Iaetsd recognition of emg based hand gestures

RECOGNITION OF EMG BASED HAND GESTURES
FOR PROSTHETIC CONTROL USING
ARTIFICIAL NEURAL NETWORKS
Sasipriya.S, Prema.P
Department of Biomedical Engineering,
PSG College of Technology, Coimbatore.
email id: sashpsg@gmail.com, ppr@bme.psgtech.ac.in
Abstract – EMG (Electromyography) is a
biological signal derived from the summation of
electrical signals produced by muscular actions.
This EMG can be integrated with external
hardware and control prosthetics in
rehabilitation. Pattern recognition plays an
important role in developing myo-electric control
based interfaces with prosthetics and Artificial
Neural Networks (ANN) are widely used for such
tasks. The main purpose of this paper is to
classify different predefined hand gestured EMG
signals (wrist up and finger flexion) using ANN
and to compare the performances of four
different Back propagation training algorithms
used to train the network. The EMG patterns are
extracted from the signals for each movement and
then ANN is utilized to classify the EMG signals
based on their features. The four different
training algorithms used were SCG, LM, GD and
GDM with different number of hidden layers
and neurons. The ANNs were trained with those
algorithms using the available experimental data
as the training set. It was found that LM
outperformed the others in giving the best
performance within short time elapse. This
classification can further be used to control
devices based on EMG pattern recognition.
Keywords - Electromyography, Myo-electric
control, Artificial Neural Network, Back
propagation algorithms, Pattern recognition.
I. INTRODUCTION
Congenital defects or accidental loss of
limbs can be corrected by the use of artificial limbs
or prostheses. The most efficient way of controlling
the prosthesis is by the use of EMG signals
obtained from the active muscles. EMG signals
obtained by different actions can be used to make
the prosthesis perform those different functions in
real time. For doing so, the EMG signal obtained is
filtered, windowed and after digital conversion,
based on the features extracted, it is classified and
given as control inputs to a prosthetic system for
activating correct functions. The classifier is an
important element in this system. Artificial neural
networks (ANN) are mathematical modeling of
biological neuronal systems and they are
particularly useful for complex pattern recognition
and classification tasks that are difficult for
conventional computers or human beings. The
nonlinear nature of neural networks, their ability
to learn from their environments in supervised as
well as unsupervised ways, as well as their
universal approximation property make them
highly suited for solving difficult signal processing
problems. But, it is critical to select the most
appropriate neural network paradigm that can be
applied for specific function. Artificial neural
networks based on Multi-Layer Perceptron Model
(MLP) are most commonly used as classifiers for
EMG pattern recognition tasks and selecting an
appropriate training function and learning
algorithm for a classifier is a crucial task. The time
INTERNATIONAL CONFERENCE ON CURRENT TRENDS IN ENGINEERING RESEARCH, ICCTER - 2014
INTERNATIONAL ASSOCIATION OF ENGINEERING & TECHNOLOGY FOR SKILL DEVELOPMENT www.iaetsd.in
115
ISBN: 378-26-138420-01

delay between the command and activation of
prosthetic function should be minimum (not more
than 100ms) for the comfort of the users [1], [2], [3],
[4]. Hence the classifier should be precise and
rapid in producing correct signals for controlling
the prosthesis, inspite of inaccuracies that may
occur during the process of detection and
acquisition of EMG signals. So it is necessary to
assess the impact of neural network algorithms on
the performance, robustness, and reliability aspects
and to identify those that work the best for solving
our problem of interest. This work has attempted
to classify EMG based on two predefined hand
gestures and to identify the better performing
training functions in Feed Forward
Backpropagation algorithm (standard strategy of
MLP) used in recognizing the patterns.
II. .METHODOLOGY
A. EMG SIGNAL ACQUISITION AND
FEATURE EXTRACTION
EMG signals used in this study are acquired
from the muscles of the forearm namely Flexor
Carpi Ulnaris (FCU), Extensor Carpi Radialis
(ECR) and Extensor Digitorum (reference) for two
types of hand movements-finger flexion and wrist
up. FCU assists in wrist flexion with ulnar
deviation and ECR assists in extension and radial
abduction of the wrist. The myoelectric signals are
acquired by means of single channel differential
electrodes (disposable Ag/AgCl surface electrodes)
which are then amplified and filtered before
further processing.
13 different statistical features extracted
from the acquired EMG signals are Integrated
EMG, Mean Absolute value, Modified Mean
Absolute value 1, Modified Mean Absolute value
2, Simple square integral(energy), Variance, Root
Mean Square, Waveform length, Zero Crossing,
Slope sign change, Willison amplitude, Difference
Absolute Mean Value and Histogram of EMG.
These 13 features are extracted from 150 samples of
EMG signals in which 75 represented finger flexion
and the rest represented wrist up functions
obtained from volunteers and inputted to the
neural network for classification.
B. NEURAL NETWORK ARCHITECHTURE
A neural network is a general mathematical
computing paradigm that models the operations of
biological neural systems [4]. Weights (which are
determined by the training algorithm), bias and
activation function are important factors for the
response of a neural network. A Multi-Layer
Perceptron model based on Feed forward
Backpropagation algorithm is used here for
classification. Fig 1 shows the basic architecture of
a neural network with hidden layers between the
input and the output layers.
Fig 1: Neural network with hidden layers.
The designed ANN for EMG pattern
recognition consists of 3-layers: input layer, tan-
sigmoid (standard activation function) hidden
layer and a linear (purelin) output layer. Each layer
except input layer has a weight matrix, a bias
vector and an output vector. The learning rule for
the propagation of neural network defines how the
weights between the layers will change [5], [6].
Here, the input is a 13 x 150 matrix and the
corresponding target is a 2 x 150 matrix, (as 13
features extracted from 150 samples and there are 2
types of gestures to be classified). The classification
was divided into 3 stages: training (70% of
samples), validation (15% of samples) and test
116
ISBN: 378-26-138420-01

(15% of samples). Four main training algorithms
used were Scaled Conjugate Gradient (SCG),
Levenberg-Marquardt (LM), Gradient descent
(GD), and Gradient descent with momentum
(GDM). The sample input vectors, their
corresponding target vectors and the output
vectors after classification are shown in Table.1.
III. RESULTS AND DISCUSSION
The classification is done by altering the
training functions, learning rate (GD and GDM),
and number of neurons in a hidden layer with a
standard performance function MSE (Mean Square
Error) and the better performing algorithm is
identified. The summary of the results of various
training algorithms used and the corresponding
number of neurons is listed in Table.2. It is found
that, on training the network with 10 neurons, SCG
gives the least error and best performance. But on
considering the fast convergence (which is the
actual need for prosthetic control), LM with 20
neurons gives good classification performance
earlier than SCG. If number of neurons is increased
to 30, performance of LM is better than SCG, with
a least time elapse. Also the classification rate of
SCG algorithm is saturated at 81.3 % after 30
neurons, no matter how many neurons increased,
the rate of correct classification remains constant.
But higher the neuron number, better the
classification rate in LM algorithm. The other two
algorithms, GD and GDM require more time and
number of iterations needed for classification.
Their performances are also undesired. Hence the
classification efficiencies of SCG and LM
algorithms alone are shown in Table 3 and Table 4
respectively.. The best performance and
classification rate of LM based ANN with 50
neurons are also shown in Fig 2 and Fig 3. Thus the
LM network outperforms the other algorithms by
providing the least response time (fastest
convergence), higher classification rate, less
number of iterations required and less error values.
Fig 2: Best Performance of LM
Fig 3: Classification Rate of LM
IV. CONCLUSION AND FUTURE WORK
This work has been carried out to find the
appropriate classifier for EMG signals. The
experimental results show that the Levenberg-
Marquardt algorithm gives good and fast
performance with least computations. The main
disadvantages of this algorithm are requirements
of large memory and increased number of neurons
which may affect the hardware design (DSP
processors) of the prosthetics. Hence the future
will be to optimize the performance and provide
interface to the hardware.
117
ISBN: 378-26-138420-01

. Fig 4: LM Training tool
Table.1 : SAMPLE INPUT DATA AND SIMULATED OUTPUT DATA
Sample 1 Sample 2 Sample 3
Input vectors 13 extracted features
from EMG
257.92 318.26 154.49
0.42987 0.53044 0.25749
0.32246 0.39805 0.19358
0.20606 0.28215 0.1467
643.08 695.51 425.03
1.0647 1.1527 0.70138
0.80911 0.88938 0.58973
67.429 88 26.857
89.086 115.55 46.511
0.14872 0.1929 0.077647
6.5714 9.1429 0.85714
67.571 68 73.857
116.29 111.86 108.86
Target
vectors
Wrist up 1 1 0
Finger flexion 0 0 1
Output
vectors
Wrist up 0.7866 0.9530 0.3495
Finger flexion 0.2339 0.0411 0.6390
118
ISBN: 378-26-138420-01

Table.2 : COMPARISON OF VARIOUS ALGORITHMS
Table.3 : SHOWING THE CLASSIFICATION RATE SATURATION OF SCG
Table.4 : SHOWING THE CLASSIFICATION RATE EFFICIENCY OF LM
GDM GD LM SCG HIDDEN
NEURONS
BEST
PERFORMANCE
0.15275 0.14585 0.18598 0.14172 10
0.27954 0.16829 0.0991680 0.13132 20
0.12168 0.1526 0.14919 0.17166 30
EPOCH AT
WHICH BEST
PERFORMANCE
IS GOT
1000 1000 5 21 10
1000 968 9 20 20
1000 1000 7 27 30
HIDDEN NEURONS 10 20 30 40 50
CLASSIFICATION
RATES
(PERCENTAGE)
Correctly classified 76.7 80 81.3 81.3 81.3
Misclassified 23.3 20 18.7 18.7 18.7
BEST PERFORMANCE 0.17323 0.10074 0.22587 0.14376 0.13385
EPOCH AT WHICH BEST PERFORMANCE
OBTAINED
11 12 19 15 6
HIDDEN NEURONS 10 20 30 40 50
CLASSIFICATION
RATES
(PERCENTAGE)
Correctly classified 78.7 76 84 87.3 88
Misclassified 21.3 24 16 12.7 12
BEST PERFORMANCE 0.20428 0.22699 0.14131 0.30642 0.20215
EPOCH AT WHICH BEST
PERFORMANCE OBTAINED
2 2 7 4 10
119
ISBN: 378-26-138420-01

REFERENCES
[1] Alcimar Soares et al., ‘Development of virtual
myo electric prosthetic controlled by EMG pattern
recognition system based on neural networks’ ,Journal
of Intelligent Information Systems,2003
[2] Md. R. Ahsan,Muhammad, I. Ibrahimy,
Othman O. Khalifa, ‘EMG Signal Classification for
Human Computer Interaction: A Review’ , European
Journal of Scientific Research, 2009.
[3] Md. Rezwanul Ahsan, Muhammad Ibn
Ibrahimy, Othman O. Khalifa, ‘Electromyography
(EMG) Signal based Hand Gesture Recognition using
Artificial Neural Network (ANN)’, 4th International
Conference on Mechatronics (ICOM), 17-19 May
2011, Kuala Lumpur, Malaysia,
[4] Yu Hen Hu, Jenq-Neng Hwang,’ Handbook of
neural network signal processing’, ‘‘Introduction to
Neural Networks for Signal processing’’, CRC PRESS,
2001.
[5] A. Ghaffari et al., ‘Performance comparison of
neural network training algorithms in modeling of
bimodal drug delivery’, International Journal of
Pharmaceutics, 2006.
[6]. Claudio Castellini, Patrick van der Smagt,
‘Surface EMG in advanced hand prosthetics’,
. Biol Cybern , 2009.
120
ISBN: 378-26-138420-01

Iaetsd recognition of emg based hand gestures

More Related Content

What's hot (20)

Viewers also liked (15)

Similar to Iaetsd recognition of emg based hand gestures (20)

More from Iaetsd Iaetsd (20)

Recently uploaded (20)

Iaetsd recognition of emg based hand gestures