A Robust and Adaptive Algorithm for Real-time Muscle Activity Interval

Detection using EMG Signals

Rabya Bahadur

1,2

and Saeed ur Rehman

Department of Electrical & Computer Engineering, Center for Advanced Studies in Engineering, Islamabad, Pakistan

Department of Electrical Engineering, COMSATS Institute of Information Technology, Abbottabad, Pakistan

Keywords:

Electromyogram, Action Potentials, Muscle Activation Interval, Onset/Offset Detection.

Abstract:

Detection of Muscle Activity Interval plays a pivotal role in the design and implementation of real-time My-

oelectric controlled devices and their applications. This paper presents an algorithm for real-time detection

of onset/offset points in the muscles activity by employing adaptive threshold technique on the Correlation

Coefﬁcient of Taeger Kaiser Energy Operator using low cost hardware. Performance of the algorithm has also

been evaluated through real-time tests carried out under various constrained scenarios and different signal to

noise ratios, revealing very promising results with a maximum accuracy of 99.9% using medium or no external

forces.

1 INTRODUCTION

In Neuro-rehabilitation, body signals are used as a

mean of controlling the assistive devices. The most

commonly used body signals for this purpose are ob-

tained through surface Electromyogram (sEMG). In

this technique, EMG signals are retrieved from the

skin surface using electrodes on the desired muscles

(non-invasively), which represent a convolution of the

electrical activity of the muscle ﬁbers known as Motor

Unit Action Potential (MUAP) (Farina et al., 2014)

and are described using eqn. 1.

emg(t) =

∑

i=1

∞

∑

j=−∞

(t)δ(t − t

i j

) (1)

Where N is the number of active muscle ﬁber sig-

nals known as Active Motor Neurons, ψ

(t) represents

the waveform of the i

motor unit, δ(t) is the impulse

function and t

i j

is the time delay of the i

motor unit.

These signals are generally used to control the speed

of a motor or position of an actuator in prosthesis.

Therefore, the accurate detection of the points where

muscles initiate to move from their relaxed states is of

great signiﬁcance in implementing real-time control

of prosthesis. The triggering of the muscular contrac-

tion is known as the onset point, while the end point of

the contraction is referred to as an offset point. These

onset/offset represents the duration of a muscle activ-

ity.

For neuro-rehabilitation active control of an external

device with maximum precision and minimum pro-

cessing time is a major requirement. Currently avail-

able myo-controlled prosthesis such as i-limb and Ot-

tobock utilize state machine based controller for se-

lection of movements (Farina et al., 2014). In such

systems real-time detection of these onsets/offsets

play a vital role. A great deal of work has been carried

out in this regard during the past two decades. (Merlo

et al., 2003) used amplitude threshold on continuous

time wavelet transform for the detection of presence

of MUAP . The method proposed is unsuitable for real

time activity detection because its accuracy depends

upon the shape of window selected which may vary

for different movements; in addition, the proposed

method is computationally intensive. (Solnik et al.,

2010) refer to the use of Teager-Kaiser Energy Op-

erator (TKEO) for the use of EMG detection, which

seems to be a very promising parameter . However,

they have used a threshold of 3 standard deviations

for 25 consecutive samples which makes the over-

all algorithm non-adaptable especially for scenarios

where signal variations due to changes in environ-

mental/physical conditions are experienced; this as-

pect is further discussed in section (6). The two stage

methodology proposed by (Drapala et al., 2010) is

based on initial estimation from the entire EMG fol-

lowed by local estimation. The ﬁrst stage proposes

energy estimation of sEMG signal using the TKEO

algorithm. The estimated energy signal acts as input

Bahadur, R. and Rehman, S.

A Robust and Adaptive Algorithm for Real-time Muscle Activity Interval Detection using EMG Signals.

DOI: 10.5220/0006536200890096

In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2018) - Volume 4: BIOSIGNALS, pages 89-96

ISBN: 978-989-758-279-0

to Expectation Maximization (EM) clustering method

for the classiﬁcation of EMG activity region vs. non

activity region . Once the active and non-active re-

gions are vividly separated they are further enhanced

using a single threshold algorithm. As the algorithm

needs complete signal for analysis and comparison,

therefore, their work is presented for ofﬂine analy-

sis, focusing more on precision as compared to com-

putational complexity. (Xu et al., 2013) present an

adaptive approach for onset/offset detection . Their

work merges Maximum Likelihood (ML) algorithm

with a look-up table, comprising of a set of thresh-

old values based on different Signal to Noise ratios

(SNR). The algorithm is not computationally inten-

sive and, therefore, can be adopted for real time signal

characterization but the major drawback is the use of

simulated data combined with Additive White Gaus-

sian Noise (AWGN) at different SNRs (a mean for

generating the look-up table) which cannot be mod-

eled for a real time EMG signal. Further investiga-

tions for onset/offset detection have also been carried

out using machine learning algorithms such as the

use of Gaussian Mixture Model and Hidden Markov

Model by (Liu et al., 2015a; Liu et al., 2015b; Naseem

et al., 2016). However, all such attempts are lim-

ited to ofﬂine sEMG and cannot be adopted for real

time analysis due to their computational complexity.

In this paper, the onset/offset points resulting in the

subsequent Muscle Activity Interval (MAI) are deter-

mined using temporal and statistical features for the

real time sEMG signals. The run-time performance of

the proposed algorithm has been evaluated using real-

time EMG signals under various constrained scenar-

ios. The algorithm was tested on signals from seven

healthy subjects using same experimental procedure

as discussed in section (2).

2 EXPERIMENTAL PROCEDURE

2.1 Testing Protocol

Seven healthy subjects volunteered in this study in-

volving three females and four males. All experi-

ments were approved by campus bioethics commit-

tee and written consent was also taken from each par-

ticipant. During these experiments, the elbow joint

was ﬁxed while wrist ﬂexion and extension were per-

formed with 0

◦

rest position. In order to check the

robustness of the proposed algorithm, three kinds of

exercises were performed by each participant.

i. Wrist ﬂexion and extension in the absence of any

external force acting on the wrist.

ii. Wrist ﬂexion and extension in the presence of an

opposing force acting on the wrist .

iii. Wrist ﬂexion and extension while holding 1 kg, 2

kg, and, 5 kg weights, respectively.

In order to further investigate the accuracy of

the proposed algorithm, AWGN was added to the

recorded signals to generate different SNRs of 1.25, 3,

6, and 9 dBs. All ofﬂine simulations were performed

using MATLAB 2015a while all online results were

obtained using NI LabVIEW 2015.

2.2 Experimental Setup

Bipolar pre-gelled Ag/AgCl disposable surface elec-

trodes were placed on the left arm at ﬂexer carpi

ulnaris and radialis muscles, as shown in Fig. 1,

following the recommendations of Surface Elec-

tromyography for Non-Invasive Assessment of Mus-

cles(SENIAM) (Hermens et al., 2000). The refer-

ence electrode was in particular placed on medial epi-

condyle of elbow joint which was kept at rest dur-

ing the entire experimentation process. This paper

presents results of a pilot study initiated at CASE,

Pakistan, to investigate the use of a low-cost hard-

ware for online detection of MAI. For this purpose, a

2 channel Olimex EKG/EMG bio-feedback shield was

used with Arduino Uno R3, despite its lower acqui-

sition signal bandwidth of 0.16Hz to 40Hz which is

much less than the required 500 Hz, for sEMG signal.

The main idea in this pilot study was to explore the

viability / suitability of information contents present

in lower frequency components of sEMG signals, for

estimation of MAI. sEMG signals were sampled at

frequency of 340Hz with 8-bit precision, and the ex-

perimental setup was interfaced with LabVIEW using

a baudrate of 57,600 bps. The EMG signals were dis-

played on a real time monitor for visual inspection to

ensure quality acquisition.

Figure 1: Electrode placement on upper limb for EMG ac-

quisition.

BIOSIGNALS 2018 - 11th International Conference on Bio-inspired Systems and Signal Processing

Figure 2: sEMG signal with energy operator (a) sEMG (b)

the Energy estimated using TKEO from eqn.2.

3 FEATURES

In order to estimate the onset/offset points following

aspects of the sampled sEMG signal have been uti-

lized in the proposed algorithm (discussed in detail in

section 4).

3.1 Teager-Kaiser Energy Operator

TKEO is a simple algorithm with two multiplication

and a single addition to determine the temporal energy

level of a signal at any instant of time as depicted in

eqn. 2. It has been used by different researchers for

EMG activity analysis with a ﬁxed threshold (Micera

et al., 2001) (Xu and Adler, 2004). Fig. 2 presents the

Energy calculated using TKEO operator for a sEMG

signal.

τ(n) = x(n)

− x(n − 1)x(n + 1) (2)

Where, τ(n) is presents TKEO energy estimator and

x(n) is the input EMG signal.

3.2 Correlation Coefﬁcient

The correlation coefﬁcient is utilized in order to es-

timate the change in current state of activity. Cor-

relation between consecutive w samples is calculated

if the correlation coefﬁcient of two consecutive win-

dows is approximately 0, it means the behavior in the

two windows is changing, referring to a change in cur-

rent state of the muscles. Thus, the transition points

of the correlation coefﬁcient denoted by ZC{r(n)}

are the candidate onset/offset points termed as ﬁdu-

cial points, as referred in eqn. 3 below.

r(n) =

cov(x(n − w : n − 1), x(n : n + w))

std(x(n − w : n − 1))std(x(n : n + w))

ZC{r(n)} =



0; |r(n)| < ε

1; Otherwise

(3)

Where r(n), is the correlation coefﬁcient, x(n) is the

current n

sample of input signal, w is the size of

Figure 3: Black solid line presents sEMG signal while grey

dotted line show activity region where the small width ac-

tive regions are false alarms.

samples selected for correlation evaluation, which in

our case is 30, cov is the covariance function and std

represents the standard deviation; ε is zero crossing

threshold and ZC{r(n)} refers to potential candidate

points for onset/offset reference.

3.3 Minimum Activity Length

In order to complete a simple activity such as wrist

ﬂexion or extension, it is expected that some deﬁnite

minimum time duration would be required. In our

case, we have assumed it to be around one half of a

second, i.e. if Fs is the total number of samples per

second then an activity interval must be at least Fs/2

samples long. Therefore, a second threshold based on

activity length is incorporated in the proposed algo-

rithm. If the length of the activity region identiﬁed

is found to be less than Fs/2, than the candidate ac-

tivity region is considered as no activity region and is

simply rejected. An example of such false alarms and

true activity regions is presented in Fig. 3.

4 ALGORITHM

During all the experiments performed, the hand was

initially kept at rest position. For this reason, the al-

gorithm was initialized with “no activity in progress”.

In order to estimate the onset/offset points, the TKEO

parameter for sEMG signal is calculated at every in-

stant, referred to as, τ(n), followed by correlation co-

efﬁcient estimation of w consecutive samples. If the

correlation coefﬁcient attains a value less than ε, it

denotes a change in current state of wrist position, re-

ferred as ZC{r(n)}. Once an initial onset point is iden-

tiﬁed, a record of energy between the current onset

points till the next ﬁducial point is tracked and time

interval for maximum energy is estimated. If the en-

ergy in current time interval is maximum so far, pre-

vious interval is cross examined. If the energy level in

the previous interval was at least half of the energy in

A Robust and Adaptive Algorithm for Real-time Muscle Activity Interval Detection using EMG Signals

the current interval, the starting point of previous in-

terval is considered as an onset else the ﬁducial point

of current band is the marked onset. The process con-

tinues till an offset is encountered. Once an onset has

been found, search for the offset is activated among

the subsequent ﬁducial points. For this purpose, en-

ergy of the two consecutive intervals, deﬁned by the

subsequent ﬁducial points, is observed. If the energy

level in any two successive intervals falls below a cer-

tain threshold, which in our case has been taken equal

to 5% of the energy of the onset interval (based on ra-

tio of maximum and minimum power observed), we

mark the ending ﬁducial point of the second interval

as the offset. As the level of threshold is not against

a hard constant value rather 5% of the energy interval

therefore this initial search for onset/offset is referred

as the ﬁrst adaptive threshold.

The candidate MAI from the ﬁrst adaptive thresh-

old is then evaluated for minimum activity length. It

has been observed as well as adopted in different liter-

ature that the minimum duration for any hand activity

requires at least half a second (Merlo et al., 2003).

Therefore, in case the number of samples of the can-

didate MAI exceeds Fs/2, the interval is marked as

a valid MAI otherwise rejected as a false alarm. A

detailed step-by-step description of the algorithm is

provided in Table1 in the form of a pseudo code.

5 RESULTS

This paper presents the implementation of double

threshold adaptive algorithm for real time sEMG ac-

tivity interval detection. The algorithm was tested on

real time sEMG signals with three different types of

external forces, as discussed in Section 2. All results

were tested for both wrist ﬂexion as well as wrist ex-

tension and while computing the qualitative measures

an average for each scenario using 30 signals for both

ﬂexion and extension by each patient were quantiﬁed.

Performance rating for all scenarios is conducted by

comparing the mean error µ, and its respective stan-

dard deviation σ. The mean error was evaluated by

comparing the estimated ﬁducial points with visually

inspected ﬁducial points.

5.1 sEMG without Any External Force

Two channel sEMG signals were acquired from upper

limb ulnaris and radialus carpi muscles using Olimex

EKG/EMG shield with a sampling frequency of 340

Hz. The elbow was placed on a wooden table with

wrist ﬂexed from 0

◦

rest position and in a similar

manner extended from the same rest area. A sample

Figure 4: sEMG for wrist ﬂexion and extension without any

external force applied.

of the ﬂexion and extension signals acquired with-

out any external force acting on the hand / wrist, as

discussed in the above fashion are presented in Fig.

4. Due to negligible external force the MAI from

sEMG signal is accurately characterized by the al-

gorithm with an error rate of 0.01 ± 0.01msec and a

maximum accuracy of 99.9%.

5.2 sEMG with Applied Opposing Force

(i.e. Frictional Force)

Flexion and extension were performed with a force

in the opposite direction of movement of hand act-

ing as a frictional force. The attributes of the ac-

quired sEMG signal with and without external force

are same in terms of data variation and other statisti-

cal means. However, an evident increase in the time

duration of MAI as well as additional harmonics are

observed as shown in Fig. 5(a& c). These oscillations

appears due to the resistance of the opposing force,

Figure 5: sEMG signals with opposing force applied; (a)

wrist ﬂexion and (b) wrist extension, with respective energy

signals.

BIOSIGNALS 2018 - 11th International Conference on Bio-inspired Systems and Signal Processing

Table 1: Proposed Algorithm.

Input: A matrix X ∈ ℜ

[N×2]

with N representing number of samples and 2 are the number of channels.

Output :s =



0;I f activityinProgress

1;Otherwise

(4)

1. Initial Values: k=1;l=1;w=30; mE=0; onset=0;

2. τ(n) = x(n)

− x(n − 1)x(n + 1); n=1,2,3,...,N-w //TKEO

3. r(n) = corr(|τ(n − w : n)|,|τ(n : n + w)|) //correlation of w consecutive elements, N is the

total current length of input samples acquired

4. if(|r (n)|=<0.01) //For any channel using or operator

on(k)=n;

if (k == 1) //For initializing onset

onset=on(k)

end

k++;

if (k > 2) //Estimating the correct onset point using

maximum Energy interval

initial.Energy =

on(k)

∑

u=on(k−1)

|τ(u)| current.Energy =

on(k+1)

∑

u=0n(k)

|τ(u)| (5)

end

5.if(max.Energy<initial.Energy)

max.Energy=initial.Energy //Maximum Energy

if(2 ×

on(k−1)

∑

u=on(k−2)

|τ(u)| < initial.Energy) (6)

onset=on(k-1);

ty=on(k+1);

else

onset=on(k);

ty=on(k-2);

end

s=1;

end

6. if(onset = 0) //If activity is initialized

if (current.Energy < initial.Energy && current.Energy > max.Energy*0.05)

offset= on(ii+1);

s=0;

end

else

off= ty;

end

7. if (s==1 && n-onset > Fs/2)

Activity In Progress

elseif (offset-onset >Fs/2)

x(onset : offset) is the detected Activity Region

mE=0;k=1;onset=0;

end //Repeat Step 1 to 5 till the process continues

A Robust and Adaptive Algorithm for Real-time Muscle Activity Interval Detection using EMG Signals

resulting in an increase in MAI as compared to nor-

mal time for the wrist to complete its 90

◦

ﬂexion or

extension from rest position. In order to choose the

correct onset, it is required to analyze the energy level

in the signal continuously. Due to the oscillatory be-

havior the TKEO based correlation results in multiple

closely detected onset candidates as highlighted with

black dots in Fig. 5. However, the energy threshold

between every two consecutive ﬁducial points tracks

the point from where a consistent increase in energy

level is observed, till the maximum energy value is

achieved, resulting in the identiﬁcation of the accu-

rate onset point. It further investigates the existence of

offset point and continuously compares the detected

MAI with the set sample threshold.

5.3 sEMG with Variable Weights

The same experiment of wrist ﬂexion and extension

has been tested with multiple weights in order to in-

vestigate the robustness of the proposed algorithm.

Multiple alarms are activated in the form of ZC{r(n)}

which is an afﬁrmation of presence of activity. How-

ever, the adaptive threshold is successful in identify-

ing the region of interest. With minimum weights

i.e. 1 kg and 2 kg the mean error observed is al-

most the same as no weight added (0.01 ± 0.01, &

0.01 ± 0.02msec, respectively). With the increase in

weight the energy level starts increasing as shown in

Fig. 6(b & d). At 5 kg as the weight becomes heavy,

there is a continuous pressure on the hand muscles re-

sulting in initial oscillations even when the wrist is at

Figure 6: Results of proposed algorithm for sEMG with its

corresponding TKEO graph for different weight lifted dur-

ing wrist extension (a, b) and ﬂexion (c, d, e, & f).

rest position, as can be clearly observed in Fig. 6(f)

between 0.5 to 1 sec duration. As the proposed algo-

rithm is based on maximum energy as well as interval

length, it selects the second region as muscle activity

region and rejects the initial energy band as a false

alarm Fig. 6(e & f).

6 DISCUSSION

Muscle contraction or relaxation is responsible for the

accurate movement of any limb in human body. In or-

der to estimate muscle activity onset/offset point de-

tection, this paper proposes the use of adaptive thresh-

old based on correlation coefﬁcient of TKEO energy

estimator. The proposed algorithm is tested for dif-

ferent scenarios such as free wrist i.e. no external

force applied, with an opposing force in the direction

of motion, and wrist movement with different weights

attached. The results for all above discussed scenarios

is presented in Fig. 7. The mean error in case of no

external force or 1 kg weight is near to zero, however,

when the weight is increased to 2 kg and 5 kg or in

case of opposing force, the error slightly increase with

a maximum error variation of 270 m sec and provides

lowest results with an acurracy of 98.3%. When the

hand muscle grasps a weight as heavy as 5 kg pressure

is exerted as a result of which additional harmonics

are observed in sEMG recorded even at rest position

of the wrist. Therefore, the results of MAI is not as

high as that in low weight sEMG signals. A compar-

ison with the onset MSE proposed by (Solnik et al.,

2010) is also presented in the Fig. 7. The onset de-

tection in no external force, lite weights and opposing

Figure 7: Performace comparison (on the basis of Mean

Square Error (MSE) and Standard Deviation values) of the

proposed algorithm with the algrithm presented by (Solnik

et al., 2010) under various loading scenarios. Bold sym-

bol shapes (i.e. Square, Diamond and Triangle) indicate

the Mean Error value while the vertical dotted lines repre-

sent the spread of error. Note that the proposed algorithm

detects both onset as well as offset points, whereas Solnik

algorithm detects onsets ONLY.

BIOSIGNALS 2018 - 11th International Conference on Bio-inspired Systems and Signal Processing

Table 2: Mean error and stand deviation comparison of algorithm results for sEMG Onset Detection with constrained and

unconstrained environment at different SNRs.

SNRs No Force Friction 1 kg 2 kg 5 kg

(dB) µ σ µ σ µ σ µ σ µ σ

x(t) 0.01 0.01 0.26 0.57 0.01 0.01 0.01 0.05 0.03 0.08

1.25 0.09 0.32 4.72 3.04 0.90 1.76 2.01 3.04 3.21 4.26

3 0.04 0.05 2.89 3.47 0.04 0.07 1.36 1.53 1.23 3.70

6 0.05 0.07 2.09 2.55 0.05 0.06 0.25 0.82 1.85 2.43

9 0.09 0.13 1.18 1.98 0.05 0.05 0.25 0.86 1.47 1.25

Table 3: Mean error and stand deviation comparison of algorithms results for sEMG Offset Detection with constrained and

unconstrained environment at different SNRs.

SNRs No Force Friction 1 kg 2 kg 5 kg

(dB) µ σ µ σ µ σ µ σ µ σ

x(t) 0.01 0.01 0.02 0.08 0.004 0.01 0.04 0.01 0.08 0.20

1.25 0.68 2.22 4.77 5.00 1.47 1.13 3.31 2.48 4.75 3.76

3 0.15 0.15 3.93 4.80 0.28 0.22 0.18 2.29 2.85 2.03

6 0.18 0.21 2.50 3.84 0.25 0.24 0.18 0.35 1.30 2.08

9 0.12 0.15 1.50 2.38 0.21 0.26 0.76 0.27 1.30 1.80

frictional force is comparable, where as in case of 5 kg

weight the proposed algorithm out performs the other

by an onset detection accuracy of 98.3% in compari-

son with 88.1% achieved by (Solnik et al., 2010).

In order to investigate the performance of algo-

rithm for external interference, AWGN with different

SNRs was added to the already acquired EMG sig-

nals. The algorithm was run ofﬂine on noise added

EMG signals to evaluate the performance for on-

set/offset detection accuracy. The results are summa-

rized in Table 2 & 3, respectively. For 3 and 9 dBs

SNR sEMG signals, with no external force, 1 kg, and

2 kg weights the mean error is below 280 msec. But

when the signal strength decreases as in case of 1.25

and 3 dB the algorithm has a slight increase in mean

error and standard deviation, especially in case of fric-

tional force and 5 kg weights applied. This is due to

the fact that when an external force is applied the mus-

cle tries to contract but the external force has a domi-

nant pressure on the muscle, resulting in muscle con-

traction noise. Further increase in SNR deteriorates

the strength of the signal thus an increase in error is

observed while estimating the MAI. sEMG acquired

in stressful scenarios have a higher frequency as com-

pare to sEMG acquired from wrist with no external

interference or lite weights such as 1 kg or 2 kg. The

results can be easily improved for such signals if the

window length w is increased. However, increase in

window length would make the algorithm less suit-

able for real time applications, therefore, an adaptive

window length selection needs to be catered in future.

All sEMG signals evaluated on sEMG focusing

easily available low cost equipment targeting lower

frequency band of 0.14 Hz-40 Hz. As the lower fre-

quencies provides information about the presence of

activity and higher frequencies tends to deﬁne the

particular movement, therefore, the use of lower fre-

quency range is sufﬁcient for MAI detection. How-

ever, it may effect if the same low frequency sEMG

signals are used for movement identiﬁcation.

7 CONCLUSION

The extraction of muscle activity interval can play

a vital role for real time Myo-electric controlled de-

vices and other clinical applications. The purpose

of the proposed work is to present an algorithm for

real time detection of muscle activity using affordable

hardware. Although the hardware selected focuses on

low frequency sEMG signal only, the detected MAI is

considerably accurate and therefore, low cost sEMG

acquisition devices such as Olimex EKG/EMG shield

can be utilized for control of automated prosthesis in

order to provide cheaper solutions.

While majority of the methods proposed in lit-

erature are limited to ofﬂine analysis and onset de-

tection; our proposed methodology is fast and easily

implementable for real-time systems as well as han-

dles both activity activation and deactivation regions.

The proposed algorithm not only detects the complete

muscle activity interval but also provides consider-

ably higher accuracy than other previously proposed

algorithms. The methodology presented adopts a dou-

A Robust and Adaptive Algorithm for Real-time Muscle Activity Interval Detection using EMG Signals

ble threshold adaptive algorithm for MAI detection.

The initial threshold is based on correlation coefﬁ-

cient and percentage of maximum energy providing

an adaptive behavior. The use of correlation coefﬁ-

cient provides better performance for different SNRs

while the percentage of maximum energy computed

over an interval selects the most optimum candidate

interval. The experimental results are achieved us-

ing real surface EMG signals in different scenarios as

well as cross checked with different levels of AWGN.

We have shown that the proposed algorithm is ro-

bust to estimate the closest correct onset and offset

point even with external interference. The proposed

algorithm performs best with lite or no external force,

achieving an accuracy of 99.9%, and improves perfor-

mance by 10.2% in worst case scenario in comparison

with previously proposed work.

REFERENCES

Drapala, J., Brzostowski, K., Szpala, A., and Kucharska,

A. (2010). Two stage emg onset detection method.

Archives of Control Science, 22:427–440.

Farina, D., Jiang, N., Rehbaum, H., Holobar, A., Grain-

mann, B., Dietl, H., and Aszmann, O. (2014). The

extraction of neural information from the surface emg

for the control of upper limb prosthesis: Emerging av-

enues and challenges. IEEE transaction on Neural

Systems and Rehabilitation Engineering, 22:797–809.

Hermens, H., Merlett, R., and Freriks, B. (2000). Surface

ElectroMyopGraphy for Non-Invasive Assessment of

Muscles(SENIAM).

Liu, J., Ying, D., and Rymer, W. (2015a). Emg burst pres-

ence probability: A joint time-frequency representa-

tion of muscle activity and its application to onset de-

tection. Journal of Biomechanics, 48.

Liu, J., Ying, D., Rymer, W. Z., and Zhou, P. (2015b). Ro-

bust muscle activity onset detection using an unsu-

pervised electromyogram learning framework. Plos

ONE, 10:1193–1197.

Merlo, A., Farina, D., and R., M. (2003). A fast and reliable

technique for muscle activity detection from surface

emg signals. IEEE transaction on Biomedical Engi-

neering, 50:316–323.

Micera, S., Vannozzi, G. andSabatini, A. M., and F., D.

(2001). Improving detection of muscle activation in-

tervals. IEEE Engineering in Medicine and Biology

Magazine, 20.

Naseem, A., Jabloun, M., Ravier, P., and Buttelli, O. (2016).

Muscular activation intervals detection using gaus-

sian mixture model gmm applied to semg signals.

In IEEE 18th International Conference on e-Health

Networking Applications and Services (Healthcom).

IEEE Xplore.

Solnik, S., Rider, P., Steinweg, K., DeVita, P., and Horto-

bagy, T. (2010). Taeger kaiser energy operator signal

conditioning improves emg onset detection. European

Journal of Applied Physiology, 110:489–498.

Xu, L. and Adler, A. (2004). An improved method for

muscle activation detection during gait. In Canadian

conference on Electrical and Computer Engineering.

IEEE Xplore.

Xu, Q., Quan, Y., Yang, L., and He, J. (2013). An adaptive

algorithm for the determination of the onset and offset

of muscle contraction by emg signal processing. IEEE

transaction on Neural Systems and Rehabilitation En-

gineering, 21:65–73.

BIOSIGNALS 2018 - 11th International Conference on Bio-inspired Systems and Signal Processing