ELECTROMYOGRAPHY BASED FINGER MOVEMENT

IDENTIFICATION FOR HUMAN COMPUTER INTERFACE

Pah Nemuel D, Kumar Dinesh K

School of Electrical and Computer Engineering RMIT University GPO Box 2476 Melbourne,VIC 3001,Australia

Keywords: Surface Electromyogram, Wavelet Network, Neural Networks, and Rehabilitation.

Abstract: This paper reports experiments conducted to classify single channel Surface Electromyogram recorded from

the forearm with the flexion and extension of the different fingers. Controlled experiments were conducted

where single channel SEMF was recorded from the flexor digitorum superficialis muscle for various finger

positions from the volunteers. A modified wavelet network called Thresholding Wavelet Networks that has

been developed by the authors (D Kumar, 2003) has been applied for this classification. The purpose of this

research was towards developing a reliable man machine interface that could have applications for

rehabilitation, robotics and industry. The network is promising with accuracy better than 85%.

1 INTRODUCTION

With greatly improved computational power, and

use of computers having exploded into every walk

of life, there is a greater need for flexible, natural

and reliable human computer interface. Hand

movement gestures play a very important role in the

interactions between people. But most of the

interaction with computers is based static events

such as a key press, and the information contained in

the dynamic gesture is lost, greatly reducing the

scope of machine interaction. There is thus need for

simple and reliable methods for human hand action

identification by machines. This paper reports a new

technique for automatic recognition of human hand

movements.

Skeletal movement is caused by or prevented by

muscle contraction. Muscle contraction is a result of

electrical stimulation received from the nerves to

individual muscle fibres. The resultant electrical

activity can be recorded by electrodes kept in the

close proximity of the muscles. Surface

electromyography (SEMG) (J Cram, 1998) is the

recording of the electrical activity of skeletal muscle

from the skin surface. It is a result of the

superposition of a large number of transients

(muscle action potentials) that have temporal and

spatial separation that is semi-random.

SEMG signal is the electrical recording from the

surface and represents the summation of the

electrical activity from all the muscle fibres and thus

the summation of all Motor Unit Action Potentials

(MUAP) in the region of the electrodes. The origin

of each of the MUAP is inherently random, non-

stationary, and the electrical characteristics of the

surrounding tissues are non-linear. Distribution of

the magnitude of SEMG can be approximated by a

Gausian function (J Cram, 1998).

SEMG is used for a number of applications

including control of Human Computer Interface

(HCI), prosthesis control (Hudgins, 1993, D Graupe,

1975,F Chan, 2000), muscle diagnostic and

biofeedback. Amplitude and spectral information of

EMG have also been exploited to estimate muscle

fatigue and force of muscle contraction and torque

(K Englehart, 1999). These applications require

automated analysis and classification of SEMG. The

complexity of the signal makes this a challenging

task. The authors have reported using combination

of three channels SEMG from the forearm to

identify the hand action. The difficulty of using

multiple channels is the need for precise positioning

of the electrodes by an expert.

For automated classification of SEMG related to

movement, it is essential to develop the system that

can extract appropriate features of SEMG with

respect to the movement and have a mechanism for

relating these features to the movement generating

the signal without the need for multiple channels.

The earlier SEMG classification techniques were

based on the statistical analysis of the signal

properties (Hudgins, 1993). Auto Regressive (AR)

221

Nemuel D. P. and Dinesh K K. (2004).

ELECTROMYOGRAPHY BASED FINGER MOVEMENT IDENTIFICATION FOR HUMAN COMPUTER INTERFACE.

In Proceedings of the First International Conference on Informatics in Control, Automation and Robotics, pages 221-226

DOI: 10.5220/0001146902210226

 SciTePress

model (Graupe, 1975), of the SEMG signals

representing limb positions and were able to classify

a single channel recording with 85% success rate.

The techniques relied on the fixed thresholding

levels determined by manual inspection of the

tendency of the signal’s parameters. But this system

was highly dependent on the subject and recording

and required high degree of manual intervention.

Hudgins et al (Hudgins, 1993) reported the first

major work of SEMG classification using Artificial

Neural Networks (ANN). The ANN was used to

introduce the flexibility and self-learning ability to

the classification technique. The accuracy of the

classification technique was ranging from 80% to

90%. However, the technique was only applied to

the initial stage of the contraction. The technique is

sensitive to the window size and the appropriate

selection of the signal’s features (F Chan, 2000).

Englehart (K Englehart, 1999,K Englehart,

2001), the authors (D Kumar, 2003) and others have

reported the results of classification of SEMG

against resultant movement and muscle status using

various signal features. Some of the features

reported include time domain features, Short Time

Fourier transform (STFT), Wavelet Transform (WT)

and Wavelet Packet Transform (WPT). It has been

reported that WT and WPT were superior in the

classification SEMG against muscle status during

steady-state contraction. It has also been reported

that the technique was sensitive to the appropriate

selection of signal features to be included in the

classification and that the technique could be

improved by including the adaptive feature selection

process.

The Thresholding Wavelet Network (TWN) has

been developed by the authors (D Kumar, 2003) and

has been applied in this paper for SEMG

classification. This combines the WT, ANN and

wavelet thresholding. The combination enables the

network to extract time-scale features from the

signal and adaptively select the appropriate features

for the classification task. This paper reports the

architecture of the network and the network’s

performance in the classification of SEMG recorded

for various finger movements.

This paper is organised into six sections. Section

2 reviews the basic concept of WT and wavelet

networks while Section 3 details the architecture and

learning process of the TWN. Section 4 presents the

experimental method and results while Section 5

discusses the results. Section 6 concludes the paper.

2 THEORETICAL BACKGROUND

A. Wavelet Networks

Wavelet network (S Mallat, 1999) is a class of ANN

(D Kumar, 2003, M Hagan, 1996) that includes WT

in its algorithm. The combination provides a tool

that can calculate wavelet coefficient in parallel

mode and adaptively select the proper wavelet

coefficients for the approximation or classification

task. The wavelet networks are grouped in two

different categories:

Approximation Wavelet Networks (AWN):

Wavelet networks are designed for the purpose of

function approximation or representation and

iteratively generate the wavelet coefficients as well

the inverse wavelet transforms to approximate the

signal.

Classification Wavelet Networks (CWN):

Wavelet networks that are designed for the purpose

of function classification. These are based on

computing and determining the wavelet coefficients

to classify the signal.

AWN optimises its parameters based on a cost

function that is sensitive to energy content. The

approximation is only useful in a classification task

if the distinguishing factors are immersed in the

high-energy region of the signal. But the features

that discriminate a class of SEMG signals from the

other classes are not necessarily immersed in the

high-energy region of the signal.

The basic principle of CWN is to iteratively

locate wavelet coefficients (scale and translation)

Figure 1: A Thresholding Wavelet Network constructed by six

thresholding nodes (three for each scale).

ICINCO 2004 - ROBOTICS AND AUTOMATION

222

that contrast the difference between signals of

different classes, while enhancing the commonality

between signals of the same class. But these

networks are temporal dependent and sensitive to the

time location of significant events in the input signal

such as the singularity points. This limits its

application in SEMG signals where the action

potentials are semi-randomly located (J Cram,

1998).

3 THRESHOLDING WAVELET

NETWORKS

The authors have introduced a new type of wavelet

network, the thresholding wavelet network (TWN)

(D Kumar, 2003). This is suitable for classifying

signals such as SEMG. The wavelet coefficients are

thresholded with an upper and lower bound. TWN

selects wavelet coefficients that identify a signal

based on magnitude for the relevant scales making it

is less sensitive to the time location of the

coefficients while more sensitive to signal features

(instantaneous frequency and singularity)

represented by the magnitude of wavelet

coefficients. An example of the network is shown in

Figure 1.

The TWN consist of four blocks of network

layer: a wavelet layer, maxima layer, thresholding

layer and neural networks layers (figure 1).

The input signal x(n) is applied to the wavelet layer,

the output is the magnitude of wavelet coefficients

|Wf(s,σ)|. The wavelet maxima layer selects wavelet

coefficients that are locally maxima.

The input to the thresholding layer is the wavelet

maxima at each scale of interest. The thresholding

levels (θ

,θ

) for one scale are the same. The scale-

dependent thresholding levels allow the network to

apply different thresholding levels to different

scales. The output of wavelet thresholding node ϕ is

the number of wavelet maxima with magnitude

between θ

and θ

The bias β determines the centre of g(x) while α

determines the width of the function. The

combination of α and β determines the upper and

lower thresholds. Figure 2 illustrates the

thresholding wavelet node.

The network is a supervised learning algorithm

based system. The parameters of this network (

and neural network weights and biases) are

initialised with random values. During the iterative

learning process, the values of these parameters are

changed to reduce the classification error using input

and target output examples. The cost function used

is the sum-squared error (SSE), the difference

between target output

and the actual output

)(

ϕϕ

−=

E (1)

In each thresholding node, the learning is the process

to locate the lower and upper threshold levels (

and

) of each node. Determining the optimum

value of the threshold parameters ensures number of

wavelet coefficients with magnitude

<|Wf(s,

)|<

can best categorise the class of input signals.

The change of the parameters for each iteration

is determined using the gradient descent algorithm

(equations 9 and 10), and the learning rate

coefficient

old

oldnew

SSE

ραα

∂

−=

(2)

old

oldnew

SSE

ρββ

∂

−=

(3)

The learning process is repeated until the sum-

squared-error SSE falls below a predefined

maximum error E

. At this stage the network is

considered as able to classify the training pattern

with an error less than E

4 EXPERIMENTS

The aim of this study was to determine the possible

use of single channel SEMG from the forearm to

Wf(u

i,1

)

-α

(

)

Wf(u

i,3

)

Figure 2: The Wavelet thresholding node.

ELECTROMYOGRAPHY BASED FINGER MOVEMENT IDENTIFICATION FOR HUMAN COMPUTER

INTERFACE

223

identify the various movements of the fingers.

Towards this aim, the TWN has been employed to

identify the hand gestures by classifying the SEMG

signal based on the difference in shape and

amplification of action potentials due to the

proximity of muscle fibers to the surface electrodes.

This problem has three levels of complexity; (i)

where all the fingers move together, (ii) where two

fingers move together and (iii) where each of the

finger are independent. As the first level of

complexity may be considered as trivial, two sets of

controlled experiments were conducted.

Single channel SEMG was recorded from the

flexor digitorum superficialis muscle using BIOPAC

System EMG100C at 2000 Hz sampling rate. The

SEMG recording system had HPF at 10 Hz, LPF at

1000 Hz and a notch filter at 50 Hz to eliminate

power-line interference and with gain of 2000. Three

male volunteers were tested on three separate

occasions.

The experiments were repeated and the

network was trained with 10 signals from each class.

The length of each signal for training purposes was

500 samples. TWN used had 500 nodes at its

wavelet layer, 498 nodes at maxima layer, 8 nodes at

its thresholding layer and 100 nodes at its hidden

layer. The TWN used Db2 wavelet at scale 1. The

experimental results are shown in Table 1.

The first experiment had two finger positions

classes. Class A signals were recorded when the

middle and ring fingers were flexed, while class B

signals were recorded when the index and little

fingers were flexed (Figure 3). These were selected

because the tendons of middle and ring fingers are

superficial compare to the tendons of index and little

fingers (J Cram, 1998, N Palastanga, 1994). The

muscle fibers of middle and ring fingers are more

superficial then the fibers of distal tendons. Thus the

shape and magnitude of recorded action potentials

when flexing the middle and ring fingers are

different to the action potentials when flexing the

index and little fingers. The results are presented in

Table 1.

The second set of experiments involved the

classification of SEMG signals for four different

fingers (flexion and extension). The finger positions

for the experiments are shown in Figure 4, 5 and 6,

while the experimental results are shown in Table 2,

3 and 4.

(a)

(b)

Figure 3: The finger positions for experiment 2.

a) The finger position when class A signals were

being recorded. b) The finger position when class

B signals were being recorded.

(a)

(b)

(c)

(d)

Figure 6: The finger positions for experiment 5.

(a)

(b)

(c)

(d)

Figure 4: Finger positions for experiment 3.

(a)

(b)

(c)

(d)

Figure 5: The finger positions for experiment 4.

ICINCO 2004 - ROBOTICS AND AUTOMATION

224

Table 1:

Table 2: The Classification Error of Experiment 3

Class A Class B Class C Class D Average

Subject 1 10.0% 6.67% 20.0% 6.67% 10.83%

Subject 2 13.33% 30.0% 13.33% 0% 14.16%

Subject 3 16.67% 26.67% 33.33% 0% 19.16%

Average 13.33% 21.11% 22.22% 2.22% 14.72%

Table 3: The Classification Error of Experiment 4

Class A Class B Class C Class D Average

Subject 1 0% 6.67% 6.67% 46.67% 15.0%

Subject 2 0% 6.67% 40.0% 6.67% 13.33%

Subject 3 40.0% 13.33% 0% 0% 13.33%

Average 13.33% 8.89% 15.56% 17.78% 13.89%

Table 4: The Classification Error of Experiment 5

5 RESULTS AND DISCUSSION

The results of the experiments are tabulated in tables

1 to table 4. From these tables it is observed that

single channel SEMG when classified using the

magnitude of the wavelet coefficients gives high

level of accuracy, ranging from 93% to 85%.

From the results it is also observed that the

classification performance using the TWN decreases

as the complexity increases (number of classes

increases). The error for the classification of two

classes of signal was 7 %, while the error for the

classification of four classes ranged between 12%

and 14%.

The experiments confirmed the effectiveness of

using wavelet transform in feature extraction stage

of the classification process. The TWN could extract

wavelet’s time-scale features of input signal and

adaptively select the proper features necessary for

the classification through wavelet thresholding

mechanism. The thresholding mechanism eliminates

the need for manual feature selection process. The

network initialisation did not require the priori

knowledge of the signal to be considered.

The experiments also demonstrated the efficacy

of TWN to classify SEMG signals recorded during

low-level, steady state contractions. All the SEMG

signals used in the experiments were recorded when

the fingers were bended with a minimal needed

contraction. This advantage enables the TWN to be

applied in SEMG classification of natural finger

movement where the contraction level is minimal.

Also this classification technique can be applied in

the system that responds to steady state contraction

rather than the transient of contraction as in Hudgins

network (Hudgins, 1993).

6 CONCLUSION

This paper presents a new technique where single

channel SEMG from the flexor digitorum

superficialis muscle is used to accurately determine

the movement (flexion and extension) of the

individual fingers. The authors have used a wavelet

network that has been developed by them (D Kumar,

2003). The network classifies SEMG signals by

extracting time-scale features with wavelet

transform, and adaptively adjusts its thresholding

level during its learning process to select wavelet

maxima with certain magnitude that characterised

the input signals.

The experimental results of the SEMG

classification using TWN are extremely promising.

From the results, it is observed that:

a) This technique provides high accuracy of

classification, accuracy ranging from 93% to

85%.

b) The accuracy of increases as the number of

signal class decreases.

c) The TWN can be applied to classify low

contraction level SEMG signals. This advantage

allows the network to be applied in SEMG-

based finger posture classifier and may find

applications for other tonic muscle contractions

such as muscles of the back.

d) The TWN is less sensitive to the window size.

REFERENCES

Cram, J.R., G.S. Kasman, and J. Holtz, (1998),

Introduction to Surface Electromyography.

Gaithersburg, Maryland: Aspen Publishers, Inc.

Class A Class B Average

Subject 1 16.67% 0% 8.33%

Subject 2 0% 10.0% 5.0%

Subject 3 16.67% 3.33% 10.0%

Average 11.11% 4.44% 7.78%

Class A Class B Class C Class D Average

Subject 1 0% 13.33% 0% 0% 3.33%

Subject 2 33.33% 6.67% 20.0% 13.33% 18.33%

Subject 3 0% 20.0% 33.33% 13.33% 16.67%

Average 11.11% 13.33% 17.78% 8.89% 12.78%

ELECTROMYOGRAPHY BASED FINGER MOVEMENT IDENTIFICATION FOR HUMAN COMPUTER

INTERFACE

225

Chan, F.H.Y., et al., (2000) Fuzzy EMG Classification for

Prosthesis Control. IEEE Transaction on

Rehabilitation Engineering. 8(3): p. 305-311.

Englehart, K., et al., (1999) Classification of the

Myoelectric Signal using Time-Frequency based

Representation. Medical Engineering and Physics. 21:

p. 431-438.

Englehart, K., B. Hudgins, and P.A. Parker, (2001) A

Wavelet-Based Continuous Classification Scheme for

Multifunction Myoelectric Control. IEEE Transaction

on Biomedical Engineering,. 48(3): p. 302-311.

Hagan, M.T., H.B. Demuth, and M. Beale, (1996) Neural

Network Design., Boston: PWS Publishing Company

Hudgins, B., P. Parker, and R.N. Scott, (1993) A New

Strategy for Multifunction Myoelectric Control. IEEE

Transaction on Biomedical Engineering,. 40(1): p. 82-

94.

Graupe, D. and W.K. Cline, (1975) Functional Separation

of SEMG Signals via ARMA Identification Methods for

Prosthesis Control Purposes. IEEE Transaction on

Systems, Man, and Cybernetics. SMC-5 (2): p. 252-

259.

Kumar D, Pah N D; (2003) Thresholding Wavelet

Networks for Signal Classification: International

Journal of Wavelets, Multiresolution and Information

Processing (IJWMIP) (No: 02-10-25, 2003).

Kumar D, Pah N D; (2003) Wavelet Analysis of Surface

Electromyography to Determine Muscle Fatigue:

IEEE Transactions on Neural Systems and

Rehabilitation Engineering, pp 400- 406.

Mallat, S., (1999) A Wavelet Tour of Signal Processing.,

London: Academic Press.

Palastanga, N., D. Field, and R. Soames, (1994)

Anatomy and Human Movement., Oxford:

Butterworth-Heinemann Ltd

Szu, H.H., B. Telfer, and S. Kadambe, (1992) Neural

Network Adaptive Wavelets for Signal Representation

and Classification. Optical Engineering,. 31(9): p.

2907-1916.

ICINCO 2004 - ROBOTICS AND AUTOMATION

226