A over damped person identification system using emg signal

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
_______________________________________________________________________________________________
Volume: 03 Issue: 10 | Oct-2014, Available @ http://www.ijret.org 54
A OVER DAMPED PERSON IDENTIFICATION SYSTEM USING EMG
SIGNAL
Suresh M1
, P G Krishnamohan2
, Mallikarjun S Holi3
1
Assistant Professor, Department of E&C, KIT, Karnataka, India
2
Retired Professor, Department of E&C, JNTUH, Hyderabad, India
3
Professor & Chairman, Department of E&I, UBDT, Karnataka, India
Abstract
EMG signal distinct from one person to another person. Body mass, muscle fiber pattern, muscle size, motion of subject,
neuromuscular activity, neurotransmitter activity in different areas within the muscle, different density of bone, changes in blood
flow in the muscle, fatigue, skin conductivity, motor unit firing pattern, motor unit paths, distribution of heat in the muscle, skin-
fat layer, motor unit recruitment order and characteristics of muscle, strength and force generated by the muscle changing from
person to person. Change in measurement of EMG signal depends on many factors. Irrespective of which factor changes the EMG
signal but physiologic dependent singularity is different and hence Electromyogram (EMG) can be used for person identification.
Useful information present in EMG signal need to be extracted to discriminate different persons. In this work, EMG signal
acquired from a single channel instrumentation system. Cepstrum technique used to extract features from EMG signal. The
Vector Quantization (VQ) is used to build person models. 50 healthy persons EMG data is recorded in three sessions at different
days. Person identification system developed using cepstrum and VQ provide good result in performance.
Keywords: Biometrics, cepstral features, feature extraction, vector quantization (VQ).
--------------------------------------------------------------------***----------------------------------------------------------------------
1. INTRODUCTION
Many parts of the body like fingers, hands, feet, ears, eyes,
faces, teeth, veins characteristics of voice, signature, styles
of typing, emotion, egos are used in biometric systems.
Physiological signal like electrocardiogram(ECG) and
electroencephalogram(EEG) can be used for person
identification[1, 2, 3]. In 1970 komi and buskirk [4] found
that the average test-retest reliability for surface electrodes
was 88%.
A train of impulses due to firing of motor units passes
through the muscle gives motor unit action potential. The
EMG signal can be modeled using
Y(n) = i(r)f(n-r) + g(n)
F(n) is impulse train, i(r) is impulse response, g(n) is
Gaussian noise and N is number of motor units. Source is
the impulse process, system is muscle, Cepstrum is used to
separate sorce and system component and then person
identification performed.
2. EMG FEATURE EXTRACTION
Cepstral Features
The time-frequency methods are used in EMG feature
extraction because of short time variability of spectrum. The
feature vectors extraction is done through cepstral
coefficients. The cepstrum is obtained by taking the inverse
Fourier transform of the log magnitude spectrum for an
EMG signal.
Consider a given EMG burst for a particular person. EMG
frames of size 50 ms (100 samples at 2 kHz sampling rate)
with a shift of 25 ms (50 samples at 2 kHz sampling rate)
are taken. Here assumption is the data within a window are
nearly stationary. 40 frames per second and therefore 400
frames for every ten second slot of EMG data obtained. In
each session five slots of ten seconds EMG data from all
subjects has been collected. So for three sessions 6000
frames per subject are obtained. For each observation
window, the first 13 basic cepstral coefficients plus 13 delta
and 13 delta-delta coefficients were used as EMG signal
features. Encouraging results obtained for 13 dimensional
base cepstrum with 13 velocity and 13 acceleration
coefficients.
3. EMG MODELING USING VQ METHOD
Person models built using the feature vectors from different
persons EMG bursts. The desirable attributes of person
model are better representation of person, must consume
less time for processing and space complexity [6]. The
objective of person identification is to obtain reliable
performance. The success of modeling techniques depends
on better cluster or captures the distribution of the feature
vectors. Earlier studies for person identification used direct
template matching between training and testing data [7, 8, 9,
10, 11, 12, 13,]. It is common to reduce the number of
feature vectors by clustering approach. The cluster centers
are known as code vectors and a set of code vectors is
known as codebook. Person identification system uses VQ
modeling has all of these attributes, which is briefly
described in this section.

_______________________________________________________________________________________________
Vector Quantization (VQ)
After obtained cepstral feature vectors, second level of
compression is done using VQ method, where a set of
nearest feature vectors are clustered together. VQ model is
formed for every person and comparison is done with the
person VQ model. The binary split algorithm is used for
obtaining codebook vectors from cepstral features, which in
turn uses the K-means clustering algorithm [14, 15, 16].
Larger codebook and EMG test data required for getting
good person identification performance. Also, VQ codebook
can be updated to alleviate the performance degradation.
The number of centroids should be at least ten times smaller
than number of feature vectors in the EMG signal.
4. RESULTS AND DISCUSSIONS
Each session contains 5 slots of 10 seconds EMG data.
There are total of 15 slots of data from all the three sessions.
First four slots of any two arbitrary sessions are used for
training. 400 feature vectors obtained for every 10 seconds
slot of EMG data. Therefore for four slots of any two
sessions data 1600×2=3200 feature vectors are used for
training. The codebooks of all persons are stored in the
EMG database. For testing, last slots (400 frames per slot)
of all sessions are used separately.
Feature vector consist of 39 dimension coefficients
including 13 dimensions of cepstral coefficients combined
with its velocity and acceleration coefficients. The 39
dimension feature vectors are trained. The result of person
identification performance for VQ technique is given in
Table 1.
Table -1: 13 dimensional base cepstra plus, 13∆ and 13∆∆
with VQ
Training
sessions
Code book
size
Testing session
1 2 3
1, 2
16 55.10 77.55 63.27
32 59.18 83.67 61.23
64 55.10 79.59 57.14
2, 3
16 44.90 69.39 85.71
32 46.94 69.39 83.67
64 4490 69.39 81.63
1, 3
16 44.90 69.39 85.71
32 46.90 71.43 83.67
64 44.90 71.43 81.63
During the course of experiment, 4 slots of EMG data from
session 1 and session 2 are used for training and no slots of
data from session 3 is used for training and remaining one
slot of data from each of the session 1, session 2 and any of
the slot of data of 10 seconds from session 3 are used for
testing. Maximum person identification performance of
83.67% is obtained for the last slot of test EMG data chosen
from trained session 2 for the 32 code book size rather than
from session 1 test EMG data. It is obvious for the reason
that the model is most recently learned with the session 2
data. It seems model is not yet generalized to yield better
performance for the last slot of test EMG data from session
1and session 3. Further, the model is not trained with
session 3 data and hence the maximum identification
performance of 63.27% is obtained for the last slot of test
EMG data selected from untrained session 3, which is less
than the person identification performance of 83.67% for the
last slot of test EMG data chosen from trained session 2.
This model is similarly trained with session 2 and session 3
EMG data without using session 1 EMG data for training. In
this case, maximum person identification performance of
85.71% is obtained for the last slot of test EMG data
selected from trained session 3 with code book size of 16.
Since the session 1 EMG data is not used for training the
model, the maximum person identification performance of
46.94% is obtained for the last slot of test EMG data
selected from untrained session 1 for the codebook size of
32 which is less than the maximum person identification
performance of 85.71% obtained for the last slot of test
EMG data selected from trained session 3 with code book
size of 16. In this case also, the model yield better
performance 85.71% for the last slot of 10 seconds of EMG
data selected from the most recently learned session 3 EMG
data.
In order to complete the experiment, the model is again
trained with session 1 and session 3 EMG data without
using the session 2 EMG data for training. In this case, the
model yields the maximum person identification
performance of 85.71% with codebook size of 16 for the
last slot of 10 seconds of test EMG data selected from
trained session 3 which is also the most recently learned
session EMG data by the model. Further, the model yields
maximum person identification performance of 71.43% with
code book size of 32 and 64 for last slot of 10 seconds of
EMG data selected from untrained session 2. This result is
again less than the result obtained from the trained session 3.
It is evident from these set of experiments that the model is
suffering from session generalization and is varying from
experiment to experiment for the test EMG data selected
from untrained sessions. In order to confirm this findings,
the experiment is again repeated with sessions used for
training is reversed and similar observation is found in Table
2.
Table -2: Cepstra-VQ with two sessions used for training is
reversed
Training
sessions
reversed
Code
book
size
Testing session
1 2 3
2,1
16 81.63 38.77 42.86
32 85.71 40.82 34.69
64 87.75 32.65 38.78
3,2
16 55.10 77.55 61.23
32 55.10 81.63 61.23
64 57.14 77.55 57.14
3,1
16 81.63 38.78 44.90
32 83.67 40.82 34.69
64 85.71 30.61 36.74

_______________________________________________________________________________________________
Fig -1: Performance of person identification system based
on13 dimensional base cepstra plus, 13∆ and13∆∆ with VQ
The Fig. 1, shows the large person identification
performance for recently trained sessions. From results, it is
well understood that the problem is in session
generalization. Further to explore, the EMG data slots are
intermixed amongst the sessions selected for training and
above experiment is repeated with an expectation of model
person identification performance improvements. For this
purpose, following mixing strategy is employed. In the
proposed mixing strategy, first slot of 10 seconds of EMG
data of session 1 and session 2 are first used for training the
model and then the second slot of 10 seconds of EMG data
of both sessions are similarly used next for training the
model and so on. In this case, after similar testing,
maximum person identification performance of 83.67% and
71.43% is obtained for last slot of 10 seconds of test EMG
data selected from session 1 and session 2 respectively with
improved session generalization amongst training session as
shown in Table 3. In this case also, the model suffers from
session generalization for test EMG data selected from
untrained session as shown in Fig.2.
Table -3: Cepstra-VQ with two sessions used for training is
intermixed
Training
sessions with
intermixed
data slots
Code
book
size
Testing session
1 2 3
1,2
16 83.67 71.43 55.10
32 79.59 71.43 46.94
64 77.55 69.40 48.98
2,3
16 51.02 85.71 71.43
32 53.10 87.76 69.40
64 53.10 85.71 71.43
1,3
16 69.39 67.40 63.27
32 77.60 65.31 71.43
64 83.67 55.10 65.31
Fig -2: Performance of person identification system with
two sessions used for training is intermixed using Cepstra-
VQ
This problem can be overcome by choosing all the three
sessions of EMG data for training. In order to explore this
fact, intermixing of slots of 10 seconds of data from all the
three sessions is employed for training. In a conventional
manner the last slot of 10 seconds of test EMG data from
each of the trained session is used for testing. In this
circumstance, the model yield maximum person
identification performance of 83.67% for a codebook size of
16 and 32 for session 1 and session 2 respectively and
65.31% for codebook size of 16 for session 3 as shown in
Table 4. Further, the same consistent person identification
performance is observed when sessions are interchanged
during training with testing procedure is repeated. It is
observed that session generalization amongst trained
sessions is improved with enhanced performance shown in
Fig.3.
Table -4: Cepstra-VQ with three sessions for training
intermixed
Training
sessions
Code
book
size
Testing session
1 2 3
1,2,3
16 83.67 81.63 65.31
32 79.59 83.67 63.27
64 73.47 79.59 63.27
2,3,1
16 83.67 81.63 65.31
32 79.59 83.67 63.27
64 73.47 79.59 63.27
3,1,2
16 83.67 81.63 65.31
32 79.59 83.67 63.27
64 73.47 79.59 63.27

_______________________________________________________________________________________________
Fig -3: Performance of person identification system with
three sessions used for training is intermixed using Cepstra-
VQ
Table 5 describes the person identification performance for
first four slots of any arbitrary chosen two sessions for
training using VQ and all the individual five slots of
remaining (untrained) session for testing separately to verify
the consistency of result for the entire slots. Promising
person identification performance has obtained for all the
slots of untrained session.
Table -5: Results for all individual slots of untrained session
The Fig.4. Shows performance of 65% to 79% with session
1, 2 trained and session 3 all slots used for testing
separately. 47%-61% performance when session 2, 3 trained
and session 1 all slots used for testing individually and 61%-
79% when session 1, 3 trained and remaining five slots of
session 2 used for testing separately. For all code book size
of VQ and slots of untrained session person identification
performance is almost similar.
Fig -4: Performance of person identification system for all
individual slots of untrained session with cepstra-VQ
To customize the EMG based person identification system
the experiment is repeated for different group of persons and
tested for up to 1024 sized code books and the results are
shown in Table 6.
Training sessions Code book
size
Testing
Slot1 Slot2 Slot3 Slot4 Slot5
1,2 16
32
64
79.59 73.47 71.43 71.44 71.43
73.47 73.47 75.51 69.39 69.39
79.59 75.51 73.47 69.39 65.31
2,3 16
32
64
53.06 46.94 57.14 57.14 57.14
53.06 48.98 59.18 61.22 53.06
46.94 51.02 53.06 57.14 51.02
1,3 16
32
64
61.22 79.59 71.43 77.55 77.53
63.26 73.47 67.35 77.55 71.42
67.35 77.55 65.31 73.46 71.42

_______________________________________________________________________________________________
Table -6: cepstra-VQ with three sessions used for training intermixed for different groups of people
The variation in performance of EMG person identification
system for different codebook size is graphically illustrated
in Fig. 5. It is evident from the graph that the performance
gradually increases from code book size of 2 and become
consistent from code book size of 64 onwards. The nature of
variation of performance with changing code book size
resembles mostly the behavior of over damped system. The
performance of EMG based person identification system
become substantially stable from code book size of 64
onwards for change in total number of persons. In order to
ensure this fact, the performance change against the
variation in total number of persons is graphically illustrated
in Fig. 6, given below. It is clearly revealed the fact that the
performance of this EMG based person identification system
substantially remains stable from codebook size of 64
onwards for change in total number of persons. Thus the
stability of EMG based person identification system that is
developed in this work demonstrated excellent in its
performance from code book size of 64 onwards and is
superior at the codebook size of 128.
Fig -5: Performance variations versus code book size for
intermixed 3 session data using cepstra-VQ.
Fig -6: Performance variations versus group of persons for
intermixed 3 session data using Cepstra-VQ.
6. CONCLUSION
An EMG based person identification system using 39
dimensional Cepstra-VQ approaches substantially remains
stable from codebook size of 64 onwards for change in total
number of persons. Thus the stability of EMG based person
identification system that is developed in this work
demonstrated excellent in its performance from code book
size of 64 onwards and is superior at the codebook size of
128.
REFERENCES
[1] M. Poulos, et al,” Neural Network based Person
Identification using EEG Features,” IEEE
International Conference on Acoustics, Speech, and
Signal Processing, volume 2: 1999, PP.1117-1120
[2] Israel. S, Irvine. J, Cheng. A, Widerhold. M,
Widerhold. B,”ECG to identify individuals”, Pattern
Recognition 38(1), 2005, PP.133-142.
Total
No. of persons
Code book size
2 4 8 16 32 64 128 256 512 1024
05 100 100 100 100 100 100 100 100 100 80
10 70 60 80 90 70 50 80 90 80 70
15 53 33 60 66 53 40 60 66 60 53
20 50 30 65 75 55 45 65 65 55 50
25 56 40 68 76 64 52 68 68 60 56
30 60 43 70 76 66 56 70 70 63 60
35 62 51 62 80 71 62 74 71 68 65
40 62 57 67 80 75 67 77 75 70 70
45 64 62 71 82 77 71 80 77 73 73
50 67 65 73 83 79 73 81 77 69 73

_______________________________________________________________________________________________
[3] L. Biel, O. Pettersson, L. Philipson, and P. Wide.
“ECG Analysis: A New Approach in Human
Identification”, IEEE Transactions on
Instrumentation and Measurement, vol. 50, no. 3,
June 2001, pp. 808 – 812.
[4] Komi.P, Buskirk.E,” Reproducibility of
electromyographic measurements with inserted wire
electrodes and surface electrodes,”
Electromyography 1970, 10:357.
[5] Suresh.M, P.G.Krishnamohan, Mallikarjun S Holi,”
Significance of EMG in Biometric Person
Authentication,” In: the Proceedings of International
Conference on Recent Innovations in Technology
(ICRIT 2011), Feb 10-12, 2011, Kottayam, pp. 113-
118.
[6] D.A.Reynolds,“An overview of automatic speaker
recognition technology,” in Proc. IEEE Int. Conf.
Acoust., Speech Signal Process, vol. 4, May 2002,
pp. 4072–4075.
[7] James E. Luck,“Automatic speaker verification
using cepstral measurements,” J. Acoust. Soc.
Amer., vol. 46(2), Nov. 1969, pp. 1026–1032.
[8] S.Pruzansky, “Pattern-matching procedure for
automatic talker recognition,” J. Acoust. Soc.
Amer., vol. 35(3), March. 1963, pp. 354–358.
[9] J.W.Glenn and N.Kleiner, “Speaker identification
based on nasal phonation ,” J. Acoust. Soc. Amer.,
vol. 43(2), June 1967, pp. 368–372.
[10] B.S.Atal, “Effectiveness of linear prediction
characteristics of the speech wave for automatic
speaker identification and verification,” J. Acoust.
Soc. Amer., vol. 55, 1974, pp. 1304–1312.
[11] M.R.Sambur,“Selection of acoustic features for
speaker identification,” IEEE Trans. Acoust.,
Speech, Signal Process., vol. ASSP-23(2), Apr.
1975, pp. 176–182.
[12] A.E.Rosenberg and M.R.Sambur, “New techniques
for automatic speaker verification,” IEEE Trans.
Acoust., Speech, Signal Process., vol. ASSP-23(2),
Apr. 1975, pp. 169–176.
[13] M.R.Sambur,“Speaker recognition using orthogonal
linear prediction,” IEEE Trans. Acoust., Speech,
Signal Process.., vol. ASSP-24(4), Aug. 1976, pp.
283–289.
[14] P.K. Soong, A.E. Rosenberg, L.R. Rabiner, B.H.
Juang “A vector quantization approach to speaker
recognition,” AT&T Tech. J., vol. 66, no. 2, 1987,
pp. 14–26.
[15] Higgins, L.Bhaler, and J.Porter, “Voice
identification using nearest neighbor distance
measure,” in Proc. IEEE Int. Conf. Acoustics,
Speech, and Signal Processing, Minneapolis, MN,
1993, pp. 375–378.
[16] Y.Linde, A.Buzo, and R.M.Gray, “An algorithm for
vector quantizer design,” IEEE Trans.
Communications, vol. COM-28(1), pp. 84–96, Jan.
1980.
[17] P.Dempster, N.M.Laird, and D.B. Rubin,
“Maximum Likelihood from Incomplete Data via
the EM algorithm,” J. Roy. Statist. Soc. B, vol. 39,
1977, pp. 1-38.
[18] E.W.Theado, G.G.Knapik and W.S.Marras,
“Modification of an EMG –assisted biomechanical
model for pulling and pushing,” in International
Journal of Industrial Ergonomics, 37(2007)825-831.
BIOGRAPHIES
M.Suresh was born in India on October 28,
1967.He received the BE degree in
instrumentation at Siddaganga Institute of
Technology, Tumkur, Karnataka and ME
degree in electronics from the BMS College of
engineering,, Bangalore, India in 1991 and 1994
respectively. From 1994 to 1999, he was a Lecturer,
onwards and currently he is Assistant Professor with
electronics and communication department, Kalpataru
Institute of Technology, Tiptur, India, pursuing PhD
programme at Jawaharlal Nehru Technological University,
Kukatpally, Hyderabad, Andhrapradesh. His research
interests include biometrics, biomedical and speech signal
processing. Prof. Suresh is a member of ISTE and member
of IETE.
Dr.P.G.Krishnamohan was born in India on
June 25, 1952. He received the B. Tech degree
in electronics and communication engineering
at Regional Engineering College, Warangal in
1975. He received the master degree in
Communication Systems at Rookee University, Roorkee,
Uttar Pradesh, India in 1977 and research degree in Signal
Processing at Indian Institute of Science, Bangalore, India in
1988. He has 36 years teaching experience, Worked as Head
of Department of Electronics and Communication
Engineering Department, JNTU, Hyderabad. he guided 30
M.Tech student project and 3 PhD students. Many research
projects carried out successfully and published 42 research
papers. He was chairman of board of studies of E.C.E
department. Currently he is professor in the department of
electronics and communication, IARE, Hyderabad. His
research interests include signal processing, signal
estimation, probability random variables, analog and digital
communications. Dr.Krishnamohan is a fellow of IETE, IE,
and member of ISTE, involved in Curriculum revision for
B.TECH and M.TECH courses from academic year 2005-
06.involved in Curriculum revision for B.TECH and
M.TECH courses from academic year 2005-06.Chairman of
Board of Studies of E.C.E department Bottom of For
Dr.Mallikarjun S Holi was Born on 15th
August 1970 at Gulbarga, completed BE in
Instrumentation Technology from PDA
College of Engineering (Gulbarga
University) in 1993. He obtained M.Tech.
and Ph.D. degree from Indian Institute
Technology, Madras in 1997 & 2004 respectively. He has
17 years of teaching experience and 8 years of Research
experience in the area of Instrumentation Technology &
Biomedical Engineering. His research interests are:
Biomechanics, Rehabilitation Engineering, Biomedical
Signal and Image Processing, Biomedical Instrumentation &
Process Control. He has published 52 research papers in

_______________________________________________________________________________________________
National & International journals and conferences. Currently
he is Professor and Chairman, Department of Electronics
and Instrumentation Engineering, University B.D.T College
of Engineering , Davangere, Karnataka guiding 5 students
for Ph.D. Programme. He is carrying out three Sponsored
Research Projects funded by AICTE and Visveswaraiah
Technological University. Dr.Mallikarjun S Holi is a life
member of Indian Society for Technical Education, Indian
Society of Biomechanics, Bio medical Engineering Society
of India, Ultrasound Society of India, Indian Association of
Biomedical Scientists. Dr.Holi has received many awards,
among prominent are:
 Institute Merit Award for the Best Academic record
in M.Tech (Biomedical Engineering) program for
the academic year 1995–1996, from I.I.T, Madras
 Received Travel Grant Award from IEEE-
Engineering in Medicine and Biology Society for
Best Research Paper in the 28th
Annual Northeast
Bioengineering Conference, Drexel University,
Philadelphia, USA, 2002.
 Received Travel Grant Award from International
Bone and Mineral Society & Japanese Society for
Bone and Mineral Research for Best Research Paper
in the First Joint Meeting of IBMS-JSBMR, Osaka,
Japan, 2003.

A over damped person identification system using emg signal

More Related Content

What's hot (13)

Viewers also liked (20)

Similar to A over damped person identification system using emg signal (20)

More from eSAT Publishing House (20)

Recently uploaded (20)

A over damped person identification system using emg signal