2014 IEEE International Conference on Robotics & Automation (ICRA)

Hong Kong Convention and Exhibition Center

May 31 - June 7, 2014. Hong Kong, China

A Comparative Study on PCA and LDA Based EMG Pattern

Recognition for Anthropomorphic Robotic Hand*
Daohui Zhang, Xingang Zhao, Jianda Han, and Yiwen Zhao

Abstract A multifunctional myoelectric prosthetic hand is a

perfect gift for an upper-limb amputee, however, the
myoelectric control for a prosthetic hand is not so good now.
Here, the paper presents a comparative study on
electromyography (EMG) pattern recognition based on PCA
and LDA for an anthropomorphic robotic hand. Four channels
of surface EMG (sEMG) signals were recorded from the
subjects forearm. Time-domain analysis, frequency-domain
analysis, wavelet transform analysis, nonlinear entropy analysis
and fractal analysis were done and fourteen kinds of features
were extracted from sEMG signals. The features were divided
into four groups, and the performances of the four groups were
compared and analyzed. In the feature projection stage, three
schemes were proposed and their performances were compared
with each other. The first one only used the principal component
analysis (PCA) for dimension reduction. And the second one
only used the linear discriminant analysis (LDA) for dimension
reduction. The third one used PCA for the first step of
dimensionality reduction, and then used LDA for the next step
of dimensionality reduction. In the classification stage,
minimum distance classifier (MDC) was employed for
identifying nine kinds of hand/wrist motions in the projected
space. Comparative experiments of four groups of features and
three projection schemes were done and evaluated. The online
experiment of real-time myoelectric control for an
anthropomorphic robotic hand was done as well.

I. INTRODUCTION
Nowadays, many disabled people who have lost limbs in
wars, car accidents, and industrial accidents are provided with
prosthetic devices to face challenges in their daily life.
Advanced commercial prosthetic hand such as Otto Bock
hand [1], i-Limb hand [2], and Smart hand [3] have been able
to achieve some motions like human hands. And these devices
could help amputee persons improve the quality of life at
physical and psychological aspects. The approach using
electromyography (EMG) signals from amputee persons
remnant muscles to control prosthetic devices is beneficial to
restore the missing functionality of amputated limbs, and it
has a history of about 40 years [4]-[6]. Previous commercial
EMG-based prosthetic devices primarily adopt a conventional
threshold or proportional control strategy to achieve a few
simple actions such as hand opening and closing [7]. In recent
*This work was supported in part by the National Science Foundation of
China under Grant 61273355, 61273356 and the Independent Subject of
State Key Laboratory of Robotics under Grant 2013-Z06.
Daohui Zhang is with the State Key Laboratory of Robotics, Shenyang
Institute of Automation, Chinese Academy of Sciences, 110016, Shenyang,
China; with the Graduate School, Chinese Academy of Sciences, 100039,
Beijing, China (e-mail: zhangdaohui@sia.cn).
Xingang Zhao, Jianda Han and Yiwen Zhao are with the State Key
Laboratory of Robotics, Shenyang Institute of Automation, Chinese
Academy
of
Sciences,
110016,
Shenyang,
China
(e-mail:
zhaoxingang@sia.cn, jdhan@sia.cn, zhaoyw@sia.cn).
978-1-4799-3685-4/14/$31.00 2014 IEEE

years, myoelectric pattern recognition for a prosthesis system

has attracted increasing attention [8].
In an EMG pattern recognition system, feature extraction
is the process that converts original sEMG signals to a
compact and informative set of features [9]. So far, many
feature analyzing methods have been applied to EMG pattern
recognition problems. L. Hargrove et al. [10] indicated that the
combination of the time-domain features and the
autoregressive (AR) features had outstanding performance in
EMG pattern recognition. In [11], discrete wavelet transform
was compared with the combination of integral of sEMG
(IEMG) and power spectral density (PSD) for six simple hand
motions. In the previous work [12], we extracted five
time-domain features including the mean absolute value
(MAV), root mean square (RMS), zero crossings (ZC),
Waveform length (WL) and slope sign changes (SSC) and
four frequency-domain features including the power spectral
density (PSD), median frequency (FMN), median frequency
(FMD) and auto-regressive coefficients (AR).
Traditional methods such as time-domain and
frequency-domain analyzing methods have been widely
utilized in EMG pattern recognition, and they have a good
capability to track muscular changes. However, these methods
are not effective in detecting the critical feature of sEMG
signals during transient human movements. A sEMG signal
which is a complex signal embedded in noise is non-stationary
and stems from a high-dimensional nonlinear system.
Therefore, nonlinear methods such as nonlinear entropy
analysis and fractal analysis recently have been proposed to
analyze sEMG signals for extracting some informative
features which can detect the changes in different muscle
statuses. Zhao Jingdong et al. [13] extracted sample entropy
and wavelet transform coefficients from three channels of
sEMG signals for classifying six fingers movements. The
fractal dimension features were utilized for identifying finger
movements in [14]. Actually, nonlinear features can only be
seen in a few literatures owing to their high complexity.
Researchers commonly just focus on one or two kinds of
nonlinear features and illustrate the experimental results in
EMG pattern recognition. This study will do nonlinear entropy
analysis and fractal analysis and extract four representative
features. The traditional methods including time-domain
analysis, frequency-domain analysis and wavelet transform
analysis will be utilized as well, and ten frequently-used
features will be extracted. The total fourteen features from five
analyzing methods will be divided into four groups, and the
recognition performances of the four groups of features will be
studied comparatively.
In order to recognize more motion patterns, more channels
of sEMG signals and more kinds of features are used to
increase the information extracted from sEMG signals. This

4850

leads to a high dimensionality of the feature vector. Thus,

feature projection is needed in EMG pattern recognition.
Feature projection could map high dimensional original
features to an appropriate low dimensional space. And feature
projection could be optimized by a learning criterion so that
the features belonging to the same class are clustered and the
generalization ability is improved. In addition, dimensionality
reduction of feature projection reduces the processing time of
pattern recognition, and is beneficial to meet the demand of
the real-time control. The principal component analysis (PCA)
is a conventional method which can project high dimension
data into a low dimension space and make the data not
relevant in the low dimension space. The paper [13] adopted
PCA to reduce the dimensionality of WPT coefficients from
four channel EMG signals. In this study, PCA is used to carry
out the first step of dimensionality reduction considering that
PCA merely generates a well-described coordinate space of
the features without considering class separation. The linear
discriminant analysis (LDA) could reduce the dimensionality
of features and meanwhile take the class separability into
account. Plus, LDA just needs a short processing time.
Therefore, in the study, LDA is used for the second step of
dimensionality reduction considering that the combination of
PCA and LDA projection may obtain a better integrated merits
deriving from both the two methods. In the classification
stage, the minimum distance classifier (MDC) is employed to
recognize the projected features.
In the previous work [12], we have used five time-domain
features and four frequency-domain features from four
channels of sEMG, and have compared the performance of
LDA and PCA+LDA for classifying five kinds of hand
motions. The goal of this study is to find an efficient feature
projection method based on PCA and LDA for EMG pattern
recognition and to compare and evaluate the performance of
four groups of features derived from the five analyzing
methods including time-domain analysis, frequency-domain
analysis, wavelet transform analysis, nonlinear entropy
analysis and fractal analysis, and finally to construct a most
efficient EMG pattern recognition system for nine kinds of
hand/wrist motions. In this study, the pattern recognition
system consists of the feature extraction including five kinds
of analysis methods, PCA feature projection, LDA feature
projection and MDC classification, as shown in Fig. 1. The
PCA+LDA projection is compared with PCA projection and
LDA projection respectively. Finally, the real-time EMG
pattern recognition system is implemented for an
anthropomorphic robotic hand, and its performance is
demonstrated.

Figure 1. Block diagram of EMG pattern recognition system.

II. DATA ACQUISITION

This study attempted to recognize nine kinds of hand/wrist
motions: hand closing (HC), hand opening (HO), index finger
pinching (IFP), middle finger pinching (MFP), wrist flexion
(WF), wrist extension (WE), wrist radial deviation (WRD),
wrist ulnar deviation (WUD) and relaxing motion (RM) as

shown in Fig. 2. Four surface electrodes were used to acquire

the sEMG signals from the extensor digitorum, flexor
digitorum superficialus, extensor carpiradialis, and flexor
carpiulnaris, respectively, which are concerned with the nine
motions. In order to collect high quality sEMG signals, the
skin was scrubbed with alcohol and shaved if necessary, and
then the electrodes with conductive gel were attached to the
corresponding positions as shown in Fig. 3.

HC

HO

IFP

WF

WE

Figure 2.

MFP

WRD

RM

WUD

Nine kinds of hand/wrist motions.

Figure 3. The placement of surface electrodes on a left forearm.

Since the amplitude of sEMG signals is at range of

100~5000uV, and the spectrum is mainly distributed in
10~500Hz, the study adopts the MyoScan EMG Sensor (see
Fig. 4), whose parameters are shown in Table I. The sEMG
signals are digitized by an A/D converter card, Advantech
PCI-1716, and each channel is sampled at a rate of 1000 Hz
with 16-bit resolution.

Figure 4. MyoScan EMG Sensor

TABLE I.
Sensors
Parameters
values

THE SENORS PARAMETERS

parameters

Power
supply

Amplifying
rate

7.2V

500

Input
signal
range

Active
frequency
range

Sampling
rate

0-2000uV

10-500Hz

2048s/s

In this study, sixteen groups of sEMG signals were

collected from one intact-limb subjects forearm. The first
eight groups were used for the training section, while the
remaining eight groups were used for the test section. For each
group, the time of data acquisition was 32 seconds, 2 seconds
for each active motion (the nine motions except the relaxing
motion), and a 2-second relaxing motion between each two
active motions. There was a two minutes rest time between
every two groups to prevent muscle fatigue.
III. FEATURE EXTRACTION
In order to control an robotic hand in real time without
perceiving a time delay, the processing time of EMG pattern

4851

recognition should be less than 300 msec. Thus, the scheme of

a sliding window with an incremental window was adopted
for the steady-state motion recognition. For the real-time
myoelectric hand control, all the processes including
transmitting control commands should be completed within an
incremental window. In this study, a 128-msec (128 samples)
sliding window with a 32-msec (32 samples) incremental
window was selected. After data segmentation, fourteen
features would be extracted in a sliding window. In this study,
we would use the five analyzing methods including the
time-domain analysis, frequency-domain analysis, wavelet
transform analysis, nonlinear entropy analysis, and fractal
analysis.

(a) Raw sEMG

(c) FMD

Time-domain analysis is one of the simplest and most

common feature analyzing methods. We selected six typical
EMG features including the mean absolute value (MAV), root
mean square (RMS), zero crossings (ZC), waveform length
(WL), slope sign changes (SSC), and autoregressive model
coefficients (ARC). Here, the 4-order autoregressive model
was adopted, and the parameters a1, a2, a3, a4 were chosen as
EMG features.

Nonlinear entropy analysis is a promising approach to

characterize non-stationary and nonlinear EMG signals. In this
study, the nonextensive entropy (TE) and sample entropy (SE)
were chosen to extract nonlinear information from sEMG
signals.
Fractal analysis is a new approach to describe the inherent
self-similarity and long-range correlation of sEMG signals.
Correlation dimension (DC) and box dimension (DB) are two
representative parameters of fractal dimension, thus they were
selected as EMG features for motion recognition in this study.
In total, 14 kinds of features deriving from five analyzing
methods were extracted from 4 channels of sEMG signals.
Some of those features are shown as Fig. 5. These features
were divided into 4 groups according to their performance and
category (see Table II). Group A includes 6 kinds of
time-domain features. Group B contains 3 kinds of
frequency-domain features and wavelet transform coefficients
(WTC). Group C consists of 2 kinds of nonlinear entropy
features and 2 kinds of fractal dimension features. Group D is
the combination of 3 kinds of time-domain features and
wavelet transform coefficients (WTC).

(d) WTC (a3)

(e) SE

(f) DC

Figure 5. Some features extracted from 4 channels of sEMG signals

Frequency-domain analysis is another traditional

analyzing method for EMG recognition. Average power (AP),
mean frequency (FMN), and median frequency (FMD) were
used to extract the frequency spectrum information from the
sEMG signals in the study.
Wavelet transform analysis is one kind of
time-frequency-domain analyzing methods, which can
simultaneously obtain time information and frequency
information from signals. In this study, 3-layer wavelet
decomposition and reconstruction were implemented, and the
mean values of coefficients d1, d2, d3, a3 were used as EMG
features. Experimental results indicated that the mean value
used less processing time yet obtained even higher
identification accuracy compared to the singular value
deriving from singular value decomposition in the study.

(b) RMS

TABLE II.
Feature
Grouping
features

THE FEATURE GROUPING

Groups

Group A

Group B

Group C

Group D

MAV, RMS,
ZC, WL, SSC,
ARC

AP, FMN,
FMD, WTC

TE, SE,
DC, DB

RMS, WL,
SSC, WTC

IV. FEATURE PROJECTION AND CLASSIFICATION

With the increase of the number of channels and features,
the feature vector becomes a high-dimensional vector. In order
to realize EMG pattern recognition efficiently, feature
projection is necessary to map a high-dimension feature vector
to an appropriate lower dimension space. Feature projection
could make the feature vectors which belong to the same class
clustered, and improve the generalization ability of the
classifier. In addition, dimensionality reduction of feature
projection could reduce the computational cost of EMG
pattern recognition. In order to find the best feature projection
scheme, the three projection schemes were compared in this
study (see Fig. 6). The first one was just using PCA for the
feature projection. The second one was that the principal
component analysis (PCA) was used for the first step of
feature projection, and then the linear discriminant analysis
(LDA) was used for the next step of feature projection. And
the third one was just using LDA for the feature projection. In
this study, the three projection schemes are written as PCA,
PCA+LDA, and LDA for short.

4852

Figure 6. Schematic diagram of the three feature projection schemes

As described in the previous work [12], PCA can project

feature vectors from a high-dimension space onto a
low-dimension space, and make the variance of data
maximum in the low-dimension space. Suppose that
X=(X1,X2,,Xn)T is an n-dimension random variable,
Y=(Y1,Y2,,Ym)T is an m-dimension random variable,
C=1/(n-1) (Xi-u)(Xi-u)T is the covariance matrix of the
samples. Assume that there is a linear transformation:
Y1 a1 1 X 1 a 2 1 X 2

Y 2 a1 2 X 1 a 2 2 X 2

Y a X a X
1m
1
2m
2
m

a n 1 X n a1 X
T

an2 X n a2 X
T

a nm X n a m X

(2)

To solve the problem the Lagrange multiplier is used, and

the formula Cai= ai is gotten. And then the principal
components could be gotten. In general, The i-th principal
component of X can be obtained by solving the corresponding
eigenvector of the i-th largest eigenvalue, and they are usually
required to be independent of each other in order to make the
information they contains not overlapping.
Finally, a set of orthogonal basis (a1, a2,, am), denoted by
A, is obtained, and it can project n-dimension random
variables X=(X1,X2,,Xn)T for m(m<n)-dimension random
variables Y=(Y1,Y2,,Ym)T by (3), achieving dimensionality
reduction meanwhile retaining the maximum information.
T

Y A X

(3)

As described in the previous work [12], LDA projects

high-dimension vectors onto an optimal discriminant space to
extract class information and reduce the vector dimension, and
makes sure that the projected vectors have the largest
between-class distance and the smallest within-class distance.
The between-class scatter matrix and the within-class scatter
matrix are defined as follows:
c

Sb

n u
i

u ui u

xk ui xk

Y A W X

(8)

The minimum distance classifier (MDC) is one of the

simple and efficient classification methods. MDC firstly
calculates the center of each class, and then finds the class
which has the minimum distance from the given input sample
to the class center, and the input sample is classified into the
minimum distance class. In this study, MDC is an appropriate
classifier because the projected feature vectors are easy to
indentify in a low-dimensional space, thus a complex
classifier is not necessary. In addition, MDC needs very short
computing time.
V. EXPERIMENTAL RESULTS AND DISCUSSIONS
A. Comparison of three feature projection schemes
In this section, PCA, LDA, and PCA+LDA were
compared with each other. The features of Group A (including
MAV, RMS, ZC, WL, SSC, and ARC) were adopted in this
comparative study. For the PCA projection scheme, an
appropriate dimensionality of the PCA-projected space was
obtained by the experiments. As can be seen in Fig. 7(a), when
the dimensionality reaches 16, the classification accuracy
mostly reaches the maximum. Thus, 16 was selected as the
optimal dimensionality of the PCA-projected space. For
PCA+LDA projection scheme, as Fig. 7(b) shows, the
classification accuracy reaches maximum when the
dimensionality is 18. Therefore, 18 was obtained as the
parameter for the PCA+LDA scheme.

(4)

Where ui=1/nx, (xclass i) is the mean of the i-th

sample, u=1/mxi is the mean of the total samples, m is the
number of the total samples, ni is the number of the sample i, c
is the number of the classes, and n1+n2++nc=m. LDA needs
the lower between-class coupling degree and higher
within-class polymerization degree. Thus, the Fisher criterion
is introduced:
T

For the PCA+LDA projection scheme, the matrices A

and W can be obtained sequentially in the training stage, and
then a sample X can be projected onto the PCA+LDA
combination projected space to obtain a new sample Y, as the
following equation

(5)

i 1 x k class i

J W

(7)

i 1

Sw

Y W X

(1)

The problem of solving Var(Yi) maximum value can be

transform into solving the following optimization problem:
T

m ax a1 C a1

s.t a1 a1 1

The optimal projection matrix W can be gotten by

maximizing J(W). It is easy to prove, to maximize J(W),
SBW=SWW must be met. A sample X can be projected onto
the LDA optimal discriminant space to get a new sample Y, as
the following equation

W S BW
T

W SW W

(6)

(a) The scheme of PCA

(b) The scheme of PCA+LDA

Figure 7. The classification accuracy vs the dimensionality in the

PCA-projected spaces of the PCA and PCA+LDA schemes.

The first eight groups of sEMG signals were used for

training the projected matrices of the three feature projection
schemes. The signals of nine kinds of hand/wrist motions were
segmented from the eight groups respectively. The features of
Group A were extracted from the segmented samples, and
then they were used for training the feature projected matrices.
After training, the feature projected matrices and the centers of
nine kinds of hand/wrist motions were obtained, and the

4853

clustering effects of the three projection schemes can be

observed in the projected space. Fig. 8 shows the clustering
effects of nine kinds of hand/wrist motions in the projected
spaces of three feature projection schemes respectively.

(a) PCA

(b) LDA

(c) PCA+LDA
Figure 8. The clustering effects in the projected spaces of three feature
projection schemes respectively.

The last eight groups of sEMG signals were used for test,
and the classification result of one group of signals is shown in
Fig. 9. In the figure, HC, HO, IFP, MFP, WF, WE, WRD,
WUD, RM are labeled as 1, 2, 3, 4, 5, 6, 7, 8, 9, respectively.
The burrs in the figure are the wrongly classified sample
points. The number of the wrongly classified samples was
counted and then was used for calculating the classification
accuracy. The classification accuracy and the processing time
are shown in Table III. The process of the processing time
consists of feature extraction of Group A, feature projection of
each scheme, and MDC classification. The experiments were
executed on 2.66GHz Intel(R) Core(TM)2 Quad CPU PC.

Experimental results showed that PCA had a poor

classification performance compared with the other two
schemes, and the classification accuracy of LDA and
PCA+LDA was very close to each other. PCA is a good tool
for dimensionality reduction, yet lacks of the ability of the
class separation. LDA considers the class separabillity when it
learns from the training samples to obtain a linear optimal
projected matrix. Thus, LDA had a higher accuracy than PCA.
Though PCA+LDA did not bring complex computation to the
projection process, it could not improve the performance of
the pattern recognition much. Because the three projection
schemes all obtained a projected linear matrix in the end, the
processing time of them had no big difference.
B. Comparison of four groups of features
In this section, a comprehensive comparative study on the
performance of four groups of features (see Table II) deriving
from five analyzing methods was done. The LDA feature
projection scheme was adopted for the comparative study.
Four groups of features were extracted, and then they were
projected to the LDA-projected space, and finally four groups
of the projected features were classified by the MDC
classifier, respectively. In the study, the first eight groups of
sEMG signals were used for training, and the remaining eight
groups were used for test. The clustering effects of four groups
of features in the LDA-projected space are shown in Fig. 10.
The classification accuracy and the processing time of four
groups of features are shown in Table IV. The process of the
processing time consists of feature extraction of each group,
LDA projection, and MDC classification. The experiments
were executed on 2.66GHz Intel(R) Core(TM)2 Quad CPU
PC.

(a) Group A

(b) Group B

(c) Group C

(d) Group D

Figure 9. the classification result of one group of sEMG signals

Figure 10. The clustering effects of four groups of features in the

LDA-projected space.

TABLE III.
AVERAGE VALUES OF CLASSIFICATION ACCURACY AND
PROCESSING TIME OF THREE FEATURE PROJECTION SCHEMES

TABLE IV.
AVERAGE VALUES OF CLASSIFICATION ACCURACY AND
PROCESSING TIME OF FOUR GROUPS OF FEATURES
Four Groups of features
Group A

Group B

Group C

Group D

97.50.7

Classification
Performance
Classification accuracy
[%]

97.4
0.8

95.1
1.2

94.1
1.5

96.2
0.9

0.51

Processing time [msec]

0.5

37.1

320.1

4.3

Feature Projection Schemes

Classification
Performance

PCA

LDA

PCA+LDA

Classification accuracy [%]

85.62.2

97.40.8

Processing time [msec]

0.478

0.495

4854

It is not easy to see differences between each group from

the clustering effects, because the classification accuracy of
the four groups had no big difference from each other. And all
the features deriving from five analyzing methods had a good
class separabillity on EMG pattern recognition. However, the
processing time of the four groups had a big difference from
each other. The time-domain features from Group A just
needed a very short processing time, yet the nonlinear entropy
and fractal analysis features from Group C needed a very long
processing time. The frequency-domain features from Group
B had a bearable processing time for some real-time
applications. The wavelet transform features from Group D
had a short processing time. The complexity of the features
has a dominant part of the processing time. If a feature with
complex calculation hasnt an outstanding classification
performance, it should not be selected for real-time
recognition application. The comparative results indicate that
the time-domain features and wavelet transform features are
good choices for real-time EMG pattern recognition owning to
their excellent performance both in the classification accuracy
and processing time, and it is not wise to focus excessively on
new complex features for real-time EMG pattern recognition.
Nevertheless, long time-consuming features may be good
choices for the applications without real-time need, for
example some frequency-domain features have a good
performance on muscle fatigue study based on EMG.
C. Real-time myoelectric control for anthropomorphic
robotic hand
In order to verify the performance of the proposed pattern
recognition system in real-time applications, the structure
consisting of Group A features, PCA+LDA feature projection,
and MDC classification was employed. An anthropomorphic
robotic hand which can implement the nine kinds of
hand/wrist motions was chosen. An accompany video is
available at http://ieeexplore.ieee.org, and Fig. 11 is one
screen shot of the video. Experimental result showed that the
anthropomorphic robotic hand could implement nine kinds of
hand/wrist motions well as the subject did without perceiving
a time delay for that the control circle was within an increment
window (32msec). The subject had an extended physiological
proprioception during a long time implementation.

including PCA, LDA and PCA+LDA were compared with

each other. Nine kinds of hand/wrist motions were selected,
and four channels of sEMG signals from humans forearm
were collected. Experimental results indicated that
PCA+LDA projection didnt improve the recognition
performance much compared to LDA projection, and the
time-domain features and wavelet transform features were
able to obtain a better performance on both the classification
accuracy and the processing time compared to the
frequency-domain features, nonlinear entropy features and
fractal analysis features. In this study, the features of Group A
going with PCA+LDA projection and MDC classification
obtained the classification accuracy of 97.50.7%, and just
needed the processing time of 0.51msec. The online real-time
myoelectric control for the anthropomorphic robotic hand
was implemented, and experiments showed that the
anthropomorphic robotic hand was synchronized nicely with
the humans hands without perceiving a time delay.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]

[8]

[9]
[10]
[11]

[12]
[13]

Figure 11. Online real-time control experiments

VI. CONCLUSION

[14]

In this paper, we have done a comparative study on an

EMG pattern recognition system based on PCA and LDA for
an anthropomorphic robotic hand. Four groups of features
deriving from five analyzing methods were studied
comparatively. The three feature projection schemes
4855

Otto Bock HealthCare, Product Brochure, Feb.2013, Available:

http://www.living-with-michelangelo.com/fileadmin/downloads/techn
iker/english/technician_product_brochure.pdf
Touch Bionics, Inc., i-Limb Ultra Data Sheet, Feb. 2013, Available:
http://www.touchbionics.com/media/2210/i-limb%20ultra%20data%2
0sheet%20lo-res.pdf
C. Cipriani, M. Controzzi, and M. C. Carrozza, The SmartHand
transradial prosthesis, NeuroEng. Rehabil. J., vol. 8, p. 29, 2011.
R. N. Scott, Myoelectric control of prostheses, Arch. Phys. Med.
Rehabil. J., vol. 47, no. 3, pp. 17481, Mar. 1966.
P. A. Parker and R. N. Scott, Myoelectric control of prostheses, Crit.
Rev. Biomed. Eng. J., vol. 13, no. 4, pp. 283310, 1986.
P. Parker, K. Englehart, and B. Hudgins, Myoelectric signal
processing for control of powered limb prostheses, Electromyogr.
Kinesiol. J., vol. 16, no. 6, pp. 5418, Dec. 2006.
L. Dipietro, M. Ferraro, J. J. Palazzolo, H. I. Krebs, B. T. Volpe, and N.
Hogan, Customized interactive robotic treatment for stroke:
EMG-triggered therapy, IEEE Trans. Neural Syst. Rehabil. Eng., vol.
13, no. 3, pp. 325334, Sep. 2005.
M. B. I. Raez, M. S. Hussain, F. Mohd-Yasin, Techniques of EMG
Signal Analysis: Detection, Processing, Classification and
Applications, Biological Procedures Online (BPO), vol. 8, pp. 1135,
2006.
I. Guyon, S. Gunn, M. Nikravesh, and L. A. Zadeh, Feature
Extraction: Foundations and Applications, Berlin, Heidelberg:
Springer-Verlag, 2006.
L. Hargrove, K. Englehart, and B. Hudgins, A comparison o f surface
and intramuscular myoelectric signal classification, IEEE Trans.
Biomed. Eng., vol. 54, no. 5, pp. 847853, May 2007.
F. Zhang, P. Li, Z. Hou, Y. Chen, F. Xu, J. Hu, Q. Li, and M. Tan,
SEMG feature extraction methods for pattern recognition of upper
limbs, in Proc. IEEE Int. Conf. Adv. Mechatron. Syst., pp. 222227,
Aug. 2011.
D. Zhang, A. Xiong, X. Zhao, and J. Han, PCA and LDA for
EMG-based Control of Bionic Mechanical hand, in Proc. IEEE
Int.Conf.Information and Automation, pp. 960964, June 2012.
J. Zhao, L. Jiang, H. Cai, H. Liu, EMG pattern recognition method for
prosthetic hand based on wavelet transformand sample entropy,
Control and Decision, vol. 22, no. 8, pp. 927930, Aug. 2007.
G. Naik, D. Kumar, and S. Arjunan, Use of sEMG in identification of
low level muscle activities: Features based on ICA and fractal
dimension, in Proc. Ann. Int. Conf. IEEE Eng. Med. Biol. Soc., pp.
364367, Sep. 2009.