

to detect the arm position. This approach is interest-

ing, however it presents some practical problems if in-

tegrated in user daily routine. The solution presented

is not integrated; the EMG board and inertial sensors

are on separate boards and need separate power sup-

ply and a communication infrastructure. Furthermore,

the system requires frequent calibrations due to the

misplacement of inertial nodes on the body.

In this section, we evaluate the performance im-

provements coming by the use of an additional train-

ing session, which combines samples of the patterns

from the different arm positions without modifying

the test setup. The use of an excessively large training

sets in traditional machine learning approach is not

recommended, because usually the number of support

vectors created by the algorithm is too high to guaran-

tee a strong classiﬁcation and a computational charge

suited for an embedded implementation.

In the case study, we identiﬁed the most common

positions of the arm in which a grip action of the hand

is required. Positions are shown in Figure 5. Initially

a training session is performed for each patient only

in the proximal position of the forearm and the clas-

siﬁcation performance is evaluated for the other limb

position in the same acquisition session without fur-

ther training.

Figure 8 plots the difference of accuracy in classi-

ﬁcation with arm positions not included in the train-

ing, showing the decreasing of performance com-

pared with the classiﬁcation in the position used for

training. Even in this case the open hand and the

point gestures suffer lack of performance. To verify

the efﬁciency of the training made for the different

positions of the arm, in Figure 9 the recognition ac-

curacy obtained by the SVM classiﬁer is shown when

training on one position compared with the training

in multiple positions. This strategy gives major bene-

ﬁts to the classiﬁcation of the two gestures with major

recognition errors because includes in the training set

the gesture with more variability.

Figure 9: Single vs multiple position training.

4.3 Multisession Classiﬁcation

Performance

The activation patterns of EMG sensors is strongly

dependent from the position and the orientation of the

electrodes (Young et al., 2012). Small displacements

of the sensors among the sessions can create big dif-

ferences in the EMG traces and consequently reduce

signiﬁcantly the accuracy of the classiﬁcation. The

difference of classiﬁcation performance in this paper

considering the training made on a single session or

on two different sessions. The work of Saponas et al.

(2010) uses an 8 electrodes system and evaluates the

performance of the system in 3 different sessions. The

proposed system has 8 electrodes and the set of ges-

tures chosen targets HCI applications, instead of our

daily use grasping types. Neverthless the results of

this work can conﬁrm the decreasing trend of the per-

formance caused by misplacement. The SVM in this

experiment is trained by merging the support vectors

determined in the previous tests and the vectors of an

additional session.

With this procedure, it is possible to observe the

behaviour of the recognition accuracy dependent from

the two sessions. The ﬁrst test uses the complete

training session of the last paragraph to evaluate the

accuracy of the classiﬁer among the different ses-

sions. The sequences of the movements are coded

in the acquisition protocol and for each trace the dif-

ferent movements can be located. The prediction of

the SVM algorithm is used on all the traces, and the

mean value of accuracy are collected for each patient.

The single training session cannot satisfy minimum

requirements for reliability of classiﬁcation. The dif-

ferences of accuracy in the 5 gestures are shown in

Figure 10.

The procedure to place the electrodes, described

in session 2, is standardized and repeated for all sub-

jects; however still little inconsistency in position of

the electrodes can cause big differences in muscular

activation pattern. To cope with this issue, we evalu-

AnalysisofRobustImplementationofanEMGPatternRecognitionbasedControl

Figure 10: Single vs double training session.

ate the use of multi-stage training and compare the ac-

curacy per arm positions in single versus double train-

ing session. The extended training merges two train-

ing sessions to create a unique training model. This

kind of training method enhances the performance of

the classiﬁcation accuracy as shown in Figure 11.

Figure 11: Single vs double training session in different po-

sitions.

5 DISCUSSION

The results of the experiment show that a robust clas-

siﬁcation for EMG signals cannot be achieved with-

out considering the issues coming from the electrodes

placement, which can present differences, even if

slight, from one use to another and/or day by day. Fur-

thermore, this work outlined the differences in classi-

ﬁcation accuracy, which occur for the same gesture

performed in diverse positions of the arm (parallel

to the ground, proximal to the body, lifted upwards,

etc.). The idea behind this paper is that multiple ses-

sions training is necessary for a realistic application

and a SVM with light signal ﬁltering can give good

accuracy in recognition of activation of muscular pat-

terns.

The proposed setup is ideal for an embedded im-

plementation, because the hardware setup is simple

and the signal processing is light weight, suitable on

a microcontroller. The EMG signal is ﬁltered and

de-noised directly by the sensors, and the hardware

signal conditioning allows to have well differentiated

patterns even with a four sensors setup and small dig-

ital ﬁltering.

The SVM training algorithm is quite heavy from

the computational point of view and runs on a PC ap-

plication ofﬂine, but it is possible to implement the

classiﬁcation on embedded platform maintaining re-

sponse time compatible with the use of a prosthesis.

The computational time of the prediction function is

in the most part dependent from the number of Sup-

port Vectors present in the model. More support vec-

tors indicates a complex model due to overlapping of

different patterns. The choice of an appropriate train-

ing set and the tuning of the algorithm parameters can

reduce the number of support vectors. Obviously the

model becomes more complicated when multiple ses-

sions and multiples arm positions are considered.

Figure 12 shows the number of support vectors in

different cases of the experiment.

Figure 12: Suport vectors number in different testing con-

ditions.

The platform used for the implementation of the

SVM classiﬁcation algorithm is an ARM Cortex M4,

with 100MHz clock. The execution time of the clas-

siﬁcation routine is measured to understand if the us-

age of an embedded low cost platform can satisfy the

speciﬁcation of the system. The test is performed with

different models, changing the kernel type and the

number of the support vectors, to evaluate the trade-

off between complexity and response time of the al-

gorithm.

Figure 13 and 14 shows the computation time ob-

tained with FPU (Floating point unit) calculations.

The response time allows to use the embedded plat-

form to implement an active control of the prosthesis,

with linear or RBF (Radial Basis Function) kernels.

The number of support vectors can be reduced with

BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

Figure 13: Linear kernel computation time.

Figure 14: RBF kernel computation time.

an appropriate training sample selection and with the

tuning of model made on the single subject. In this ex-

periment we did not perform optimization of the pa-

rameters or differentiated selection of the samples in

creating the training models. This choice is intended

to show the starting point in implementing an embed-

ded application and wants to contribute in evidenc-

ing and focusing on the problems and the issues that

become critical in the implementation of this kind of

devices.

6 CONCLUSIONS

In this analysis we found that the main challenge for

the classiﬁcation accuracy in a real application of an

EMG interface is the acquisition interface, intended

as the choice of the kind of sensor used and placement

strategy. We proved that 4 sensors correctly placed

and a light signal processing can give high classiﬁ-

cation accuracy, comparable with systems with more

sensors and a signal processing with higher computa-

tional cost. It is not possible now to obtain good per-

formance without considering the difference of classi-

ﬁcation in different arm positions and among multiple

sessions, if EMG surface sensors, which can be re-

moved, are used. The strategy of a double trining ses-

sion, in different days, is compatible with the clinical

scenario, because it is normal for a patient with an up-

per limb prosthesis to have periodical check-up of the

prosthesis with technicians and doctors. Future works

will test this methodology on subjects with transradial

amputation of forearm, to optimize and standardize

the placement methodology.

Another important challenge that we intend to

achieve is the optimization of the parameter of the

classiﬁcation algorithm, which can speed up commu-

tation time and enhance classiﬁcation ratio, for exam-

ple with a proper data thresholding and scaling and

test with different kernels. The ﬁnal goal is to inves-

tigate all the problem related to a real and reliable ap-

plication and to come to an embedded implementation

of the control interface.

ACKNOWLEDGEMENTS

This work is partially funded by Centro protesi, IN-

AIL in Vigorso di Budrio, Italy that also supported

the deﬁnition of requirements and of the acquisition

protocol.

REFERENCES

Bebionics (2012). http://bebionic.com/.

Boser, B. E., Guyon, I. M., and Vapnik, V. N. (1992). A

training algorithm for optimal margin classiﬁers. In

Proceedings of the ﬁfth annual workshop on Compu-

tational learning theory, pages 144–152.

Burges, C. (1998). A tutorial on support vector machines

for pattern recognition. Kluwer Academic Publishers.

Castellini, C., Gruppioni, E., Davalli, A., and Sandini, G.

(2009). Fine detection of grasp force and posture by

amputees via surface electromyography. volume 103,

pages 255–262.

Castellini, C. and Smaag, P. (2009). Surface emg in ad-

vanced hand prosthetics. pages 35–47. Springer-

Verlag.

Chen, X. and Wang, Z. J. (2013). Pattern recognition of

number gestures based on a wireless surface emg sys-

tem.

Englehart, K., Hudgin, B., and Parker, P. (2001). A wavelet-

based continuous classiﬁcation scheme for multifunc-

tion myoelectric control. In IEEE Transactions on

Biomedical Engineering, pages 302–311.

Hargrove, L., Englehart, K., and Hudgins, B. (2007). A

comparison of surface and intramuscular myoelectric

signal classiﬁcation. volume 54, pages 847–853.

libsvm (2011). http://www.csie.ntu.edu.tw/ cjlin/libsvm.

Matsumura, Y., Mitsukura, Y., Fukumi, M., Akamatsu, N.,

Yamamoto, Y., and Nakaura, K. (2002). Recognition

of emg signal patterns by neural networks. In Neural

AnalysisofRobustImplementationofanEMGPatternRecognitionbasedControl

Information Processing, 2002. ICONIP ’02. Proceed-

ings of the 9th International Conference on, volume 2,

pages 750–754 vol.2.

Milosevic, B., Farella, E., and Benini, L. (2010). Continu-

ous gesture recognition for resource constrained smart

objects. In UBICOMM 2010, The Fourth Interna-

tional Conference on Mobile Ubiquitous Computing,

Systems, Services and Technologies, pages 391–396.

Oskoei, M. and Hu, H. (2008). Support vector machine-

based classiﬁcation scheme for myoelectric control

applied to upper limb.

Ottobock (2009). http://www.ottobock.com.

Reaz, M., Hussain, M., and Mohd-Yasin, F. (2006). Tech-

niques of emg signal analysis: detection, processing,

classiﬁcation and applications. Springer-Verlag.

Touchbionics (2013). http://www.touchbionics.com.

Young, A., Hargrove, L., and Kuiken, T. (2012). Improv-

ing myoelectric pattern recognition robustness to elec-

trode shift by changing interelectrode distance and

electrode conﬁguration. In Biomedical Engineering,

IEEE Transactions on, pages 645–652.

Young, A., Smith, L., Rouse, E., and Hargrove, L. (2013).

Classiﬁcation of simultaneous movements using sur-

face emg pattern recognition. In Biomedical Engi-

neering, IEEE Transactions on. IEEE.

BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing