riewarticle semg based human motion intention...

Review ArticlesEMG Based Human Motion Intention Recognition

Li Zhang Geng Liu Bing Han ZheWang and Tong Zhang

Shaanxi Engineering Laboratory for Transmissions and Controls Northwestern Polytechnical University Xirsquoan China

Correspondence should be addressed to Geng Liu npuliugnwpueducn

Received 7 May 2019 Accepted 17 July 2019 Published 5 August 2019

Academic Editor Shahram Payandeh

Copyright copy 2019 Li Zhang et alThis is an open access article distributed under the Creative Commons Attribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Human motion intention recognition is a key to achieve perfect human-machine coordination and wearing comfort of wearablerobots Surface electromyography (sEMG) as a bioelectrical signal generates prior to the corresponding motion and reflectsthe human motion intention directly Thus a better human-machine interaction can be achieved by using sEMG based motionintention recognition In this paper we review and discuss the state of the art of the sEMG based motion intention recognitionthat is mainly used in detail According to the method adopted motion intention recognition is divided into two groups sEMG-drivenmusculoskeletal (MS)model basedmotion intention recognition andmachine learning (ML)model basedmotion intentionrecognition The specific models and recognition effects of each study are analyzed and systematically compared Finally adiscussion of the existing problems in the current studies major advances and future challenges is presented

1 Introduction

Along with worldwide population aging and increasing num-ber of the disabled and amputee the wearable robots recentlyget extensive research For the wearable robots human-machine interface is a research hotspot which acquireshuman motion intention by collecting and analyzing relatedinformation and assists external devices to develop effectivecontrol strategies [1] Accurate and real-time recognitionof human motion intention is the key to achieve perfecthuman-machine coordination and wearing comfort [2 3]As a bioelectrical signal surface electromyography (sEMG) isactivated when a neuron carrying human intention informa-tion is transmitted to related muscles and reflects the humanmotion intention directly [4 5] Hence the motion intentioncan be fully estimated without any information delay andlose [6 7] Because of containing rich information matureacquisition technology and noninvasiveness the humanmotion intention recognition based on sEMG is about to gomainstream [8 9]

The methods of sEMG based motion intention recog-nition can be divided into two groups sEMG-driven mus-culoskeletal (MS) model based and machine learning (ML)based For the former a function between sEMG andjoint moment angular velocity or angular acceleration canbe established by biomechanics model of muscles An

explanation of motion production process is the advantageof this method [3 10] For the latter the sEMG featureor processed sEMG is provided as input to the ML Thediscrete-motion classification or continuous-motion estima-tion is realized by establishing the mapping between inputand human motion intention The ML commonly usedfor motion intention recognition includes support vectormachine (SVM) linear discriminant analysis (LDA) back-propagation neural network (BPNN) and deep learning (DL)[11 12] Compared to the former the ML model possessesthe characteristics of lower computational complexity shortoperation time and real-time performance With the devel-opment of deep learning (DL) research in recent years DLis increasingly used for human motion intention recognitionCompared to the others DL greatly improves the nonlinearityof model the ability of solving complex problem and theaccuracy of recognition [13] The DL model commonlyused for motion intention recognition includes deep beliefnetwork (DBN) convolutional neural network (CNN) andstacked auto-encoder (SAE) [14 15]

There are several related review papers appearing inrecent years Nazmi et al [16] reviewed the classifica-tion methods of motion patterns based on sEMG Abrief comparison of the different methods for preprocess-ing feature extraction and classifying sEMG signals wasprovided Chowdhury [17] analyzed the signal processing

HindawiJournal of RoboticsVolume 2019 Article ID 3679174 12 pageshttpsdoiorg10115520193679174

2 Journal of Robotics

sEMGAcquisition Pre-processing Feature Extraction

KinematicsKinetics Date Acquisition

Discrete-Motion Classification

Continuous-Motion Regression

sEMG-Driven Musculoskeletal Model

Controller HMI

Motion Intention Recognition

Figure 1 The process of sEMG based human-machine interaction

Muscle Activation

Model

Muscle Contraction

Model

Musculoskeletal Geometry

(Moment Arms)

ProcessedEMGs Forward

Dynamicsai(t) Fmt

i (t)

agon(t)

antag(t) (t) (t) (t) (t)

int int

lmti (at)i (at)i(at)

Figure 2 The sEMG-driven musculoskeletal model [10]

of sEMG and evaluated the pros and cons of differentclassification models Singh et al [2] discussed the cur-rent development and challenges of the sEMG based con-trol schemes which are employed in designing exoskele-ton in stroke rehabilitation However the sEMG basedcontinuous-motion intention regression and sEMG-drivenmusculoskeletal model based motion intention recognitionare rarely reviewed And the two methods are more valuableto realize the smooth control of wearable robot movements[3 18 19] In order to further understand the knowledgeof the sEMG based motion intention recognition thispaper presents the review of all commonly used methodsof human motion intention recognition for last decadebriefly

The rest of this paper is organized as follows In Sec-tion 2 the motion intention recognition methods basedon the sEMG-driven musculoskeletal model is reviewed InSection 3 we discuss the various ML methods for discrete-motion classification and continuous-motion regressionIn Section 4 a succinct conclusion of this paper ispresented

2 sEMG-Driven MS Model Based MotionIntention Recognition

The sEMG is a nonstationary and microelectric signal whichamplitude is concentrated in 001-10mV and frequency isconcentrated in 20-500Hz especially in 50-150Hz [20]Because of about 30-150ms prior to the correspondingmotion generated sEMG is an ideal choice for motionintention estimation [20 21] Figure 1 shows the human-machine interaction process based on sEMG In the wholeprocess human motion intention recognition is the mostcritical part It can be achieved through two ways sEMG-driven MS model and ML model

The sEMG-driven musculoskeletal model can be dividedinto three submodels ie activation model contractionmodel and musculoskeletal geometry model as shown inFigure 2 [10] To serve as input to the model the raw sEMGsignal should be preprocessed by high-pass filtering full-wave rectification low-pass filtering and normalization [3]For the activation model the relationship between muscularactivation (119886119894(t)) and the processed sEMG signal (119906119894(119905)) of

Journal of Robotics 3

Fmt(t)SE

lt(t)

lmt(t)

lm(t)

(t)

PE

CE

Fmt(t)

Figure 3 The Hill-type muscle model [23]

muscle 119894 at time 119905 can be expressed as the following equation[3 22]

119886119894 (119905) = 119890119860119894119906119894 (119905) minus 1119890119860119894 minus 1 (1)

where minus3 lt 119860 119894 lt 0 is a nonlinear shape factor of musclenumber i For the contraction model Hill-type musclemodelwas always used as shown in Figure 3 [23] The forceproduced by the muscle-tendon unit (119865119898119905119894 (t)) can be given by

119865119898119905119894 (119905) = 119865119905119894 (119905)= 119865119898119886119909119894 [119891119894 (119897) 119891119894 (V) 119886119894 (119905) + 119891119901119894 (119897)] cos (120593119894 (119905))

(2)

where119865119905119894 (119905) and 119865119898119886119909119894 denote the tendon force andmaximumisometric muscle force 119891119894(119897) 119891119894(V) and 119891119901119894(119897) are the genericforce-length generic force-velocity and parallel passive elas-tic force-length curves of muscle number i respectively120593119894(119905) represents the pennation angle which is defined as theangle between the muscle fiber and the tendon [3 22] Forthe musculoskeletal geometry model the moment arms ofmuscle-tendon unit (119903119894(119905)) can be defined as

119903119894 (119905) = 120597119897119898119905119894 (119905)120597120579 (3)

where 120579 is the joint angle and 119897119898119905119894 (119905) is the muscle-tendonlength 119897119898119905119894 (119905) can be calculated by

119897119898119905119894 (119905) = 119897119905119894 (119905) + 119897119898119894 (119905) cos (120593119894 (119905)) (4)

where 119897119905119894 (119905) and 119897119898119894 (119905) represent the lengths of tendon andmuscle fiber respectively [3 10] Thus the joint moment canbe given by the following equation [3]

120591 (119905) =1003816100381610038161003816100381610038161003816100381610038161003816119899

sum119894=1

119865119894 (119905) 119903119894 (119905)1003816100381610038161003816100381610038161003816100381610038161003816119886119892119900119899119894119904119905minus10038161003816100381610038161003816100381610038161003816100381610038161003816119898

sum119895=1

119865119895 (119905) 119903119895 (119905)10038161003816100381610038161003816100381610038161003816100381610038161003816119886119899119905119886119892119900119899119894119904119905

(5)

where 119899 and 119898 denote the number of agonist and antagonistmuscles acting on the joint respectively The joint angular

acceleration ( 120579(119905)) can be calculated by the joint forwarddynamics [23 24]

120579 (119905) = (120591 (119905) minus 120591119890119909 (119905))119868 (6)

where 119868 represents the joint inertia and 120591119890119909(119905) includes theexternal torque and the limbs gravity torque Consequentlythe joint angular velocity ( 120579(119905)) and angle (120579(119905)) can becalculated by

120579 (119905) = int 120579 (119905) 119889120579 (7)

120579 (119905) = int 120579 (119905) 119889120579 (8)

There are several unknown parameters in the sEMG-drivenmusculoskeletal model so the parameters identificationthrough the preliminary experiment is necessary Han et al[23] and Ding et al [24] developed a state-space sEMGmodelto estimate the continuous motion of elbow joint directly anda closed-loop prediction-correction approach was employedThe results of preliminary experiment showed that the rootmean squared error (RMSE) of angle and angular velocitybetween estimated and actual values is around 010 rad and015 rads and the correlation coefficient (CC) is around 099and 091 respectively Lloyd et al [22] utilized a modifiedHill-type muscle model to estimate muscle forces and kneemoments An average CC of 091 and mean residual error(MRE) of 12 Nm was observed Karavas [3] employed thecommon sEMG-driven musculoskeletal model to estimatethe knee torque trajectory and stiffness trend The resultsshowed that the normalized RMSE was about 012 betweenthe estimated and actual values In the study of Sartori etal [10] a multi-DOF sEMG-driven model was developedto estimate the muscle force and joint moment of lowerextremity The results showed that the average normalizedmean absolute error (MAE) of joint moment of three lowerextremities was around 015

3 Machine Learning Based MotionIntention Recognition

Machine learning (ML) based motion intention recognitioncan be divided into two groups discrete-motion classificationand continuous-motion regression For the former a map-ping between sEMG and discrete-motion of upperlowerlimbs needs to be established The common classifiedlower limbs motions include walking running sit-to-stand stand-to-stand stair ascent and stair descent Andthe common classified upper limbs motions includeshoulder flexionextensionadductionabduction elbow flex-ionextension wrist flexionextensionradial deviationulnardeviation thumb flexionextensionadductionabductionindex flexionextensionmiddle finger flexionextension ringfinger flexionextension litter finger flexionextension handgrasp and pinch grasp [25 26] For the latter a mappingbetween sEMG and continuous-motion of upperlowerlimbs needs to be constructed The common regressed limbs


motions include angle angular velocity angular accelerationforce and moment of hip knee ankle shoulder elbow andwrist joint Compared to the former a mature method thelatter is more valuable for the smooth control of wearingrobots and will be the focus of future research [20]

31 Machine Learning Based Discrete-Motion ClassificationTable 1 reviewed the most recent studies about discrete-motion classification As shown in Figure 1 feature extrac-tion and classification model construction are two mostimportant and key steps in discrete-motion classificationThe commonly used feature can be mainly divided intotime domain feature frequency domain feature and time-frequency domain feature For the time domain featuremeanabsolute value (MAV) [27ndash32] root mean square (RMS)[29 31] variance (VAR) [29 31] standard deviation (SD)[29] zero count (ZC) [27 29 32] waveform length (WL)[27 29 32] slope sign change (SSC) [29 32] integrated EMG(IEMG) [33] and difference of mean absolute value (DMAV)[27] are commonly utilized Although the calculation of timedomain feature is simple it is not enough to describe theinformation of signals For the frequency domain featurepeak frequency (PF) median frequency (MF) and meanpower frequency (MPF) are commonly utilized It is onlyused to analyze the fatigue of muscle [34] For the time-frequency domain feature Fourier Transform Features [27]and Wavelet Transform Features [35] are commonly usedAlthough the comprehensive information of signal can beobtained the extraction process of sEMG is complex and timeconsuming When multichannel sEMG signals are used forfeature extraction feature redundancy often existsThereforedimensionality reduction algorithmwhich is usually adoptedprincipal component analysis needs formultichannel featureextraction [20]

SVM based classification model has the ability to resolvethe nonlinear binary classification problem by constructingan optimal classification hyperplane with the largest marginto separate the two classes of samples [25] For resolvingthe multiclassification problem one-versus-one SVM one-versus-rest SVM multistep SVM etc are common utilizedBabita et al [36] employed linear SVM and wavelet packettransform to classify binary elbow flexion and extension A911 classification accuracy was observed for this methodYang et al [37] classified eight hand motions including palmextension palm turn downwards palm turn upwards palmenstrophe palm ectropion fist turn downwards fist turnupwards and clenching by using genetic algorithmoptimizedSVM Power spectral density was used for feature extractionThe results showed that the training and testing recognitionaccuracy could reach 9937 and 9033 respectively Suiet al [38] utilized an improved SVM to classify six upperlimb motions namely elbow flexion elbow extension wristinternal rotation wrist external rotation fist clenching andfist unfolding The energy and variance of the wavelet packetcoefficients were selected as feature vectors The resultsshowed that the average recognition accuracy could reach9066 Cai et al [25] adopted one-versus-one SVM toclassify five upper limb motions namely shoulder flexion

shoulder abduction internal rotation external rotation andelbow flexion The results showed that the classification accu-racy could reach 9418 Pan et al [39] classified six fingermotions namely thumb bending index finger bendingmiddle finger bending ring finger bending and litter fingerbending by using one-versus-one SVM Relative energycoefficient of wavelet packet was selected as the input featureof classifier The results showed that the recognition accuracyreached 9778 Chen et al [40] utilized two-step SVM toclassify seven upper limb motions namely shoulder flexionshoulder extension shoulder adduction shoulder abductionelbow flexion and elbow extension By extracting RMS asinput feature a shorter classification time and more accurateresults could be obtained Naik et al [41] developed a twinSVM to classify seven motions including wrist flexion ringand middle finger flexion wrist flexion toward litter fingerwrist flexion toward thumb finger and wrist flexion fingerand wrist flexion toward litter finger and finger and wristflexion toward thumb An 8483 classification accuracy wasobserved for this method

LDA k-nearest neighbour (K-NN) naive Bayes (NB)quadratic discriminant analysis (QDA) random tree (RT)randomForest (RF) etc are also commonutilized as classifierlike SVM Liu et al [42] employed mixed LDA to classifythirteen hand motions including fist open hand radialdeviation ulnar deviation wrist flexion wrist extensionpronation supination fine pinch key grip ball grasp andcylinder grasp An average classification accuracy could reach8874 for this method Dhindsa et al [43] compared fourclassifiers namely LDA NB K-NN and SVM in classifyingfive classes of knee angle Fifteen features including timedomain features frequency domain features and autore-gressive coefficients were used as input vectors The resultsshowed that the classification accuracy with LDA NB K-NN and SVM classifier could reach 716 751 879and 922 respectively Pancholi et al [33] classified sevenhandmotions including hand open hand close wrist flexionwrist extension soft gripping medium gripping and hardgripping by using LDA K-NN QDA SVM RT and RF Ninetime domain features and seven frequency domain featureswere extracted as input vectors The results showed thatthe RF had the maximum classification accuracy (9954)and the LDA had the minimum classification accuracy(7538) Bian et al [11] utilized LDA RF NB and SVM toclassify eight hand motions including twist a water bottlecap turn a key press an automatic pencil press a nailclipper preform ldquoshootrdquo gesture preform ldquorockrdquo gesturepreform ldquookrdquo gesture and preform ldquoyeahrdquo gesture IEMG SDRMS MPF and MF were selected as the input features A9167 classification accuracy for LDA 8750 classificationaccuracy for RF 8683 classification accuracy for NB and9225 classification accuracy for SVM were obtained in thisstudy Alomari et al [12] compared LDA QDA and K-NNin classifying eight hand motions namely wrist flexion wristextension ulnar deviation radial deviation grip open handpinch and catch cylindrical subject Sample entropy RMSmyopulse percentage rate (MYOP) and difference absolutestandard deviation value (DASDV) were selected as featuresThe results showed that the classification accuracy with LDA


Table 1 Results from most recent studies for discrete-motion classification

Study Classification motions Features selected Classificationmethods Accuracy

Babita et al [36] Elbow flexion and extension Wavelet packettransform Linear SVM 911

Yang et al [37]Fist turn downwardsupwards Palm

extensionenstropheectropionturn upwardsturndownwards and clenching

Power spectraldensity

Genetic algorithmoptimized SVM 9033

Sui et al [38] Elbow flexionextension wrist internalexternalrotation and fist clenchingunfolding

The energy andvariance of thewavelet packetcoefficients

Improved SVM 9066

Cai et al [25] Elbow flexion and shoulder flexionabductioninternalrotationexternal rotation

RMS VAR WLMAV etc

One-versus-oneSVM 9418

Pan et al [39] Thumbindexmiddleringlitter finger bendingRelative energycoefficient ofwavelet packet


Chen et al [40] Elbow flexionextension and shoulderflexionextensionadductionabduction RMS Two-step SVM mdash

Naik et al [41]Wrist flexion ring-middle finger flexion wrist flexiontoward litter fingerthumb finger and wrist flexionfinger and wrist flexion toward litter fingerthumb

RMS Twin SVM 8483

Liu et al [42]Fist open hand radialulnar deviation wrist

flexionextension pronation supination fine pinchkey grip ballcylinder grasp

6-order ARcoefficients Mixed LDA 8874

Dhindsa et al [43] Five classes of knee angle

IEMG SSI RMSZC WL WA

MNF MF PF MPSM1 4 ARcoefficients

LDA NB K-NNand SVM

716 (LDA)751 (NB) 879(K-NN) and 922

(SVM)

Pancholi et al [33] Softmediumhard gripping wrist flexionextensionand hand openclose

IEMG MAVMMAV1 MMAV2WAMP RMS WLZC SSI MNF

MDF PKF MFDFMD FMN and

MFMD

LDA K-NN QDASVM RT and RF 7538-9954

Bian et al [11]Preform ldquoshootrdquoldquorockrdquoldquookrdquoldquoyeahrdquo gesture twist awater bottle cap turn a key press an automatic pencil

and press a nail clipper

IEMG SD RMSMPF and MF

LDA RF NB andSVM

9167 (LDA)8750 (RF)

8683 (NB) and9225 (SVM)

Alomari et al [12] Wrist flexionextension ulnarradial deviation gripopen hand pinch and catch cylindrical subject

Sample entropyRMS MYOP and

DASDV

LDA QDA andK-NN

9856 (LDA)9342 (QDA) and9425 (K-NN)

Oleinikov et al [27] Different hand motions MAV DMAV ZCWL PF MPF etc Three layers ANN 91

Oweis et al [44] grasping extension flexion ulna deviation and radialdeviation

Seventeen time andtime-series domain

featuresThree layers ANN 967

Mane et al [35] Open palm closed palm and wrist extension Discrete wavelettransform Three layers ANN 9325

Gandolla et al [28] Pinching grasp an object and grasping mdash Three layers ANN 76

Ahsan et al [29] Different hand motionsMAV RMS VARSD ZC SSC and

WLThree layers ANN 884

Shen et al [21] The phases of sit-to-stand motion mdashThree

back-propagationneural networks

9348


Table 1 Continued


Park et al [14]Tip pinch grasp prismatic four fingers grasp powergrasp parallel extension grasp lateral grasp and

opening a bottle with a tripod graspmdash Convolutional

neural network 90

Asai et al [15] Thumb openclose fingers except thumb openclose mdash Convolutionalneural network 83

Bu et al [45] Flexion extension pronation supination grasping andopening mdash Five layers

recurrent ANN 884

Orjuela et al [46] Five wrist positions Discrete wavelettransform

Auto-encoderANN 7341

EMG Processing

Feature Extraction

Input Layer Hidden Layer Output Layer

Discrete Motion Types

MAV RMS ZC VARWL PF

Artificial Neural Network (ANN)

MF MPFmiddot middot middot

Figure 4 The process of ANN based discrete-motion classification

QDA and K-NN classifier could reach 9856 9342 and9425 respectively

As shown in Figure 4 ANN based classification modelhas the ability of learning complex nonlinear patterns byadjusting a set of free parameters known as synaptic weightsTypical shallow ANN architecture consists of an input layera hidden layer and an output layer Each layer has a weightmatrix a bias vector and an output vector Number ofneurons in the input is given by the number of featuresobtained from the above methods and in the output is givenby the number of motions needed to be classified Oleinikovet al [27] classified the hand motions by using ANN Theinput features include four time domain features (MAVDMAV ZC and WL) and two frequency domain featuresfor two samples The hyperbolic tangent sigmoid transferfunction was used for twenty-five hidden neurons and Soft-Max function for output neurons The results showed 82of offline classification accuracy for eight hand motions and91 accuracy for six hand motions Oweis et al [44] adoptedANN to classify five motions including grasping extensionflexion ulna deviation and radial deviation Seventeen timeand time-series domain features were used as input neuronsThe proposed ANN includes 30 neurons in hidden layer and5 neurons in output layerThe results showed that the averageclassification accuracy could reach 967 Mane et al [35]

utilized ANN to classify open palm closed palm and wristextension of hand motion Discrete wavelet transform wasused for feature extraction TheANNarchitecture consideredin this study was comprised of two neurons in input layerten neurons in hidden layer and three neurons in outputlayer Average 9325 recognition rate was observed by theproposed method Two cascaded ANNs were exploited inthe study of Gandolla et al [30] to detect three hand graspmotions namely pinching grasp an object and graspingThetwo ANNs have the same 1025 neurons ie pattern vectorsin the input layer 25 neurons in the hidden layer and 2neurons in the output layer In the first ANN pattern vectorwas classified in clusters And in the secondANN the clusterscontaining more than one task were then classified Thepreliminary experiment results illustrated that the proposedmethod had 76 accuracy for hand motion intention Ahsanet al [29] designed an optimal ANN structure with sevenneurons (MAV RMS VAR SD ZC SSC and WL) in inputlayer ten tan-sigmoid neurons in hidden layer and four linearneurons in output layer An average success rate of 884was obtained for classifying single channel sEMG signalsShen et al [21] utilized neural network ensemble and threeback-propagation neural networks to recognize the phasesof sit-to-stand motion The sEMG characteristics from fourmuscles of lower limbs and two floor reaction force (FRF)


Input Layer ConvolutionalLayer 1

PoolingLayer 1

ConvolutionalLayer 2

PoolingLayer 2

Fully ConnectedLayer

OutputLayer

Figure 5 A typical architecture of convolutional neural network (CNN)

characteristics were used as input to the proposed networksFor each BP network there are six neurons in input layerand five neurons in linear output layer And the tan-sigmoidhidden layer for three BP networks was 12 13 and 15 neuronsrespectively The preliminary experiment result showed thatthe recognition accuracy of the proposed method was about9348

DL is greatly employed to classify human motions inrecent years because it improves the nonlinearity of modeland the accuracy of recognition The common methods formotion classification include convolutional neural network(CNN) recurrent neural network (RNN) and stacked auto-encoder (SAE) For the CNN a typical architecture is shownin Figure 5 which consists of input layer convolutionallayer pooling layer fully connected layer and output layerPark [14] employed a deep feature learning model based onconvolutional neural network to classify six different handmotions including tip pinch grasp prismatic four fingersgrasp power grasp parallel extension grasp lateral grasp andopening a bottle with a tripod graspThe proposedmodel wascomposed of one input layer four convolutional layers fourpooling layers and two fully connected layers The resultsshowed that the classification accuracy of this method couldbe up to 90 Asai et al [15] estimated four finger motionsnamely thumb open thumb close fingers except thumbopen and fingers except thumb close based on the frequencyconversion of sEMG using convolutional neural networkThe proposed method contained two pairs of convolution-pooling layers and two fully connected layers A preliminaryexperimental result illustrated that the accuracy of motionestimation reached 83 For the RNN Bu et al [45] utilizedfive-layer recurrent log-linearized Gaussian mixture net-work (R-LLGMN) to classify six motions including flexionextension pronation supination grasping and opening Anaverage recognition accuracy of 884 was observed for thismethod For the SAE Orjuela et al [46] employed an auto-encoder based deep ANN to classify the five classes of wrist

angles Discrete wavelet transform was used to achieve theextraction of twelve featuresTheDNNarchitecture consistedof a sixty-neuron input layer a five-neuron auto-encoderlayer a four-neuron hidden layer and a five-neuron outputlayer The results showed that the classification accuracy wasaverage about 7341 for five wrist positions

In general the motion description of discrete-motionclassification is relatively simple and there is no uniformclassification standard In addition the types of motion usedfor classification are predefined The unclassifiable conditionwill happen when the undefined motion type appears [20]

32 Machine Learning Based Continuous-Motion RegressionThe motion classification can only recognize a few discretebody motion and not be used for smooth control of wearablerobotsTherefore continuous-motion regression which esti-matesmoremotion information than the former will becomethe new focus Similar to the sEMG-driven musculoskeletalmodel based motion intention recognition the mappingbetween sEMG and joint angle angular velocity angularacceleration or joint moment can also be established by MLThe common used ML based continuous-motion regressionmethods include shallow ANN and DL Therefore the tworegression methods will be mainly discussed in this sectionTable 2 reviewed the most recent studies about continuous-motion regression

321 Mapping between sEMG and Joint Kinematics For jointkinematics regression the mapping between sEMG and jointangle is commonly established The estimated angle is usedas an input signal in the control system of wearable robotsto achieve accurate angle trajectory track Compared to deepANN the shallow ANN is the most common method forsEMG based joint kinematics regression The deep ANN isstill in the development stage now and will be widely used inthe future


Table 2 Results from most recent studies for continuous-motion regression

Study Regression motions Regression methods AccuracyLuh et al [47] Elbow joint angle BPNN Satisfactory accuracy

Chen et al [48] Elbow joint angle Hierarchical projectionregression (HPR)

Regression error less than98 deg

Raj et al [49] Human forearm kinematics Radial basis functionneural network (RBFNN)

CC more than 076 forangle and 039 for angular

velocity

Wang et al [50] Elbow joint angle RBFNN RMSE less than 0043 andCC more than 0905

Kwon et al [51] Elbow and shoulder jointangles

Feed forward neuralnetwork (FFNN) mdash

Ngeo et al [52] Finger joint angles FFNN CCmore than 092 andNRMSE less than 85 deg

Xia et al [13] Upper limbs movement Recurrent convolutionalneural network (RCNN) CC more than 93

Zhang et al [53] Anklekneehip joint angles BPNN Average error less than9 deg

Jiang et al [5] Knee joint angle Four-layer FFNNmodel CC more than 0963

Anwar et al [54] Knee joint angle Generalized regressionneural network (GRNN) MSE less than 157

Mefoued [18] Knee joint angle RBFNN RMS less than 134 degZiai et al [55] Wrist joint torque ANN NRMSE less than 28Yokoyama et al [8] Handgrip-force ANN CCmore than 084Naeem et al [11] Arm muscle force BPNN CCmore than 099

Pena et al [19] Knee joint torque andstiffness

Multilayer perceptronneural network mdash

Chandrapal et al [56] Knee joint torque ANN Error more than 1046

Ardestani et al [57] Lower extremity jointmoment

Multi-dimensional waveletneural network (WNN)

NRMSE less than 10 andCCmore than 094

Khoshdel et al [4] Knee joint force Optimized ANN Error less than 345

For upper limb motion estimation Luh et al [47] esti-mated the angle of elbow joint by using BPNN The firstlayer consisted of sixteen filtered sEMG features nodes Thesecond hidden layer was constructed by 240 nodes andthe third layer had one angle output node The simulationresults illustrated that the proposed method was capable ofestimating the elbow angle with satisfactory accuracy Chenet al [48] adopted a hierarchical projection regression (HPR)for estimation of elbow angle using sEMGThe HPR projectsthe original date into a lower feature space to achieve a localrefined mapping between sEMG and the human motionAn average regression error of 98 deg was observed forpreliminary experiment Raj et al [49] utilized multilayeredperceptron neural network (MLPNN) and radial basis func-tion neural network (RBFNN) to identify the human forearmkinematics The features of IEMG and ZC were extractedas the input signals The results indicated that the RBFNNhave a better identification with an average CC of 076 and039 for angle and angular velocity respectively Wang et al[50] also utilized RBFNN to map the relationship betweensEMG and elbow joint angleTheGaussian function was usedbetween the input layer and hidden layer The experimentalresults showed that the RMSE and CC were around 0043

and 0905 respectively Kwon et al [51] estimated upper limbmotion by using feed forward neural network (FFNN) Thenetwork input terms were the MAV of sEMG and angularvelocities The output terms were estimated the angles ofelbow and shoulder joints Ngeo et al [52] employed FFNN tobuild the nonlinear relationship between finger joint anglesand sEMG signals The proposed network consisted of aneight-nodes input layer a tan-sigmoid hidden layer withactivation function and a fourteen-nodes linear output layerThe results showed that the correlation between predictedand actual finger joint angles was up to 092 And thirtyneurons were used in hidden layer The results showedthat the average NRMSE was around 85 deg Xia et al[13] implemented recurrent convolutional neural network(RCNN)which combined the properties of RNN andCNN toestimate the movement of upper limbs As shown in Figure 6the proposed RCNN architecture was composed of one inputlayer three convolutional layers two pooling layers two longshort-term memory (LSTM) layers and one output layer Anaverage CC of 93 was obtained for the proposed RCNNmethod

For lower limb motion estimation Zhang et al [53]employed BPNN to establish the mapping between sEMG




PoolingLayer 1


Long Short-Term

MemoryLayer 2

OutputLayer

Input Layer

Long Short-Term

MemoryLayer 1

PoolingLayer 2

Upper Limb Motion

Figure 6 The architecture of the recurrent convolutional neural network (RCNN) model [13]

and joint angles of ankle knee and hip The proposednetwork consisted of sixty-neuron input layer twenty-neuronhidden layer and three-neuron output layer The resultsshowed that the average error of different leg motions wasless than 9 deg Jiang et al [5] developed a sEMG based real-time control method The raw sEMG signal was processedand then input to a four-layer FFNN model to establishthe mapping relation between sEMG and knee angle In theproposed network five sEMG signals were the neurons ofinput layer and knee joint angle was the neuron of outputlayer The neuron number of the first hidden layer was 23and the second hidden layer was 13The results of preliminaryexperiment showed that the average value of CC was about0963 Anwar et al [54] estimated the knee joint angle basedon generalized regression neural network (GRNN) Theexperiment results illuminated that the MSE by using GRNNwith multiscale wavelet transform feature was around 157Mefoued [18] developed a RBFNN to map the nonlinearitiesbetween sEMG signal and desired knee angle The RBFNNarchitecture considered in this study was comprised of twoneurons in input layer five neurons in hidden layer and oneneuron in linear output layer And the nonlinear radial basisfunction was utilized as activation function The maximalRMS error of knee position estimation was equal to 134 deg

322 Mapping between SEMG and Joint Kinetics For jointkinetics regression the mapping between sEMG and jointforce or moment is commonly constructed On the one handthe estimated force ormoment is used as an input signal in thecontrol system of wearable robots to achieve accurate torquetrajectory track On the other hand the estimated momentand the estimated angles from the previous section are usedas the input signals to achieve accurate double closed-loopimpedance control Compared to joint kinematics regressionthe researches of kinetics regression are relatively rare

For upper limb motion estimation Ziai et al [55] esti-mated the wrist joint torques using sEMG based ANN Theproposed network used FFBPNN with one 8-neuron inputlayer two hidden layers and one torque output layer Anaverage NRMSE of 28 was observed Yokoyama et al [8]utilized sEMG based ANN to predict the handgrip-forceThe proposed network consisted of one input layer fourhidden layers and one output layer The RMS features fromfour sEMG signals were used as input layer and estimated

handgrip-force was used as output layer For the hiddenlayer 64 32 16 and 8 neurons were used in each hiddenlayer respectively The experimental results showed that theaverage CC was 084 between the predicted and observedforces Naeem et al [11] estimated human arm muscle forceby implementing a BPNNThe proposed network utilized therectified smoothed sEMG as input to generate the estimatedmuscle force as output The results illustrated that the CC ofthe proposed model and Hill-type model can exceed 099

For lower limb motion estimation Pena et al [19] pro-posed a multilayer perceptron neural network to map thesEMG signals to the knee torque and stiffness The inputsignals were the sEMG signals knee angle and angularvelocities and the output signals were estimated knee torqueand stiffness A second-order sliding mode control wasdeveloped to control the assistive device by using the desiredknee angle Chandrapal et al [56] established a mappingbetween five sEMG signals and knee torque by implementingANN There are three neurons in the hidden layer of themultilayer perceptron (MLP) and three neurons in the fullyconnected cascade (FCC) network The results showed thatthe mean lowest estimation error can achieve 1046 forthe proposed methods Ardestani et al [57] developed ageneric multidimensional wavelet neural network (WNN) topredict the moment of human lower extremity joint A totalof ten inputs including eight sEMG signals and two GRFcomponents were determined as the inputs for theWNN andthree joint moments of lower extremity were determined asthe output The results showed that the proposed WNN canestimate joint moments to a high level of accuracy NRMSEless than 10 and CC more than 094 Khoshdel et al [4]developed an optimized ANN (one input layer two hiddenlayers and one output layer) for knee force estimation Theinput layer consists of four preprocessed sEMG signals andthe output layer consisted of estimated force A total error of345 was obtained for the proposed optimized ANN

4 Conclusions

In this study the latest advanced researches in sEMG basedmotion intention recognition were discussed based on twomethods sEMG-driven musculoskeletal model and machinelearning based model For the sEMG-driven musculoskeletalmodel fundamental modelling theory and the performance


of models from different studies have been analyzed For themachine learning based model feature extraction and classi-fication model construction of discrete-motion classificationand mapping establishment between sEMG and joint kine-maticskinetics of continuous-motion regression have beendiscussed Additionally the advantages and disadvantages ofthe existed different motion intention recognition methodshave been discussed according to their different purposes inapplication

One can notice that it is hard to find a sEMG basedrecognition method that can estimate all human motionintentions completely and thoroughly Because of the lackof day-to-day repeatability and long training procedure thecurrent existed sEMG based motion intention recognitionmethods are still in the laboratory application stage and fewof them are truly marketized Deep learning based methodshave an important adverse impact on enhancing recognitionaccuracy and will become the trend of future development Ingeneral the proposed methods are only applicable to the spe-cific users and movement patterns Improving the robustnessand practicability of recognition methods is very importantAnd developing more precise and real-time human motionintention recognition methods will still be a crucial challengein the future

Conflicts of Interest

The authors declare that they have no conflicts of interest

References

[1] K A Strausser and H Kazerooni ldquoThe development andtesting of a human machine interface for a mobile medicalexoskeletonrdquo in Proceedings of the 2011 IEEERSJ InternationalConference on Intelligent Robots and Systems Celebrating 50Years of Robotics IROSrsquo11 pp 4911ndash4916 September 2011

[2] R M Singh S Chatterji and A Kumar ldquoA review on surfaceEMG based control schemes of exoskeleton robot in strokerehabilitationrdquo in Proceedings of the 2013 International Con-ference on Machine Intelligence Research and AdvancementICMIRA 2013 pp 310ndash315 India December 2013

[3] N Karavas A Ajoudani N Tsagarakis J Saglia A Bicchi andD Caldwell ldquoTele-impedance based assistive control for a com-pliant knee exoskeletonrdquoRobotics andAutonomous Systems vol73 pp 78ndash90 2015

[4] V Khoshdel and A Akbarzadeh ldquoAn optimized artificial neuralnetwork for human-force estimation consequences for rehabil-itation roboticsrdquo Industrial Robot An International Journal vol45 no 3 pp 416ndash423 2018

[5] J Jiang Z Zhang Z Wang and J Qian ldquoStudy on real-timecontrol of exoskeleton knee using electromyographic signalrdquoin Life System Modeling and Intelligent Computing vol 6330 ofLecture Notes in Computer Science pp 75ndash83 Springer BerlinHeidelberg 2010

[6] R A R C Gopura D S V Bandara K Kiguchi and G KI Mann ldquoDevelopments in hardware systems of active upper-limb exoskeleton robots a reviewrdquo Robotics and AutonomousSystems vol 75 pp 203ndash220 2016

[7] W Huo S Mohammed J C Moreno and Y Amirat ldquoLowerlimb wearable robots for assistance and rehabilitation a state ofthe artrdquo IEEE Systems Journal vol 10 no 3 pp 1068ndash1081 2016

[8] M Yokoyama R Koyama and M Yanagisawa ldquoAn evalua-tion of hand-force prediction using artificial neural-networkregressionmodels of surface emg signals for handwear devicesrdquoJournal of Sensors vol 2017 Article ID e3980906 12 pages 2017

[9] P Artemiadis ldquoEMG-based robot control interfaces pastpresent and futurerdquo Advances in Robotics amp Automation vol 1no 2 pp 1ndash3 2012

[10] M SartoriMReggianiD FarinaDG Lloyd andP LGribbleldquoEMG-driven forward-dynamic estimation of muscle force andjoint moment about multiple degrees of freedom in the humanlower extremityrdquo PLoS ONE vol 7 no 12 Article ID e52618 pp1ndash11 2012

[11] F Bian R Li and P Liang ldquoSVM based simultaneous handmovements classification using sEMG signalsrdquo in Proceedingsof the 14th IEEE International Conference on Mechatronics andAutomation ICMA 2017 pp 427ndash432 Japan August 2017

[12] F AlOmari and G Liu ldquoAnalysis of extracted forearm sEMGsignal using LDA QDA K-NN classification algorithmsrdquo OpenAutomation and Control Systems Journal vol 6 no 1 pp 108ndash116 2014

[13] P Xia J Hu and Y Peng ldquoEMG-based estimation of limbmovement using deep learning with recurrent convolutionalneural networksrdquo Artificial Organs vol 42 no 5 pp E67ndashE772017

[14] K-H Park and S-W Lee ldquoMovement intention decodingbased on deep learning for multiuser myoelectric interfacesrdquoin Proceedings of the 4th International Winter Conference onBrain-Computer Interface BCI 2016 pp 1-2 Republic of KoreaFebruary 2016

[15] K Asai and N Takase ldquoFinger motion estimation based onfrequency conversion of EMG signals and image recognitionusing convolutional neural networkrdquo in Proceedings of the 17thInternational Conference on Control Automation and SystemsICCAS 2017 pp 1366ndash1371 Republic of Korea October 2017

[16] N Nazmi M Abdul Rahman S Yamamoto S Ahmad HZamzuri and S Mazlan ldquoA review of classification techniquesof EMG signals during isotonic and isometric contractionsrdquoSensors vol 16 no 1304 pp 1ndash28 2016

[17] R H ChowdhuryM B I Reaz M A B Ali et al ldquoSurface elec-tromyography signal processing and classification techniquesrdquoSensors vol 13 no 9 pp 12431ndash12466 2013

[18] S Mefoued ldquoA second order sliding mode control and a neuralnetwork to drive a knee joint actuated orthosisrdquo Neurocomput-ing vol 155 pp 71ndash79 2015

[19] GG Pena L J ConsoniWM Santos andAA Siqueira ldquoFea-sibility of an optimal EMG-driven adaptive impedance controlapplied to an active knee orthosisrdquo Robotics and AutonomousSystems vol 112 pp 98ndash108 2019

[20] Q Ding A Xiong and X Zhao ldquoA review on researchesand applications of sEMG-based motion intent recognitionmethodsrdquoActa Automatics Sinica vol 42 no 1 pp 13ndash25 2016

[21] H Shen Q Song X Deng et al ldquoRecognition of phases in sit-to-standmotion byNeuralNetwork Ensemble (NNE) for powerassist robotrdquo in Proceedings of the 2007 IEEE InternationalConference on Robotics and Biomimetics ROBIO pp 1703ndash1708China December 2007

[22] D G Lloyd and T F Besier ldquoAn EMG-driven musculoskeletalmodel to estimate muscle forces and knee joint moments invivordquo Journal of Biomechanics vol 36 no 6 pp 765ndash776 2003


[23] J Han Q Ding A Xiong and X Zhao ldquoA state-space EMGmodel for the estimation of continuous joint movementsrdquo IEEETransactions on Industrial Electronics vol 62 no 7 pp 4267ndash4275 2015

[24] Q C Ding A B Xiong X G Zhao and J D Han ldquoA novelEMG-driven state spacemodel for the estimation of continuousjointmovementsrdquo in Proceedings of the International Conferenceon Systems pp 2891ndash2897 2011

[25] S Cai Y Chen S Huang et al ldquoSVM-based classificationof sENG signals for upper-limb self-rehabilitation trainingrdquoFrontiers in Neurorobotics vol 13 no 31 pp 1ndash10 2019

[26] M Atzori A Gijsberts C Castellini et al ldquoElectromyographydata for non-invasive naturally-controlled robotic hand pros-thesesrdquo Scientific Data vol 53 no 140053 2014

[27] A Oleinikov B Abibullaev A Shintemirov and M Fol-gheraiter ldquoFeature extraction and real-time recognition of handmotion intentions from EMGs via artificial neural networksrdquoin Proceedings of the 6th International Conference on Brain-Computer Interface BCI 2018 pp 1ndash5 Republic of KoreaJanuary 2018

[28] S Kyeong W D Kim J Feng and J Kim ldquoImplementationissues of EMG-basedmotion intentiondetection for exoskeletalrobotsrdquo inProceedings of the 27th IEEE International Symposiumon Robot and Human Interactive Communication RO-MAN2018 pp 915ndash920 China August 2018

[29] M R Ahsan M I Ibrahimy and O O Khalifa ldquoEMG motionpattern classification through design and optimization of neuralnetworkrdquo in Proceedings of the 2012 International Conferenceon Biomedical Engineering ICoBE 2012 pp 175ndash179 MalaysiaFebruary 2012

[30] M Gandolla S Ferrante G Ferrigno et al ldquoArtificial neuralnetwork EMG classifier for functional hand grasp movementspredictionrdquo Journal of International Medical Research vol 45no 6 pp 1831ndash1847 2017

[31] F A S Gomez D E G Villamarin W A R Ruiz et al ldquoCom-parison of advanced control techniques for motion intentionrecognition using EMG signalsrdquo in Proceedings of the 2017 IEEE3rd Colombian Conference on Automatic Control (CCAC) pp1ndash7 October 2017

[32] S Kyeong W Shin and J Kim ldquoPredicting Walking Intentionsusing sEMG andMechanical sensors for various environmentrdquoin Proceedings of the 40th Annual International Conference of theIEEE Engineering in Medicine and Biology Society EMBC 2018pp 4414ndash4417 USA July 2018

[33] S Pancholi A M Joshi et al ldquoPortable EMG data acquisitionmodule for upper limb prosthesis applicationrdquo IEEE SensorsJournal vol 18 no 8 pp 3436ndash3443 2018

[34] J L Pons Wearable Robots Biomechatronic Exoskeleton JohnWiley and Sons Ltd West Sussex 2008

[35] S M Mane R A Kambli F S Kazi and N M Singh ldquoHandmotion recognition from single channel surface EMG usingwavelet amp artificial neural networkrdquoProcedia Computer Sciencevol 49 pp 58ndash65 2015

[36] Babita P Kumari Y Narayan and L Mathew ldquoBinary move-ment classification of sEMG signal using linear SVM andWavelet Packet Transformrdquo in Proceedings of the 1st IEEE Inter-national Conference on Power Electronics Intelligent Control andEnergy Systems ICPEICES 2016 pp 1ndash4 India July 2016

[37] S Yang Y Chai J Ai S Sun and C Liu ldquoHand motionrecognition based on GA optimized SVM using sEMG signalsrdquoin Proceedings of the 2018 11th International Symposium on

Computational Intelligence and Design (ISCID) pp 146ndash149Hangzhou China December 2018

[38] X Sui K Wan Y Zhang et al ldquoPattern recognition of SEMGbased on wavelet packet transform and improved SVMrdquo Optik- International Journal for Light and Electron Optics vol 176 pp228ndash235 2019

[39] J Pan B Yang S Cai et al ldquoFinger motion pattern recognitionbased on sEMG support vector machinerdquo in Proceeding of theIEEE International Conference onCyborg andBionic Systems pp1ndash7 2017

[40] Y Chen Y Zhou X Cheng and Y Mi ldquoUpper limb motionrecognition based on two-step SVM classification method ofsurface EMGrdquo International Journal of Control and Automationvol 6 no 3 pp 249ndash266 2013

[41] G R Naik D K Kumar and Jayadeva ldquoTwin SVM forgesture classification using the surface electromyogramrdquo IEEETransactions on Information Technology in Biomedicine vol 14no 2 pp 301ndash308 2010

[42] J Liu X Sheng D Zhang and X Zhu ldquoBoosting trainingfor myoelectric pattern recognition using Mixed-LDArdquo inProceedings of the 2014 36th Annual International Conferenceof the IEEE Engineering in Medicine and Biology Society EMBC2014 pp 14ndash17 USA August 2014

[43] I S Dhindsa R Agarwal and H S Ryait ldquoPerformanceevaluation of various classifiers for predicting knee angle fromelectromyography signalsrdquo Expert Systems with Applicationsvol 11 pp 1ndash14 2019

[44] R J Oweis R Rihani and A Alkhawaja ldquoANN-based EMGclassification for myoelectric controlrdquo International Journal ofMedical Engineering and Informatics vol 6 no 4 pp 365ndash3802014

[45] N Bu O Fukuda and T Tsuji ldquoEMG-muscle motion dis-crimination using a novel recurrent neural networkrdquo Journal ofIntelligent Information Systems vol 21 no 2 pp 113ndash126 2003

[46] A D Orjuela-Canon A F Ruız-Olaya and L Forero ldquoDeepneural network for EMG signal classification of wrist positionpreliminary resultsrdquo in Proceedings of the 2017 IEEE LatinAmerican Conference on Computational Intelligence LA-CCI2017 pp 1ndash5 November 2017

[47] G-C Luh J-J Cai and Y-S Lee ldquoEstimation of elbowmotionintension under varing weight in lifting movement using anEMG-Angle neural network modelrdquo in Proceedings of the 16thInternational Conference on Machine Learning and CyberneticsICMLC 2017 pp 640ndash645 China July 2017

[48] Y Chen X Zhao and J Han ldquoHierarchical projection regres-sion for online estimation of elbow joint angle using EMGsignalsrdquoNeural Computing andApplications vol 23 no 3-4 pp1129ndash1138 2013

[49] R Raj R Rejith and K Sivanandan ldquoReal time identificationof human forearm kinematics from surface EMG signal usingartificial neural network modelsrdquo Procedia Technology vol 25pp 44ndash51 2016

[50] S Wang Y Gao J Zhao T Yang and Y Zhu ldquoPrediction ofsEMG-based tremor joint angle using the RBF neural networkrdquoin Proceedings of the 2012 9th IEEE International Conferenceon Mechatronics and Automation ICMA 2012 pp 2103ndash2108China August 2012

[51] S Kwon and J Kim ldquoReal-time upper limb motion estima-tion from surface electromyography and joint angular veloc-ities using an artificial neural network for humanndashmachinecooperationrdquo IEEE Transactions on Information Technology inBiomedicine vol 15 no 4 pp 522ndash530 2011


[52] J Ngeo T Tamei and T Shibata ldquoContinuous estimation offinger joint angles using muscle activation inputs from surfaceEMG signalsrdquo in Proceedings of the International Conference ofthe Engineering in Medicine and Biology Society IEEE pp 2756ndash2759 2012

[53] F Zhang P Li Z Hou et al ldquosEMG-based continuous estima-tion of joint angles of human legs by using BP neural networkrdquoNeurocomputing vol 78 no 1 pp 139ndash148 2012

[54] T Anwar Y M Aung and A Al Jumaily ldquoThe estimationof knee joint angle based on generalized regression neuralnetwork (GRNN)rdquo in Proceedings of the 2015 IEEE InternationalSymposium on Robotics and Intelligent Sensors (IRIS) pp 208ndash213 Langkawi Malaysia October 2015

[55] A Ziai and C Menon ldquoComparison of regression models forestimation of isometric wrist joint torques using surface elec-tromyographyrdquo Journal of NeuroEngineering and Rehabilitationvol 8 no 56 pp 1ndash12 2011

[56] M Chandrapal A Chen W H Wang et al ldquoInvestigatingimprovements to neural network based EMG to joint torqueestimationrdquo Journal of Behavioral Robotics vol 2 no 4 pp 185ndash192 2011

[57] M M Ardestani X Zhang L Wang et al ldquoHuman lowerextremity joint moment prediction a wavelet neural networkapproachrdquo Expert Systems with Applications vol 41 no 9 pp4422ndash4433 2014

International Journal of

AerospaceEngineeringHindawiwwwhindawicom Volume 2018

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018


Active and Passive Electronic Components

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawiwwwhindawicom

Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Control Scienceand Engineering

Journal of



Journal ofEngineeringVolume 2018

SensorsJournal of



RotatingMachinery


Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation


Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

Submit your manuscripts atwwwhindawicom


sEMGAcquisition Pre-processing Feature Extraction

KinematicsKinetics Date Acquisition

Discrete-Motion Classification

Continuous-Motion Regression

sEMG-Driven Musculoskeletal Model

Controller HMI

Motion Intention Recognition

Figure 1 The process of sEMG based human-machine interaction

Muscle Activation

Model

Muscle Contraction

Model

Musculoskeletal Geometry

(Moment Arms)

ProcessedEMGs Forward

Dynamicsai(t) Fmt

i (t)

agon(t)

antag(t) (t) (t) (t) (t)

int int

lmti (at)i (at)i(at)

Figure 2 The sEMG-driven musculoskeletal model [10]

of sEMG and evaluated the pros and cons of differentclassification models Singh et al [2] discussed the cur-rent development and challenges of the sEMG based con-trol schemes which are employed in designing exoskele-ton in stroke rehabilitation However the sEMG basedcontinuous-motion intention regression and sEMG-drivenmusculoskeletal model based motion intention recognitionare rarely reviewed And the two methods are more valuableto realize the smooth control of wearable robot movements[3 18 19] In order to further understand the knowledgeof the sEMG based motion intention recognition thispaper presents the review of all commonly used methodsof human motion intention recognition for last decadebriefly

The rest of this paper is organized as follows In Sec-tion 2 the motion intention recognition methods basedon the sEMG-driven musculoskeletal model is reviewed InSection 3 we discuss the various ML methods for discrete-motion classification and continuous-motion regressionIn Section 4 a succinct conclusion of this paper ispresented

2 sEMG-Driven MS Model Based MotionIntention Recognition

The sEMG is a nonstationary and microelectric signal whichamplitude is concentrated in 001-10mV and frequency isconcentrated in 20-500Hz especially in 50-150Hz [20]Because of about 30-150ms prior to the correspondingmotion generated sEMG is an ideal choice for motionintention estimation [20 21] Figure 1 shows the human-machine interaction process based on sEMG In the wholeprocess human motion intention recognition is the mostcritical part It can be achieved through two ways sEMG-driven MS model and ML model

The sEMG-driven musculoskeletal model can be dividedinto three submodels ie activation model contractionmodel and musculoskeletal geometry model as shown inFigure 2 [10] To serve as input to the model the raw sEMGsignal should be preprocessed by high-pass filtering full-wave rectification low-pass filtering and normalization [3]For the activation model the relationship between muscularactivation (119886119894(t)) and the processed sEMG signal (119906119894(119905)) of


Fmt(t)SE

lt(t)

lmt(t)

lm(t)

(t)

PE

CE

Fmt(t)



119886119894 (119905) = 119890119860119894119906119894 (119905) minus 1119890119860119894 minus 1 (1)


119865119898119905119894 (119905) = 119865119905119894 (119905)= 119865119898119886119909119894 [119891119894 (119897) 119891119894 (V) 119886119894 (119905) + 119891119901119894 (119897)] cos (120593119894 (119905))

(2)


119903119894 (119905) = 120597119897119898119905119894 (119905)120597120579 (3)


119897119898119905119894 (119905) = 119897119905119894 (119905) + 119897119898119894 (119905) cos (120593119894 (119905)) (4)


120591 (119905) =1003816100381610038161003816100381610038161003816100381610038161003816119899

sum119894=1

119865119894 (119905) 119903119894 (119905)1003816100381610038161003816100381610038161003816100381610038161003816119886119892119900119899119894119904119905minus10038161003816100381610038161003816100381610038161003816100381610038161003816119898

sum119895=1

119865119895 (119905) 119903119895 (119905)10038161003816100381610038161003816100381610038161003816100381610038161003816119886119899119905119886119892119900119899119894119904119905

(5)



120579 (119905) = (120591 (119905) minus 120591119890119909 (119905))119868 (6)


120579 (119905) = int 120579 (119905) 119889120579 (7)

120579 (119905) = int 120579 (119905) 119889120579 (8)




















Improved SVM 9066


RMS VAR WLMAV etc






RMS Twin SVM 8483







LDA NB K-NNand SVM

716 (LDA)751 (NB) 879(K-NN) and 922

(SVM)




MFMD





LDA RF NB andSVM

9167 (LDA)8750 (RF)

8683 (NB) and9225 (SVM)



DASDV

LDA QDA andK-NN

9856 (LDA)9342 (QDA) and9425 (K-NN)











9348


Table 1 Continued




neural network 90



recurrent ANN 884



EMG Processing

Feature Extraction



MAV RMS ZC VARWL PF









PoolingLayer 1


PoolingLayer 2


OutputLayer















velocity























PoolingLayer 1


Long Short-Term

MemoryLayer 2

OutputLayer

Input Layer

Long Short-Term

MemoryLayer 1

PoolingLayer 2

Upper Limb Motion







4 Conclusions







References































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



Fmt(t)SE

lt(t)

lmt(t)

lm(t)

(t)

PE

CE

Fmt(t)



119886119894 (119905) = 119890119860119894119906119894 (119905) minus 1119890119860119894 minus 1 (1)


119865119898119905119894 (119905) = 119865119905119894 (119905)= 119865119898119886119909119894 [119891119894 (119897) 119891119894 (V) 119886119894 (119905) + 119891119901119894 (119897)] cos (120593119894 (119905))

(2)


119903119894 (119905) = 120597119897119898119905119894 (119905)120597120579 (3)


119897119898119905119894 (119905) = 119897119905119894 (119905) + 119897119898119894 (119905) cos (120593119894 (119905)) (4)


120591 (119905) =1003816100381610038161003816100381610038161003816100381610038161003816119899

sum119894=1

119865119894 (119905) 119903119894 (119905)1003816100381610038161003816100381610038161003816100381610038161003816119886119892119900119899119894119904119905minus10038161003816100381610038161003816100381610038161003816100381610038161003816119898

sum119895=1

119865119895 (119905) 119903119895 (119905)10038161003816100381610038161003816100381610038161003816100381610038161003816119886119899119905119886119892119900119899119894119904119905

(5)



120579 (119905) = (120591 (119905) minus 120591119890119909 (119905))119868 (6)


120579 (119905) = int 120579 (119905) 119889120579 (7)

120579 (119905) = int 120579 (119905) 119889120579 (8)




















Improved SVM 9066


RMS VAR WLMAV etc






RMS Twin SVM 8483







LDA NB K-NNand SVM

716 (LDA)751 (NB) 879(K-NN) and 922

(SVM)




MFMD





LDA RF NB andSVM

9167 (LDA)8750 (RF)

8683 (NB) and9225 (SVM)



DASDV

LDA QDA andK-NN

9856 (LDA)9342 (QDA) and9425 (K-NN)











9348


Table 1 Continued




neural network 90



recurrent ANN 884



EMG Processing

Feature Extraction



MAV RMS ZC VARWL PF









PoolingLayer 1


PoolingLayer 2


OutputLayer















velocity























PoolingLayer 1


Long Short-Term

MemoryLayer 2

OutputLayer

Input Layer

Long Short-Term

MemoryLayer 1

PoolingLayer 2

Upper Limb Motion







4 Conclusions







References































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia


















Improved SVM 9066


RMS VAR WLMAV etc






RMS Twin SVM 8483







LDA NB K-NNand SVM

716 (LDA)751 (NB) 879(K-NN) and 922

(SVM)




MFMD





LDA RF NB andSVM

9167 (LDA)8750 (RF)

8683 (NB) and9225 (SVM)



DASDV

LDA QDA andK-NN

9856 (LDA)9342 (QDA) and9425 (K-NN)











9348


Table 1 Continued




neural network 90



recurrent ANN 884



EMG Processing

Feature Extraction



MAV RMS ZC VARWL PF









PoolingLayer 1


PoolingLayer 2


OutputLayer















velocity























PoolingLayer 1


Long Short-Term

MemoryLayer 2

OutputLayer

Input Layer

Long Short-Term

MemoryLayer 1

PoolingLayer 2

Upper Limb Motion







4 Conclusions







References































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



Table 1 Continued




neural network 90



recurrent ANN 884



EMG Processing

Feature Extraction



MAV RMS ZC VARWL PF









PoolingLayer 1


PoolingLayer 2


OutputLayer















velocity























PoolingLayer 1


Long Short-Term

MemoryLayer 2

OutputLayer

Input Layer

Long Short-Term

MemoryLayer 1

PoolingLayer 2

Upper Limb Motion







4 Conclusions







References































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




PoolingLayer 1


PoolingLayer 2


OutputLayer















velocity























PoolingLayer 1


Long Short-Term

MemoryLayer 2

OutputLayer

Input Layer

Long Short-Term

MemoryLayer 1

PoolingLayer 2

Upper Limb Motion







4 Conclusions







References































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia









velocity























PoolingLayer 1


Long Short-Term

MemoryLayer 2

OutputLayer

Input Layer

Long Short-Term

MemoryLayer 1

PoolingLayer 2

Upper Limb Motion







4 Conclusions







References































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia





PoolingLayer 1


Long Short-Term

MemoryLayer 2

OutputLayer

Input Layer

Long Short-Term

MemoryLayer 1

PoolingLayer 2

Upper Limb Motion







4 Conclusions







References































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia







References































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia










































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia


riewarticle semg based human motion intention...

Documents