Grasp Recognition From Myoelectric Signals

Simon Ferguson
Department of Mechanical Engineering
University of Canterbury, Christchurch
PB4800, New Zealand.
s.ferguson@mech.canterbury.ac.nz
G Reg Dunlop
Department of Mechanical Engineering
University of Canterbury, Christchurch
PB4800, New Zealand.
r.dunlop@canterbury.ac.nz



Abstract
Current myoelectric prosthetic hands enable the
user to enact a simple grasp with a strength
proportional to the contraction of certain muscle
groups. This paper discusses the development of
a system that will allow complex grasp shapes to
be identified based on natural muscle movement.
The application of this system can be extended to
a general device controller where input is
obtained from forearm muscles, measured using
unobtrusive surface electrodes. This system
provides the advantage of being less fatiguing
than traditional input devices. The
instrumentation hardware, computer software and
algorithms developed to achieve this task are
described in this paper.

1 Introduction
Hand amputees at present have the option of several kinds
of prosthetic device. Apart from cosmetic hands or
mechanically activated grippers there is a range of
electrically driven hands available. Many of these are
activated by the user contracting a single muscle with the
resulting electromyographic signal activating the opening
or closing of a single degree of freedom device. The speed
and strength of the grip may be proportional to the signal
that is read. The work presented in this paper stems from
the desire to create a more advanced prosthetic system.
The anticipated use for the artificial hand is as a robotic
anthropomorphic device with multiple degrees of
freedom. Such a hand is being developed at the University
of Canterbury, but there are many suitable devices being
researched worldwide [Jacobsen et al., 1986; Kyberd et
al., 1998]. To control these more complex hands requires a
more advanced human-machine interface involving
multiple EMG (electromyographic) electrodes and a
suitable pattern recognition system to interpret the raw
signal. Ideally this system would require the user to
undergo minimal training and therefore the control inputs
should be as natural as possible. Muscles in the forearm
are responsible for the coarse grasp shape made by a
human hand, while muscles intrinsic to the hand are
responsible for the fine grip and movement. After a hand
amputation many of the muscles in the forearm remain
and can still be used by the amputee. EMG signals can be
read from these and used as the control source for the
prosthetic device.
The system can also be used by able-bodied individuals
as a machine interface to replace traditional controllers
such as joysticks. The advantage to this includes less
fatigue and a more natural movement. A successful
controller will need minimal training of the operator.
Many studies have looked into controllers of this type
with mixed results. Early work mostly involved simple,
single channel signal identification. Hudgins et al., [1993]
first showed that the myoelectric signal exhibited a
deterministic pattern during the early phases of muscle
contraction. Researchers have used various methods for
identifying limb motion based on EMG signals, including
multilayer neural networks [Englehart et al., 2001],
Hidden Markov Models [Jorgensen et al., 2000] and
Fuzzy ARTMAP Networks [Vuskovic and Du, 2002].
Control of an artificial hand, whether a prosthetic
device or a robot manipulator, is achieved by placement of
a number of EMG electrodes at specific locations on the
surface of the forearm. Each sensor measures the surface
potential at two points, the potential difference being
directly proportional to the amount of contraction of the
underlying muscle. In the Canterbury system information
from these sensors is sampled by a desktop PC via a USB
interface. Custom software written in C++ is used to
process this data in a Windows based environment. The
algorithms used are designed to be portable for future use
on a fast DSP when the entire system moves to a stand-
alone arrangement.
The aim of the software is to recognize the basic grasp
shape being enacted by the user, based solely on the EMG
measurements.
2 System Description
The object of the work presented here is to have a self-
contained device taking signals from EMG electrodes on
the arm and providing control signals to a prosthetic hand.
Proc. 2002 Australasian Conference on Robotics and Automation
Auckland, 27-29 November 2002
Copyright ARAA 2002 83
For the laboratory stage of the work, a standard PC
running Windows is being used. A USB interface connects
the PC with the ADCs and EMG electrodes.
Figure 1 shows the flow of data from raw signal
through to the identification of a grasp. The information
obtained from electrodes on the forearm undergoes
filtering to isolate the usable spectrum of the EMG signal.
The next phase is feature extraction, where the overall
multichannel signal is described by a smaller set of data.
This information is then used for either training or
classification purposes. When a subject uses the electrode
system for the first time they need to train the software to
recognize their muscle movements. This is accomplished
by getting them to repeatedly perform each of the six
grasp types to be identified. The data obtained during
these movements is saved for offline processing where it
undergoes multiple feature extraction techniques. The
pattern recognition phase uses the extracted features as the
training set for either a neural network or as the basis for
statistical identification.
It is accepted that not every user will get the best
results from the same feature extraction and pattern
recognition techniques. For use of the system as either a
prosthetic controller or a bioelectric interface, the user
needs to have complete confidence in the system.
Therefore successful grasp classification rates of above
90% are desired after initial software training. When the
user has had experience using the system they will also
adapt with their muscle movements becoming more
consistent thus leading to even higher recognition rates.
The algorithms used in the identification process also
have to be fast. For real-time hand control a maximum of
200 milliseconds for classification is suitable. Too long a
delay in enacting the required grasp will make the system
impractical. There is a trade-off between processing time
and classification success rate. A complicated algorithm
might be very accurate but may take too long to be
practical.
There is also the issue of muscle selection, for able-
bodied users the same muscles can be used for all subjects
but amputees will have varying degrees of muscle
removal depending on the level of amputation. In some
cases it will be necessary to choose different electrode
locations which may not provide as much information as
the typical site selection. These users will need to undergo
more extensive training to reach suitably high levels of
correct grasp classification.
3 Instrumentation
The first stage of the project was to get a signal from the
forearm muscles. This analogue signal then needed to be
sampled and made available to software on a standard
desktop PC.
3.1 EMG Electrodes
Muscle contraction is created by twitching of muscle
fibres, the level of contraction being determined by both
the number of fibres activated and the rate at which they
twitch. A single motor unit is a collection of a number of
muscle fibres activated by a single neuron. The electrical
potential of individual motor units can be measured with
fine wire or needle electrodes to determine the level of
activity present. While these electrodes are widely used in
clinical applications they are unsuitable for the practical
applications under consideration. The EMG signal as
measured at the skin surface is the spatial and temporal
sum of individual motor units within multiple muscles in
the vicinity of the recording electrode. The signal is
stochastic with a typical amplitude of 0-6mV. The usable
frequency range is 0-500 Hz with most energy
concentrated from 50-150 Hz [De Luca, 1997].
Ambient electrical noise is a serious problem as the
human body is an excellent antenna and the amplitudes of
this noise, particularly in the 50Hz range, are far greater
than the EMG signal. Unfortunately this is exactly the
area of the EMG spectrum where the most useful
information is located. For our system we found the best
results were obtained by using instrument amplifiers with
a high CMRR (120dB) and feeding the inverted common-
mode signal from each electrode pair back onto the arm
(Figure 2). This very effectively cancels out all unwanted
electrical signals, there has been no need to carry out any
50 Hz filtering after gathering the signal.


Training
Classification
Raw Multichannel
EMG Input
Low-pass
Filter
Feature Extraction
AR
STFT
Wavelets
Pattern Recognition
Neural Net
Statistical
Clustering
Grasp
Selection
Figure 1. Grasp recognition from raw EMG signal.
84
+
-
Vcm
1...4
Vcm
Vo
+
+
-
-

Figure 2. EMG instrumentation showing electrode pair and
feedback of common-mode signal to arm

For the electrodes a system using disposable
silver/silver chloride disc electrodes was initially used but
this proved inadequate for several reasons. These needed
to be applied in pairs at specific locations on the skin.
Repeatedly locating electrodes at the correct spot is
essential as even small differences in location can lead to
large changes in the resulting signal. Positioning the
disposable electrodes accurately proved a difficult task.
An electrolytic gel needed to be used with the silver/silver
chloride discs to establish electrical contact with the skin
tissue. This led to the electrodes slipping on the now wet
skin surface. This caused motion artefacts in the signal,
these appeared as large voltage spikes whenever one of
the disposable discs moved. The other major problem with
this system was that amplification of the signal was
performed some distance from the actual electrode
surface. Because the EMG signal has such a low
amplitude this meant that electrical noise induced in the
connecting cabling could significantly degrade the signal.
Active Electrodes
To counter many of these problems an active electrode
system was developed. To improve the signal-to-noise
ratio it was desired to have amplification occurring as
close as possible to the electrodes themselves. A fixed
geometry of the electrode surfaces would assist in
relocating the electrodes at the correct location. If the
electrodes could be applied dry, without any conductive
gel, this would aid in the application phase as well as
reducing the likelihood of motion artefacts appearing in
the signal.
Guidelines for electrode geometry were obtained from
De Luca [1997] and resulted in the creation of the active
electrode
1
shown in Figure 3. The electrode detection
areas consist of two gold plated rectangular surfaces on
the bottom of a small board containing the amplification
circuitry. These each measure 10mm x 2mm and are
spaced 12mm apart. These bars are placed on the skin
surface parallel to the direction of the muscle fibres.

1
De Luca defines an active electrode as an EMG sensor
where the differential amplifier is located as close as
possible to the detection surfaces.


Figure 3. Active electrodes with NZ 5c coin for scale.

As illustrated in Figure 2 each surface is connected to an
instrumentation amplifier and the output voltage (Vo) is
sampled by the computer. The common mode voltage
(Vcm) of all electrodes is fed into a summing amplifier
and the resulting inverted signal is then fed back onto the
arm. Usually a silver/silver chloride electrode is used as
the connection for this. It is placed at a point on the skin
with minimal underlying muscle, typically in the elbow
region.
Muscles
Each electrode is placed above a specific muscle in the
forearm. The system developed can handle up to eight
electrodes simultaneously. Typically four electrodes are
used, placed on extensor muscles of the upper forearm.
Vuskovic et al., [1995] use these as they are associated
with the preshaping of the hand. Extensor Pollicis Longus
and Extensor Pollicis Brevis are responsible for thumb
movement. Extensor Communis Digitorum is related to
index and middle finger motion while Extensor Carpi
Ulnaris indicates little finger activity.
This selection is suitable for able-bodied subjects but
amputees will not necessarily have all these muscles
intact.
3.2 Computer Interface
The output from each active electrode is fed into a custom
made A/D converter sampling at a rate of 2000 Hz. A USB
connection is used to enable the hardware to be easily
used with any modern desktop PC. Data is stored in a
buffer after sampling, with the computer being able to
retrieve information at any stage. This eliminates the need
for a real-time operating system.
Software
Initially all data processing was done offline using
MATLAB. EMG data was recorded straight to hard disk
for later use. More recently all programming has switched
to C++ in a Windows environment. With faster computer
hardware it has been possible to display a plot of the
incoming multichannel signals. This is a very useful aid
when locating electrodes on the users arm as it is possible
to see where the best signal is being obtained.
Once electrodes are properly positioned the display can
be switched off in order to allow the computer to perform
the task of identifying the grasp being enacted.
4 Grasp Recognition
For the control of a robot hand we aim to identify the
85
basic grasp shape being made by the user. In order to
grasp and manipulate an object the human hand will first
form a basic shape suitable for gripping the required item.
This phase is known as preshaping. Once the style of grip
being enacted is known the robot hand is commanded to
move the fingers to the correct orientation. Recognizing
this initial shape is the first task, completing the grasp can
then be accomplished using feedback sensors within the
fingers [Vuskovic et al., 1998].
4.1 Grasps
There are a number of basic grasp shapes that the human
hand can form. There are many studies each with their
own definitions of a basic set of grasps, an extensive list is
available [MacKenzie and Iberall, 1994]. We have settled
on those used by Vuskovic et al., [1995]; large and small
cylindrical, large and small spherical, pinch and key as
illustrated in Figure 4.


Spherical Pinch
Cylindrical Key

Figure 4. Basic grasp shapes

4.2 Feature Extraction
The quantity of data being gathered makes it cumbersome
to work directly with the raw EMG signal at the pattern
recognition phase. The strategy is to reduce this
information down to a smaller set of identifying
characteristics by means of various feature extraction
algorithms. The information extracted is still enough to
describe the signal, but with a smaller set of data.
Each channel of the raw EMG signal is a discrete time
series y(k). Many methods have been trialled to
characterize this time series for later pattern recognition.
Simple methods such as moving average filters and zero-
crossing rate have been looked at but only provided low
grasp recognition rates.
Frequency Domain
Looking at the signal in the frequency domain reveals
useful information which can be used to identify the
signal. The Short-Time Fourier Transform (STFT) breaks
the incoming signal into small blocks and performs an
FFT on each segment. This gives a measure of both time
and frequency information, showing at which point in
time the various frequencies are occurring. There is a
trade-off between frequency and time resolution. To
increase time resolution, smaller segments of data must be
used which decreases the resolution of the resulting FFT.
In using this technique for feature extraction, many
different segment sizes and overlapping time windows
were trialled. The resulting data set is still quite large but
does give greater recognition rates than a filtered time
series.
Parametric Modelling
The power spectrum of the EMG signal can be described
as a pole-zero model. An autoregressive (AR), or all-pole,
model was used to model the time series as:

+ =
=
P
i
i
k e i k y a k y
1
) ( ) ( ) (


Where a
i
are the AR coefficients, P is the order of the
model and e(k) an error term. The incoming signal is
divided into overlapping segments, as in the STFT, and
the AR coefficients computed for each segment. Various
window lengths and model orders have been tested. The
best results are obtained from a model of order P = 20
with a window length of 256 samples.
Wavelet Decomposition
An interesting field of mathematics which has been
applied to many different practical applications is the use
of wavelets. These provide a method of extracting time-
frequency information from a signal.
The Discrete Wavelet Tranform (DWT) is a fast, linear
operation. It is similar to the FFT in that can be viewed as
a rotation in function space that tranforms the time domain
to a different domain. With the FFT the new domain has
basis functions of sines and cosines while the DWT bases
are more complicated and known as wavelets. There are
many different families of wavelets and several of these
were investigated.
Deconvolution
One method trialed involved an attempt to deconvolve the
surface EMG signal to find the underlying train of pulses
sent from the nervous system to the muscles. It was
thought that this would provide an easily recognizable
pattern for each grasp. While this worked well for a
simulated signal its effectiveness reduced with the
introduction of even a small amount of noise. When
applied to real data it proved completely ineffective. It is
an interesting approach but is unsuitable for surface
electrodes.

4.3 Pattern Recognition
Once the incoming signals have been reduced in the
feature extraction phase the next step is to identify this
data based on pretrained information. The training phase
involves applying EMG electrodes to a subject while
getting them to repeatedly perform each of the six
different grasp types. This data is all recorded, then the
various feature extraction techniques are performed on it.
The resulting information is to be used as the training set
for either a neural network or a statistical model. This
forms the basis for identifying grasps made by the same
86
user.
There are several different options for pattern
recognition. There are many kinds of neural network
which are suitable for this application, and there are also
statistical methods. Both these types of identification have
been used with mixed success.
The statistical methods were based on those used by
Elliott, [1998] and Vuscovic [1995]. These involve
classifying an unknown vector as belonging to a cluster to
which it is closest. The clusters are based on values
obtained from the training data set. Deciding which cluster
is closest to the vector is made by selecting an
appropriate norm. The Euclidean, Mahalanobis and
Transformed Euclidean norms were tested in Elliott
[1998], with the Mahalanobis distance providing the best
results in that study. This method was used during this
project. It has proved quite effective and has some
advantages over the neural network apporach, mainly in
speed of training.
For the neural network pattern recognition a two layer
network with back propagation learning was used. During
the training the convergence of the learning process took a
substantial amount of time while offering results similar to
those of the statistical clustering. Back propogation
networks are unsuited for incremental learning, so once
trained there is not much scope for changing the network.
5 Results and Future Work
The major testing at this stage has involved data gathered
with the original EMG electrode system being processed
offline. Successful recognition of grasp varies with the
individual subject being tested, the feature extraction
technique and parameters used during this phase, and the
pattern recognition method and parameters.
Mean values of the success rates are around 75-80%
depending on which features and extraction parameters
are being used. So far there has been no one technique has
performed consistently highly for all test subjects. Some
techniques work well on certain subjects, but no better
than average on others.
For this reason it has been decided not to rely solely on
one algorithm for all users. During the training phase each
user now records a substantial amount of data. The entire
data set undergoes multiple feature extractions of different
types. A subset of these results is used to train the pattern
recognition stage, while the remaining data is used as a
test of the recognition success. Based on the results of
these tests the best performing feature extraction
algorithm is selected and used for online testing with that
user. Online testing is where the computer software
identifies the grasp as the user is making it, rather than
working with pre-recorded data. As the user becomes
more familiar with the system they will also adapt their
movement patterns. This should lead to a very reliable,
easy to use system.

Acknowledgements
The authors would like to acknowledge the technical
support of Mr Andy Cree, and Mr Rodney Elliot, scientific
officers in the Mechanical Engineering Department of
University of Canterbury, New Zealand.
References
[De Luca, 1997] Carlo J. De Luca. Surface
Electromyography Detection and Recording.
Neuromuscular Research Centre, Boston University,
Boston, Massachusetts, 1997.
[Elliott, 1998] Rodney B. Elliott. Feature Extraction
Techniques for Grasp Classification. University of
Canterbury, Christchurch, 1998.
[Englehart et al., 2001] K. Englehart, B. Hudgin and P.A.
Parker. A wavelet-based continuous classification
scheme for multifunction myoelectric control.
Biomedical Engineering, IEEE Transactions on.
48(3):302-311, 2001.
[Hudgins et al., 1993] B. Hudgins, P. Parker and R.N.
Scott. A new strategy for multifunction myoelectric
control. Biomedical Engineering, IEEE Transactions
on. 40(1):82-94, 1993.
[Jacobsen et al., 1986] S. Jacobsen, E. Iversen, D.
Knutti, R. Johnson and K. Biggers. Design of the
Utah/MIT Dextrous Hand. In Proceedings of the 1986
IEEE International Conference on Robotics and
Automation, San Fransisco, April 7-10 1986. IEEE.
[Jorgensen et al., 2000] Charles Jorgensen, Kevin Wheeler
and Slawomir Stepniewski. Bioelectric Control of a 757
Class High Fidelity Aircraft Simulation. In Hawaii,
June 11-16 2000.
[Kyberd et al., 1998] Peter J. Kyberd, Mervyn Evans and
Stefan te Winkel. An Intelligent Anthropomorphic Hand
With Automatic Grasp. Robotica. 16 :531-536, 1998.
[MacKenzie and Iberall, 1994] Christine L. MacKenzie
and Thea Iberall. The grasping hand. North-Holland,
Amsterdam ; New York, 1994.
[Vuskovic and Du, 2002] M. Vuskovic and Sijiang Du.
Classification of prehensile EMG patterns with
simplified fuzzy ARTMAP networks. In Proceedings of
the International Joint Conference on Neural Networks,
2002, 2539-2544, 2002.
[Vuskovic et al., 1998] M. I. Vuskovic, G. R. Dunlop and
K. Filali-Adib. Controlling the Power Grasp With
Incomplete Touch Sensor Information. In Proceedings
of the 1998 IEEE International Conference on Robotics
& Automation, 3639-3645, Leuven, Belgium, May
1998.
[Vuskovic et al., 1995] M.I. Vuskovic, A.L. Pozos and R.
Pozos. Classifcation of Grasp Modes Based on
Electromyographic Patterns of Preshaping Motions. In
Proceedings of the 1995 IEEE International
Conference on Systems, Man and Cybernetics, 89-95,
1995.

87