Research ArticleHuman Activity Recognition Based on the Hierarchical FeatureSelection and Classification Framework

Yuhuang Zheng

Department of Physics Guangdong University of Education Guangzhou 510303 China

Correspondence should be addressed to Yuhuang Zheng zhyhaa126com

Received 31 January 2015 Revised 17 April 2015 Accepted 19 May 2015

Academic Editor Sos Agaian

Copyright copy 2015 Yuhuang Zheng This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Human activity recognition via triaxial accelerometers can provide valuable information for evaluating functional abilities In thispaper we present an accelerometer sensor-based approach for human activity recognition Our proposed recognition methodused a hierarchical scheme where the recognition of ten activity classes was divided into five distinct classification problems Everyclassifier used the Least Squares Support VectorMachine (LS-SVM) andNaive Bayes (NB) algorithm to distinguish different activityclasses The activity class was recognized based on the mean variance entropy of magnitude and angle of triaxial accelerometersignal features Our proposed activity recognition method recognized ten activities with an average accuracy of 956 using onlya single triaxial accelerometer

1 Introduction

Recently activity recognition has become an emerging field ofresearch and one of the challenges for pervasive computingA typical application for activity recognition is in healthcare Activity recognition is also an important research issuein building a pervasive and smart environment to providepersonalized support

Computer vision-based techniques and body-fixed accel-erators are the main methodologies used for activity recog-nition Computer vision-based techniques for activity recog-nition should be conducted in a well-controlled environmentand be subject to the limitations of the environment How-ever theymay significantly fail in an environmentwith clutterand variable lighting [1ndash3] Body-fixed accelerators offer apractical and relatively low-cost method to measure humanmotion

The existing literature demonstrates many studies onactivity recognition that use accelerometers However thereare three primary challenges in these studies

(1) The large muscles of the body are controlled for walk-ing running sitting and other activities The glutes are theprimary muscles that drive lower-body movement becauseof their natural strength and leverage advantage on the legsLower-body movement includes activities such as running

jumping and walking Sleeping sitting standing walkingrunning and jumping must be recognized as typical physicalactivities The activity recognition algorithm in Khan et al[4] did not consider jumping Running and jumping wereexcluded from the experiments in the research of Trabelsi etal [5] Tang and Sazonov [6] Lee et al [7] andDeng et al [8]Gupta and Dallas [9] did not report how to recognize stand-ing and sleeping and Tao et al [10] did not describe tests forrecognizing sitting and sleeping Alshurafa et al [11] studiedonly walking and running recognition These studies wereincomplete in recognizing typical physical activities [12]

(2) Some studies [6 9 13 14] required the combination ofmultiple sensors to increase recognition performance How-ever a user is less likely to wear a more complex operatingsystem at all times People may not feel comfortable wearingmultiple sensors Nevertheless the multisensor systems donot have an enormous advantage over the single-sensor sys-tem on the recognition accuracy if the single-sensor systemuses a higher sampling rate suitable features amore sophisti-cated classifier and the correct sensor position which has thebest performance for recognizing activities A single sensormounted at the right position can also obtain good recog-nition performance For typical physical activities multiplesensors are not helpful for significantly improving recogni-tion performance [15ndash17]

Hindawi Publishing CorporationJournal of Electrical and Computer EngineeringVolume 2015 Article ID 140820 9 pageshttpdxdoiorg1011552015140820

2 Journal of Electrical and Computer Engineering

(3) A series of lectures [18ndash20] have been given on thetopic of recognizing so-called ADL (activities of daily living)which is not physical-activity recognition ldquoActivities of dailylivingrdquo is a term used in healthcare to refer to daily self-careactivities such as cooking and hair drying within an individ-ualrsquos place of residence or in outdoor environments Physicalactivity included any body movement that works the musclesand requires more energy than resting and it simply impliesa movement of the body that uses energy such as runningor walking [21ndash23] Physical-activity recognition is discussedin this paper

Many researchers have used particular devices to collectthe raw accelerometer data for a set ofmovements and variousactivity recognition algorithms including Artificial NeuralNetworks (ANN) [4 7 13] 119896-Nearest Neighbor (119870NN) [810 11 19] Support Vector Machines (SVM) [6 14 18] andHidden Markov Model (HMM) [5 20] In our study weaddressed the activity recognition algorithm using SVM forthree reasons

(1) SVM and ANN have been broadly used in humanactivity recognition although they do not include a set ofrules understandable by humans [24] As two different algo-rithms SVM and ANN share the same concept of using thelinear learning model for pattern recognition The differenceis mainly on how nonlinear data are classified ConsequentlySVM models have preferable prediction performances toANN models SVMs have been demonstrated to have supe-rior classification accuracies to neural classifiers in manyexperiments The generalization performance of neural clas-sifiers considers the structure size and the selection of anappropriate structure relies on cross validation [25] Theperformance of SVMs depends on the selection of kernelfunction type and parameters but this dependence is lesseffective [26]

(2) 119870NN does not perform well when the size of datasetincreases and it is suitable for small datasets SVM is acomplicated classifier here we implement the leaner kernelfunction We conclude that the accuracy and other perfor-mance criteria do not significantly depend on the dataset sizebut they depend on the number of training cycles among allfactorsThe number of training cycles is the best classifier foractivity recognition [27]

(3) When a continuous HMM approach to activities isused the length of the event sequence that gives the bestpredictions uses sequential data A HMM is used to modelthe sequential information in multiaspect target signaturesThe parameter-learning task in HMMs is to determine thebest set of state transition and emission probabilities given anoutput sequence or a set of such sequencesThe task is usuallyto derive the maximum likelihood estimate of the parametersof the HMM for the set of output sequences Typical physicalactivities are nonsequential and it is not easy to use HMM torecognize a single physical activity [28]

The traditional SVM [29] is formulated for binary non-linear classification problems How to effectively extend theSVM for multiclassification remains a hot topic The LeastSquares Support Vector Machine (LS-SVM) is an advancedversion of the standard SVM and LS-SVM defines a differ-ent cost function from the classical SVM and changes its

inequation restriction to an equation restriction Recentlythere have been relatively few studies that use LS-SVM torecognize activities using a triaxial accelerometer Nasiriet al [30] addressed the Energy-Based Least Square TwinSupport Vector Machine (ELS-TSVM) algorithm which isan extended LS-SVM classifier that performs classificationusing two nonparallel hyper planes instead of a single hyperplane which is used in the conventional SVM ELS-TSVMwas used to recognize activities using computer vision insteadof a triaxial accelerometer Altun et al [31] compared theperformances of the least squares method (LSM) and theSVM but did not include the LS-SVM The LS-SVM formulticlassification is decomposed into multiple binary classi-fication tasksThe LS-SVM formulticlassification reduces thecomputational complexity by using a small number of classi-fiers and effectively eliminates the unclassifiable regions thatpossibly affect the classification performance of this algo-rithm [32ndash34]

In this paper we aimed to overcome the limitations of theexisting physical-activity recognition system and intended todevelop a new method that could recognize a set of typicalphysical activities using only a single triaxial accelerometerThis method consisted of three parts six features for activityrecognition the hierarchical recognition scheme and theactivity estimator based on the LS-SVM and NB algorithmsThis method could recognize ten physical activities with ahigh recognition rate

The remainder of the paper is organized as followsSection 2 describes the experimental dataset and hierarchicalclassification framework in this paper Section 3 involvesfeature extraction to improve the classification accuracyusing feature data over raw sensor data Section 4 focuseson an activity estimator for multiclassification to estimatethe human activity from the feature data The experimentalresults and conclusion are presented in Sections 5 and 6respectively

2 Hierarchical Classification Framework

21 Activities Dataset For this work the used dataset was theUniversity of Southern California Human Activity Dataset(USC-HAD) The USC-HAD was specifically designed toinclude themost basic and common human activities in dailylife from a large and diverse group of human subjects Theactivities in the dataset were applicable to many scenariosThe activity data were captured using a high-performanceinertial sensing device which isMotionNode [35] MotionN-ode integrates a 3-axis accelerometer a 3-axis gyroscope anda 3-axis magnetometer and the measurement range for eachaxis of the accelerometer and gyroscope is plusmn6 g and plusmn500 dpsrespectively MotionNode was firmly attached onto the par-ticipantrsquos right front hip The sampling rates of this datasetfor both accelerometer and gyroscope were set to 100HzThe dataset included 10 activities walking (forward leftand right) walking (upstairs downstairs) jumping runningstanding sitting and sleeping [36ndash38]

The main goal of this paper was to identify ten activitieswhich were divided into four groups 2D walking (walkingforward left and right) 3D walking (walking upstairs

Journal of Electrical and Computer Engineering 3

Table 1 Classified states and activities recognized in this study

State Activities Act Label

Walking 2DWalking forward WFWalking left WLWalking right WR

Walking 3D Walking upstairs WUWalking downstairs WD

Plane motion Jumping JURunning RU

Static activityStanding STSitting SISleeping SL

Table 2 A preliminary investigation of 119899 selection

119899-classclassifier

Numberof classifiers

Average accuracyrate of eachclassifier

2 9 ge903 6 ge904 5 ge905 3 asymp80

downstairs) plane motion (jumping running) and staticactivities (standing sitting and sleeping) The division wasperformed using a single triaxial accelerometerThe activitiesare listed in Table 1

22 Hierarchical Classification Framework To achieve higherscalability than the single-layer framework a multilayer clas-sification framework was presented In the first layer becausethe walking-related activities (walking forward walking leftwalking right walking upstairs and walking downstairs)jumping running and static activities were differentiatedfrom one another we classified the activities into two subsets(walking and all static activities) and two activities (jumpingand running) based on feature selection In the second layerthe walking-related activities subset included plane motionand 3Dmotion In this layer the static activity subset could beclassified by standing sitting and sleeping In the third layerall detailed activities of 2D and 3D walking were recognized[39 40]

Figure 1 illustrates the structure of the hierarchical clas-sification framework The yellow boxes represent the activityset and the green boxes represent the ten types of activitiesto recognize Now the problem of recognizing ten activityclasses was broken down to 119899 distinct classification problemsand the red boxes represent the classifiers A preliminaryinvestigation of 119899 selection is reported in Table 2 The four-class classifier was the best selection in this hierarchicalclassification framework because of the small number ofclassifiers and high average accuracy rate of each classifierThe four-class classifier was used in this paper

In the hierarchical classification framework of the four-class classifier classifier 1 at the top layer distinguishes

walking-related activities jumping running and static activ-ities Walking-related activities include walking forwardwalking left walking right walking upstairs and walkingdownstairs Static activities include standing sitting andsleeping [37] Classifier 2 at the second layer distinguishesplane motions and 3Dmotions Classifier 3 recognizes activi-ties from plane motion and classifier 4 distinguishes walkingupstairs and downstairs from 3Dmotions Finally classifier 5focuses on recognizing different static activities

3 Feature Design and Selection

Recent related work in feature selection was performedin a filter-based approach using Relief-F and a wrapper-based approach using a variant of sequential forward floatingsearch Because different features were on different scalesall features were normalized to obtain the best results for119870NN or Naive Bayes classifiers which were used for errorestimation and ensure equal weight to all potential features[1ndash6 8ndash10 13 18 24 29]

In our approach according to the elementary mechanicsof walking running jumping and sleeping we used themeans and variances of magnitudes and angles as the activityfeatures and the magnitudes and angles that were producedby a triaxial acceleration vectorThe reasons for this approachare as follows First according to [41ndash43] the musclesproduce different forces when people walk run jump andsleep Normally the forces increase in the order of sleepingwalking running and jumping Based on Newtonrsquos secondlaw the resultant accelerations of these activities also increasein that order Second as in [44] a model of persistent2D random walks can be represented by drawing turningangles Detailed features are described belowThird Shannonentropy in the time domain can measure the accelerationsignal uncertainty and describe the information-relatedproperties for an accurate representation of a given acceler-ation signal

The triaxial acceleration vector (119905) is

(119905) = 119909119886(119905) 119890119909+119910119886(119905) 119890119910+ 119911119886(119905) 119890119911 119905 = 1 119899 (1)

where 119909119886(119905) 119910

119886(119905) and 119911

119886(119905) represent the 119905 acceleration

sample of the 119909 119910 and 119911 axes This feature is independentof the orientation of the sensing device and measures theinstantaneous intensity of human movements at index 119905

We computed the mean variance and entropy of magni-tude and of the angle of over the window and used them assix features 119872mag 119881mag 119864mag 119872ang 119881ang and 119864ang where 119879

is the window length 120579 is the angle between vectors (119905 minus 1)and (119905) as shown in the following Let 119894 = 1 2 119899119879 then

119872mag = [119872mag (1) 119872mag (2) ]

where 119872mag (119894) =1119879

119879

sum

119905=(119894minus1)119879+1

10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816

4 Journal of Electrical and Computer Engineering

7 8

9

RU JU

WF

WL WR

WU WD

SI ST

SL

WF WL WR WU WDRU JU ST SI SL

1

32

4 5 6

(a) Two-class classifier

1

2 4

5 6

3

RU JU

WF WL WR WU WD

SI ST SL

WF WL WR WU WDRU JU ST SI SL

(b) Three-class classifier

1

2 3

4 5

Activity setsClassifierActivity

RU JU

WF WL WR WU WD

SI ST SL

WF WL WR WU WDRU JU ST SI SL

(c) Four-class classifier

1

2 3

Activity setsClassifierActivity

RU JU

WF WL WR WU WD SI ST SL

WF WL WR WU WDRU JU ST SI SL

(d) Five-class classifier

Figure 1 Structure of the hierarchical classification framework

119881mag = [119881mag (1) 119881mag (2) ]

where 119881mag (119894) =1119879

119879

sum

119905=(119894minus1)119879+1[10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816minus 119872mag (119894)]

2

119864mag = [119864mag (1) 119864mag (2) ]

where 119864mag (119894) = minus

119879

sum

119905=(119894minus1)119879+1[10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816

2log2(10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816

2)]

119872ang = [119872ang (1) 119872ang (2) ]

where 119872ang (119894) =1

119879 minus 1

119879

sum

119905=(119894minus1)119879+2120579 (119905)

119881ang = [119881ang (1) 119881ang (2) ]

where 119881ang (119894) =1

119879 minus 1

119879

sum

119905=(119894minus1)119879+2[120579 (119905) minus 119872ang (119894)]

2

119864ang = [119864ang (1) 119864ang (2) ]

where 119864ang (119894) = minus

119879

sum

119905=(119894minus1)119879+2[120579 (119905)

2 log2120579 (119905)

2]

(119905 minus 1) sdot (119905) =10038161003816100381610038161003816 (119905 minus 1)10038161003816100381610038161003816

10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816cos 120579 (119905)

(2)

To explore the performance and correlation among thesesix features a series of scatter plots in a 2D feature space isshown in Figure 2The horizontal and vertical axes representtwo different featuresThe points in different colors representdifferent activities In Figure 2(a) the relationship between119872mag and 119881mag is described and the running jumpingwalking and static activities are clustered In Figure 2(b)the straight line between 2D walking (forward left andright) and 3D walking (upstairs and downstairs) implies that119872ang is an available feature Figure 2(c) illustrates that the119864mag and 119872mag features successfully partition the triaxialacceleration data samples fromwalking forward walking left

Journal of Electrical and Computer Engineering 5

08 1 12 14 16 18 20

05

1

15

2

25

3

AccMean

AccV

ar

RunningJumping

Static activitiesWalking

Centroids

(a) 119872mag versus 119881mag

14 15 16 17 18 19 20

002

004

006

008

01

012

AccAngleMean

AccA

ngle

Var

Walking forward left and rightWalking upstairs downstairs

Centroids

(b) 119872ang versus 119881ang

AccM

ean

minus80 minus70 minus60 minus50 minus40 minus30 minus20 minus10 0 10095

1

105

11

115

12

125

AccEntropy

Walking forwardWalking left

Walking rightCentroids

(c) 119864mag versus119872mag

AccM

ean

minus280 minus240 minus200 minus160 minus120 minus8008

085

09

095

1

105

11

115

12

125

AccAngleEntropy

Walking upstairsWalking downstairs

Centroids

(d) 119864ang versus119872ang

AccM

ean

minus6 minus5 minus4 minus3 minus2 minus1 0 1099

1

101

102

103

104

105

106

107

AccEntropy

SittingSleeping

StandingCentroids

(e) 119864mag versus119872mag

Figure 2 Scatter plots in the 2D feature space (119879 = 50)

6 Journal of Electrical and Computer Engineering

Least Squares Support Vector Machines classifier

Mmag Vmag Emag Mang Vang Eang

Maximum Act_Labellikelihood estimation

resultrec

Act_Label1 Act_Label2 Act_Label1Activity =

Figure 3 Activity estimator for multiclassification

and walking right into three isolated clusters where eachcluster contains data samples roughly from one single activityclass Figure 2(d) demonstrates the discrimination power ofthe 119864ang and 119872ang features to differentiate walking upstairsand walking downstairs Figure 2(e) shows that the triaxialacceleration signal can be classified into standing sitting andsleeping based on the 119864mag and 119872mag features

In this study we used 119872mag 119881mag 119864mag 119872ang 119881ang and119864ang as the best features for the classifiers in each layer [45]

4 Activity Estimation for Multiclassification

We presented an activity estimator for multiclassificationto estimate the human activity from the feature data Eachactivity estimator for the multiclassification included oneLS-SVM classifier and a maximum Act Label frequencyestimator (Figure 3)

We used the LS-SVM [34] method to cluster the fea-ture data After loading the testing data into Matlab webuilt an activity-recognizing model from the data After theparameters of the model were calculated we estimated theactivity by inputting some test feature data [46]The functiontrainlssvm() was used to train the support features of an LS-SVM for classification and the function simlssvm() was usedto evaluate the LS-SVM for some test feature data

Because 119872mag 119881mag 119864mag 119872ang 119881ang and 119864ang have(119899119879) elements the LS-SVM for the multiclassifier outputsan activity set which includes 119899119879 elements of Act LabelThe activity set may have different Act Labels and we mustestimate the Act Label maximum likelihood in this activityset We used the Naive Bayes algorithm to compute allAct Label likelihoods and obtained the human activity usingthe maximum Act Label likelihood The following describedhow to mathematically compute the maximum Act Labellikelihood

119860119888119905119894V119894119905119910 = [ 119886119888119905119894 ]

= LS SVM (119872mag 119881mag 119864mag119872ang 119881ang 119864ang)

Raw accelerometer data(sample)

Accelerometer datanormalization normalization

Feature extraction

Training sample

Raw accelerometer data

Accelerometer data

Feature extraction

Testing data

Estimator for classification

Result

Training stage Testing stage

Figure 4 Activity estimator working process

119901 (119860119888119905 119871119886119887119890119897119895| 119888119897119886119904119904119894119891119894119890119903

119896)

= 119873 119861119886119910119890119904 119860119888119905119894V119894119905119910 | 119888119897119886119904119904119894119891119894119890119903119896

119895 = 1 10 119896 = 1 5

119903119890119904119906119897119905119903119890119888

= max 119901 (119886119888119905 =119860119888119905 119871119886119887119890119897119895| 119888119897119886119904119904119894119891119894119890119903

119896)

119895 = 1 10 119896 = 1 5 119886119888119905 isin 119860119888119905119894V119894119905119910

(3)

Figure 4 shows the activity estimator working processwhich includes the training stage and testing stage (onlineactivity recognition) In the training stage the labeled dataof triaxial acceleration were normalized and the statisticalfeatures were extracted from those synthesized-accelerationdataThen themulticlassification estimator was used to buildthe classification model In the testing stage unlabeled rawdata of the triaxial accelerometer were processed with themethod that was used in the training stageThese synthesizeddata were classified using the multiclassification estimatorand the recognized result was obtained [47 48]

5 Experiment

Theactivity recognition dataset was the USCHumanActivityDataset The activity dataset included ten activities andcollected data from 14 subjects To capture the day-to-dayactivity variations each subject was asked to perform 5 trialsfor each activity on different days at various indoor andoutdoor locations Although the duration of each trial variesfor different activities it was sufficiently long to capture allinformation for each performed activity [37] In this sectionwe estimated the performances of the five activity classifiersin this activity recognition scheme Table 3 shows the results

Journal of Electrical and Computer Engineering 7

Table 3 Activity classifier accuracy test

Classifiers Activities recognition accuracy rate () Classifier averageaccuracy ()WF WL WR WU WD RU JU ST SI SL

Classifier 1 983 971 971 100 982Classifier 2 981 991 mdash mdash mdash mdash mdash 986Classifier 3 986 971 957 mdash mdash mdash mdash mdash mdash mdash 971Classifier 4 mdash mdash mdash 971 971 mdash mdash mdash mdash mdash 971Classifier 5 mdash mdash mdash mdash mdash mdash mdash 986 971 986 981

Table 4 Confusion matrix for average recognition accuracy for allactivities

InputAccuracy rate () 956

OutputWF WL WR WU WD RU JU ST SI SL

WF 957 12 31 0 0 0 0 0 0 0WL 28 929 43 0 0 0 0 0 0 0WR 34 52 914 0 0 0 0 0 0 0WU 0 0 0 929 71 0 0 0 0 0WD 0 0 0 57 943 0 0 0 0 0RU 08 0 0 0 0 971 21 0 0 0JU 0 0 0 0 0 29 971 0 0 0ST 0 0 0 0 0 0 0 986 08 06SI 0 0 0 0 0 0 0 21 971 08SL 0 0 0 0 0 0 0 04 10 986

of five activity recognition classifiersThese activity classifiershad over 95 accuracy [24] and were acceptable

The results of these folds are summarized in Table 4The average recognition accuracy of 956 indicates that ourproposed human activity recognition scheme can achievehigh recognition rates for a specific subject Because 2Dwalking and 3Dwalking are similar the recognition accuracyof the five walking activities is low We will attempt to obtainhigher recognition accuracy using an adequate amount oftraining data in future research

We compared the accuracy rate and running time forcommon multiclassification methods All algorithms wererun on a computer with CPU i7-2670QM 22G 8G ram andMatlab 2013a The LS-SVM performed notably well in thetestsThe average running time for the hierarchical classifica-tion framework with the LS-SVM recognizing activities was0021 seconds whichwas less than theANN (Artificial NeuralNetwork) DT (Decision Tree) and 119870NN (119896-Nearest Neigh-bor) algorithmsWe performed the ANNDT and119870NNclas-sifier tests with the built-in functions of MatlabThe LS-SVMmethod was also better than ANN DT and 119870NN in termsof the average recognition accuracy rate for the ten activitiesTable 5 shows the results

6 Conclusion and Future Work

This paper aims to provide an accurate and robust humanactivity recognition scheme The scheme used triaxial

acceleration data a hierarchical recognition scheme andactivity classifiers based on the LS-SVM and the NB algo-rithm The mean variance entropy of magnitude and angleof triaxial acceleration data were used as the features ofthe activity classifiers The scheme effectively recognizeda typical set of daily physical activities with an averageaccuracy of 956 It could distinguish walking (forward leftright upstairs and downstairs) running jumping standingsitting and sleeping activities using only a single triaxialaccelerometer The experimental results of the hierarchicalrecognition scheme show significant potential in its abilityto accurately differentiate activities using triaxial accelerationdata Although the scheme remains to be tested with USC-HAD datasets the core of this scheme is independent of thefeatures of other activity datasets therefore it is applicable toany dataset

The novelty of the proposed human activity recognitionscheme is the introduction of the LS-SVMmethod as the clas-sifier algorithm The LS-SVM is an advanced version of thestandard SVM and there are recently relatively few studiesusing LS-SVM to recognize activities with only one triaxialaccelerometer The human activity recognition scheme withLS-SVM classifiers simplifies the construction of the hierar-chical classification framework and has a lower running timethan other commonmulticlassification algorithms Accuracyis the basic element thatmust be consideredwhen any activityrecognition system is implemented and this recognitionscheme has a high success rate for which it can recognizeten different types of activities with an average accuracy of95

The next stage of our research has two parts First thealgorithms are improved to recognize these activities and theuser will not have to worry about placing the sensors at thecorrect positions to correctly detect the activities Second anunsupervised approach for automatic activity recognition isconsidered An unsupervised learning framework of humanactivity recognition will automatically cluster a large amountof unlabeled acceleration data into discrete groups of activitywhich implies that the human activity recognition can benaturally performed

Conflict of Interests

The author declares that there is no conflict of interestsregarding the publication of this paper

8 Journal of Electrical and Computer Engineering

Table 5 Accuracy rates and running times of the classification methods

Method Accuracy rate () Average rate () Running time (s)WF WL WR WU WD JU RU ST SI SL

ANN 961 914 902 905 854 775 986 967 952 991 921 0085DT 939 946 917 912 908 849 941 957 972 944 929 0411119870NN 935 921 901 882 867 886 938 961 957 938 919 0183LS-SVM 957 929 914 929 943 971 971 986 971 986 956 0021

Acknowledgments

This work was partially supported by AppropriativeResearching Fund for Professors and Doctors GuangdongUniversity of Education under Grant 11ARF04 and Guang-dong Provincial Department of Education under Grants2013LYM 0063 and 2014GXJK161

References

[1] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014

[2] J Hernandez R Cabido A S Montemayor and J J PantrigoldquoHuman activity recognition based on kinematic featuresrdquoExpert Systems vol 31 no 4 pp 345ndash353 2014

[3] J Yin G Tian Z Feng and J Li ldquoHuman activity recognitionbased on multiple order temporal informationrdquo Computers ampElectrical Engineering vol 40 no 5 pp 1538ndash1551 2014

[4] A M Khan Y-K Lee S Y Lee and T-S Kim ldquoA triaxialaccelerometer-based physical-activity recognition via aug-mented-signal features and a hierarchical recognizerrdquo IEEETransactions on Information Technology in Biomedicine vol 14no 5 pp 1166ndash1172 2010

[5] D Trabelsi S Mohammed F Chamroukhi L Oukhellou andY Amirat ldquoAn unsupervised approach for automatic activityrecognition based on hidden markov model regressionrdquo IEEETransactions onAutomation Science and Engineering vol 10 no3 pp 829ndash835 2013

[6] W L Tang and E S Sazonov ldquoHighly accurate recognitionof human postures and activities through classification withrejectionrdquo IEEE Journal of Biomedical and Health Informaticsvol 18 no 1 pp 309ndash315 2014

[7] M-W Lee A M Khan and T-S Kim ldquoA single tri-axialaccelerometer-based real-time personal life log system capableof human activity recognition and exercise information gener-ationrdquo Personal amp Ubiquitous Computing vol 15 no 8 pp 887ndash898 2011

[8] W-Y Deng Q-H Zheng and Z-M Wang ldquoCross-personactivity recognition using reduced kernel extreme learningmachinerdquo Neural Networks vol 53 pp 1ndash7 2014

[9] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014

[10] D P Tao L Jin Y Wang and X Li ldquoRank preserving discrim-inant analysis for human behavior recognition on wireless sen-sor networksrdquo IEEE Transactions on Industrial Informatics vol10 no 1 pp 813ndash823 2014

[11] N Alshurafa W Xu J J Liu et al ldquoDesigning a robust activityrecognition framework for health and exergaming using wear-able sensorsrdquo IEEE Journal of Biomedical and Health Informat-ics vol 18 no 5 pp 1636ndash1646 2014

[12] A Chang S Mota and H Lieberman ldquoGestureNet a commonsense approach to physical activity similarityrdquo in Proceedings ofthe Conference on Electronic Visualisation and the Arts LondonUK July 2014

[13] O Banos M Damas H Pomares F Rojas B Delgado-Marquez and O Valenzuela ldquoHuman activity recognitionbased on a sensor weighting hierarchical classifierrdquo Soft Com-puting vol 17 no 2 pp 333ndash343 2013

[14] J Cheng X Chen andM Shen ldquoA framework for daily activitymonitoring and fall detection based on surface electromyogra-phy and accelerometer signalsrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 1 pp 38ndash45 2013

[15] N Kern B Schiele and A Schmidt ldquoMulti-sensor activity con-text detection for wearable computingrdquo in Ambient Intelligencevol 2875 of Lecture Notes in Computer Science pp 220ndash232Springer Berlin Germany 2003

[16] L Gao A K Bourke and J Nelson ldquoSensor positioning foractivity recognition using multiple accelerometer-based sen-sorsrdquo in Proceedings of the 21st European Symposium on Arti-ficial Neural Networks Computational Intelligence and MachineLearning pp 425ndash430 April 2013

[17] L Gao A K Bourke and J Nelson ldquoEvaluation of accelerome-ter based multi-sensor versus single-sensor activity recognitionsystemsrdquo Medical Engineering and Physics vol 36 no 6 pp779ndash785 2014

[18] S Liu R X Gao D John J W Staudenmayer and P S Freed-son ldquoMultisensor data fusion for physical activity assessmentrdquoIEEE Transactions on Biomedical Engineering vol 59 no 3 pp687ndash696 2012

[19] J Wan M J OrsquoGrady and G M P OrsquoHare ldquoDynamic sensorevent segmentation for real-time activity recognition in a smarthome contextrdquo Personal amp Ubiquitous Computing vol 19 no 2pp 287ndash301 2015

[20] Y Zhan and T Kuroda ldquoWearable sensor-based human activityrecognition from environmental background soundsrdquo Journalof Ambient Intelligence amp Humanized Computing vol 5 no 1pp 77ndash89 2014

[21] httpenwikipediaorgwikiPhysical exercise[22] httpwwwnhlbinihgovhealthhealth-topicstopicsphys[23] httpenwikipediaorgwikiActivities of daily living[24] O D Lara and M A Labrador ldquoA survey on human activity

recognition using wearable sensorsrdquo IEEE CommunicationsSurveys amp Tutorials vol 15 no 3 pp 1192ndash1209 2013

[25] S Arora D Bhattacharjee M Nasipuri L Malik M Kunduand D K Basu ldquoPerformance comparison of SVM and ANNfor handwritten devnagari character recognitionrdquo InternationalJournal of Computer Science Issues vol 18 pp 63ndash72 2010

Journal of Electrical and Computer Engineering 9

[26] J Ren ldquoANN vs SVM which one performs better in classifi-cation of MCCs in mammogram imagingrdquo Knowledge-BasedSystems vol 26 pp 144ndash153 2012

[27] J S Raikwal and K Saxena ldquoPerformance evaluation of SVMand k-nearest neighbor algorithm over medical data setrdquoInternational Journal of Computer Applications vol 50 no 14pp 12ndash24 2012

[28] M Eastwood and B Gabrys ldquoA non-sequential representationof sequential data for churn predictionrdquo in Knowledge-Basedand Intelligent Information and Engineering Systems pp 209ndash218 Springer Berlin Germany 2009

[29] Y Nam and J W Park ldquoChild activity recognition based oncooperative fusion model of a triaxial accelerometer and abarometric pressure sensorrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 2 pp 420ndash426 2013

[30] J A Nasiri N MoghadamCharkari and K Mozafari ldquoEnergy-based model of least squares twin Support Vector Machines forhuman action recognitionrdquo Signal Processing vol 104 pp 248ndash257 2014

[31] K Altun B Barshan and O Tuncel ldquoComparative studyon classifying human activities with miniature inertial andmagnetic sensorsrdquo Pattern Recognition vol 43 no 10 pp 3605ndash3620 2010

[32] R Wang S Kwong D Chen and J Cao ldquoA vector-valued sup-port vectormachinemodel formulticlass problemrdquo InformationSciences vol 235 pp 174ndash194 2013

[33] N Zhang and C Williams ldquoWater quantity prediction usingleast squares support vector machines (LSSVM) methodrdquoJournal of Systemics Cybernetics and Informatics vol 2 no 4pp 53ndash58 2014

[34] K D Brabanter and P Karsmakers ldquoLS-SVMlab Toolbox UserrsquosGuiderdquo 2011 httpwwwesatkuleuvenbesistalssvmlabdown-loadstutorialv1 8pdf

[35] httpwwwmotionnodecom[36] M Zhang and A A Sawchuk ldquoA feature selection-based frame-

work for human activity recognition using wearable multi-modal sensorsrdquo inProceedings of the International Conference onBody Area Networks (BodyNets rsquo11) Beijing China November2011

[37] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearable sen-sorsrdquo in Proceedings of the ACM International Conference onUbiquitous Computing (UbiComp rsquo12) International Workshopon Situation Activity and Goal Awareness Pittsburgh Pa USASeptember 2012

[38] httpsipiusceduHAD[39] O D Incel M Kose and C Ersoy ldquoA review and taxonomy of

activity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013

[40] Y Liang X Zhou Z Yu and B Guo ldquoEnergy-efficient motionrelated activity recognition on mobile devices for pervasivehealthcarerdquoMobile Networks and Applications vol 19 no 3 pp303ndash317 2014

[41] R Cross ldquoStanding walking running and jumping on a forceplaterdquo American Journal of Physics vol 67 no 4 pp 304ndash3091999

[42] A L Hof J P Van Zandwijk and M F Bobbert ldquoMechanics ofhuman triceps surae muscle in walking running and jumpingrdquoActa Physiologica Scandinavica vol 174 no 1 pp 17ndash30 2002

[43] G Cola A Vecchio and M Avvenuti ldquoImproving the per-formance of fall detection systems through walk recognitionrdquo

Journal of Ambient Intelligence amp Humanized Computing vol 5no 6 pp 843ndash855 2014

[44] H-I Wu B-L Li T A Springer and W H Neill ldquoModellinganimal movement as a persistent random walk in two dimen-sions expected magnitude of net displacementrdquo EcologicalModelling vol 132 no 1-2 pp 115ndash124 2000

[45] C Li M Lin L T Yang and C Ding ldquoIntegrating the enrichedfeaturewithmachine learning algorithms for humanmovementand fall detectionrdquo Journal of Supercomputing vol 67 no 3 pp854ndash865 2014

[46] N Zhang C Williams E Ososanya and W MahmoudldquoStreamflow Prediction Based on Least Squares Support VectorMachinesrdquo 2013 httpwwwaseeorgdocumentssectionsmid-dle-atlanticfall-201311-ASEE2013 Final20Zhangpdf

[47] D Rodriguez-Martin A Sama C Perez-Lopez A Catala JCabestany and A Rodriguez-Molinero ldquoSVM-based postureidentificationwith a single waist-located triaxial accelerometerrdquoExpert Systems with Applications vol 40 no 18 pp 7203ndash72112013

[48] J P Varkey D Pompili and T AWalls ldquoHumanmotion recog-nition using a wireless sensor-based wearable systemrdquo Personalamp Ubiquitous Computing vol 16 no 7 pp 897ndash910 2012

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

2 Journal of Electrical and Computer Engineering

(3) A series of lectures [18ndash20] have been given on thetopic of recognizing so-called ADL (activities of daily living)which is not physical-activity recognition ldquoActivities of dailylivingrdquo is a term used in healthcare to refer to daily self-careactivities such as cooking and hair drying within an individ-ualrsquos place of residence or in outdoor environments Physicalactivity included any body movement that works the musclesand requires more energy than resting and it simply impliesa movement of the body that uses energy such as runningor walking [21ndash23] Physical-activity recognition is discussedin this paper

Many researchers have used particular devices to collectthe raw accelerometer data for a set ofmovements and variousactivity recognition algorithms including Artificial NeuralNetworks (ANN) [4 7 13] 119896-Nearest Neighbor (119870NN) [810 11 19] Support Vector Machines (SVM) [6 14 18] andHidden Markov Model (HMM) [5 20] In our study weaddressed the activity recognition algorithm using SVM forthree reasons

(1) SVM and ANN have been broadly used in humanactivity recognition although they do not include a set ofrules understandable by humans [24] As two different algo-rithms SVM and ANN share the same concept of using thelinear learning model for pattern recognition The differenceis mainly on how nonlinear data are classified ConsequentlySVM models have preferable prediction performances toANN models SVMs have been demonstrated to have supe-rior classification accuracies to neural classifiers in manyexperiments The generalization performance of neural clas-sifiers considers the structure size and the selection of anappropriate structure relies on cross validation [25] Theperformance of SVMs depends on the selection of kernelfunction type and parameters but this dependence is lesseffective [26]

(2) 119870NN does not perform well when the size of datasetincreases and it is suitable for small datasets SVM is acomplicated classifier here we implement the leaner kernelfunction We conclude that the accuracy and other perfor-mance criteria do not significantly depend on the dataset sizebut they depend on the number of training cycles among allfactorsThe number of training cycles is the best classifier foractivity recognition [27]

(3) When a continuous HMM approach to activities isused the length of the event sequence that gives the bestpredictions uses sequential data A HMM is used to modelthe sequential information in multiaspect target signaturesThe parameter-learning task in HMMs is to determine thebest set of state transition and emission probabilities given anoutput sequence or a set of such sequencesThe task is usuallyto derive the maximum likelihood estimate of the parametersof the HMM for the set of output sequences Typical physicalactivities are nonsequential and it is not easy to use HMM torecognize a single physical activity [28]

The traditional SVM [29] is formulated for binary non-linear classification problems How to effectively extend theSVM for multiclassification remains a hot topic The LeastSquares Support Vector Machine (LS-SVM) is an advancedversion of the standard SVM and LS-SVM defines a differ-ent cost function from the classical SVM and changes its

inequation restriction to an equation restriction Recentlythere have been relatively few studies that use LS-SVM torecognize activities using a triaxial accelerometer Nasiriet al [30] addressed the Energy-Based Least Square TwinSupport Vector Machine (ELS-TSVM) algorithm which isan extended LS-SVM classifier that performs classificationusing two nonparallel hyper planes instead of a single hyperplane which is used in the conventional SVM ELS-TSVMwas used to recognize activities using computer vision insteadof a triaxial accelerometer Altun et al [31] compared theperformances of the least squares method (LSM) and theSVM but did not include the LS-SVM The LS-SVM formulticlassification is decomposed into multiple binary classi-fication tasksThe LS-SVM formulticlassification reduces thecomputational complexity by using a small number of classi-fiers and effectively eliminates the unclassifiable regions thatpossibly affect the classification performance of this algo-rithm [32ndash34]

In this paper we aimed to overcome the limitations of theexisting physical-activity recognition system and intended todevelop a new method that could recognize a set of typicalphysical activities using only a single triaxial accelerometerThis method consisted of three parts six features for activityrecognition the hierarchical recognition scheme and theactivity estimator based on the LS-SVM and NB algorithmsThis method could recognize ten physical activities with ahigh recognition rate

The remainder of the paper is organized as followsSection 2 describes the experimental dataset and hierarchicalclassification framework in this paper Section 3 involvesfeature extraction to improve the classification accuracyusing feature data over raw sensor data Section 4 focuseson an activity estimator for multiclassification to estimatethe human activity from the feature data The experimentalresults and conclusion are presented in Sections 5 and 6respectively

2 Hierarchical Classification Framework

21 Activities Dataset For this work the used dataset was theUniversity of Southern California Human Activity Dataset(USC-HAD) The USC-HAD was specifically designed toinclude themost basic and common human activities in dailylife from a large and diverse group of human subjects Theactivities in the dataset were applicable to many scenariosThe activity data were captured using a high-performanceinertial sensing device which isMotionNode [35] MotionN-ode integrates a 3-axis accelerometer a 3-axis gyroscope anda 3-axis magnetometer and the measurement range for eachaxis of the accelerometer and gyroscope is plusmn6 g and plusmn500 dpsrespectively MotionNode was firmly attached onto the par-ticipantrsquos right front hip The sampling rates of this datasetfor both accelerometer and gyroscope were set to 100HzThe dataset included 10 activities walking (forward leftand right) walking (upstairs downstairs) jumping runningstanding sitting and sleeping [36ndash38]

The main goal of this paper was to identify ten activitieswhich were divided into four groups 2D walking (walkingforward left and right) 3D walking (walking upstairs

Journal of Electrical and Computer Engineering 3

Table 1 Classified states and activities recognized in this study

State Activities Act Label

Walking 2DWalking forward WFWalking left WLWalking right WR

Walking 3D Walking upstairs WUWalking downstairs WD

Plane motion Jumping JURunning RU

Static activityStanding STSitting SISleeping SL

Table 2 A preliminary investigation of 119899 selection

119899-classclassifier

Numberof classifiers

Average accuracyrate of eachclassifier

2 9 ge903 6 ge904 5 ge905 3 asymp80

downstairs) plane motion (jumping running) and staticactivities (standing sitting and sleeping) The division wasperformed using a single triaxial accelerometerThe activitiesare listed in Table 1

22 Hierarchical Classification Framework To achieve higherscalability than the single-layer framework a multilayer clas-sification framework was presented In the first layer becausethe walking-related activities (walking forward walking leftwalking right walking upstairs and walking downstairs)jumping running and static activities were differentiatedfrom one another we classified the activities into two subsets(walking and all static activities) and two activities (jumpingand running) based on feature selection In the second layerthe walking-related activities subset included plane motionand 3Dmotion In this layer the static activity subset could beclassified by standing sitting and sleeping In the third layerall detailed activities of 2D and 3D walking were recognized[39 40]

Figure 1 illustrates the structure of the hierarchical clas-sification framework The yellow boxes represent the activityset and the green boxes represent the ten types of activitiesto recognize Now the problem of recognizing ten activityclasses was broken down to 119899 distinct classification problemsand the red boxes represent the classifiers A preliminaryinvestigation of 119899 selection is reported in Table 2 The four-class classifier was the best selection in this hierarchicalclassification framework because of the small number ofclassifiers and high average accuracy rate of each classifierThe four-class classifier was used in this paper

In the hierarchical classification framework of the four-class classifier classifier 1 at the top layer distinguishes

walking-related activities jumping running and static activ-ities Walking-related activities include walking forwardwalking left walking right walking upstairs and walkingdownstairs Static activities include standing sitting andsleeping [37] Classifier 2 at the second layer distinguishesplane motions and 3Dmotions Classifier 3 recognizes activi-ties from plane motion and classifier 4 distinguishes walkingupstairs and downstairs from 3Dmotions Finally classifier 5focuses on recognizing different static activities

3 Feature Design and Selection

Recent related work in feature selection was performedin a filter-based approach using Relief-F and a wrapper-based approach using a variant of sequential forward floatingsearch Because different features were on different scalesall features were normalized to obtain the best results for119870NN or Naive Bayes classifiers which were used for errorestimation and ensure equal weight to all potential features[1ndash6 8ndash10 13 18 24 29]

In our approach according to the elementary mechanicsof walking running jumping and sleeping we used themeans and variances of magnitudes and angles as the activityfeatures and the magnitudes and angles that were producedby a triaxial acceleration vectorThe reasons for this approachare as follows First according to [41ndash43] the musclesproduce different forces when people walk run jump andsleep Normally the forces increase in the order of sleepingwalking running and jumping Based on Newtonrsquos secondlaw the resultant accelerations of these activities also increasein that order Second as in [44] a model of persistent2D random walks can be represented by drawing turningangles Detailed features are described belowThird Shannonentropy in the time domain can measure the accelerationsignal uncertainty and describe the information-relatedproperties for an accurate representation of a given acceler-ation signal

The triaxial acceleration vector (119905) is

(119905) = 119909119886(119905) 119890119909+119910119886(119905) 119890119910+ 119911119886(119905) 119890119911 119905 = 1 119899 (1)

where 119909119886(119905) 119910

119886(119905) and 119911

119886(119905) represent the 119905 acceleration

sample of the 119909 119910 and 119911 axes This feature is independentof the orientation of the sensing device and measures theinstantaneous intensity of human movements at index 119905

We computed the mean variance and entropy of magni-tude and of the angle of over the window and used them assix features 119872mag 119881mag 119864mag 119872ang 119881ang and 119864ang where 119879

is the window length 120579 is the angle between vectors (119905 minus 1)and (119905) as shown in the following Let 119894 = 1 2 119899119879 then

119872mag = [119872mag (1) 119872mag (2) ]

where 119872mag (119894) =1119879

119879

sum

119905=(119894minus1)119879+1

10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816

4 Journal of Electrical and Computer Engineering

7 8

9

RU JU

WF

WL WR

WU WD

SI ST

SL

WF WL WR WU WDRU JU ST SI SL

1

32

4 5 6

(a) Two-class classifier

1

2 4

5 6

3

RU JU

WF WL WR WU WD

SI ST SL

WF WL WR WU WDRU JU ST SI SL

(b) Three-class classifier

1

2 3

4 5

Activity setsClassifierActivity

RU JU

WF WL WR WU WD

SI ST SL

WF WL WR WU WDRU JU ST SI SL

(c) Four-class classifier

1

2 3

Activity setsClassifierActivity

RU JU

WF WL WR WU WD SI ST SL

WF WL WR WU WDRU JU ST SI SL

(d) Five-class classifier

Figure 1 Structure of the hierarchical classification framework

119881mag = [119881mag (1) 119881mag (2) ]

where 119881mag (119894) =1119879

119879

sum

119905=(119894minus1)119879+1[10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816minus 119872mag (119894)]

2

119864mag = [119864mag (1) 119864mag (2) ]

where 119864mag (119894) = minus

119879

sum

119905=(119894minus1)119879+1[10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816

2log2(10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816

2)]

119872ang = [119872ang (1) 119872ang (2) ]

where 119872ang (119894) =1

119879 minus 1

119879

sum

119905=(119894minus1)119879+2120579 (119905)

119881ang = [119881ang (1) 119881ang (2) ]

where 119881ang (119894) =1

119879 minus 1

119879

sum

119905=(119894minus1)119879+2[120579 (119905) minus 119872ang (119894)]

2

119864ang = [119864ang (1) 119864ang (2) ]

where 119864ang (119894) = minus

119879

sum

119905=(119894minus1)119879+2[120579 (119905)

2 log2120579 (119905)

2]

(119905 minus 1) sdot (119905) =10038161003816100381610038161003816 (119905 minus 1)10038161003816100381610038161003816

10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816cos 120579 (119905)

(2)

To explore the performance and correlation among thesesix features a series of scatter plots in a 2D feature space isshown in Figure 2The horizontal and vertical axes representtwo different featuresThe points in different colors representdifferent activities In Figure 2(a) the relationship between119872mag and 119881mag is described and the running jumpingwalking and static activities are clustered In Figure 2(b)the straight line between 2D walking (forward left andright) and 3D walking (upstairs and downstairs) implies that119872ang is an available feature Figure 2(c) illustrates that the119864mag and 119872mag features successfully partition the triaxialacceleration data samples fromwalking forward walking left

Journal of Electrical and Computer Engineering 5

08 1 12 14 16 18 20

05

1

15

2

25

3

AccMean

AccV

ar

RunningJumping

Static activitiesWalking

Centroids

(a) 119872mag versus 119881mag

14 15 16 17 18 19 20

002

004

006

008

01

012

AccAngleMean

AccA

ngle

Var

Walking forward left and rightWalking upstairs downstairs

Centroids

(b) 119872ang versus 119881ang

AccM

ean

minus80 minus70 minus60 minus50 minus40 minus30 minus20 minus10 0 10095

1

105

11

115

12

125

AccEntropy

Walking forwardWalking left

Walking rightCentroids

(c) 119864mag versus119872mag

AccM

ean

minus280 minus240 minus200 minus160 minus120 minus8008

085

09

095

1

105

11

115

12

125

AccAngleEntropy

Walking upstairsWalking downstairs

Centroids

(d) 119864ang versus119872ang

AccM

ean

minus6 minus5 minus4 minus3 minus2 minus1 0 1099

1

101

102

103

104

105

106

107

AccEntropy

SittingSleeping

StandingCentroids

(e) 119864mag versus119872mag

Figure 2 Scatter plots in the 2D feature space (119879 = 50)

6 Journal of Electrical and Computer Engineering

Least Squares Support Vector Machines classifier

Mmag Vmag Emag Mang Vang Eang

Maximum Act_Labellikelihood estimation

resultrec

Act_Label1 Act_Label2 Act_Label1Activity =

Figure 3 Activity estimator for multiclassification

and walking right into three isolated clusters where eachcluster contains data samples roughly from one single activityclass Figure 2(d) demonstrates the discrimination power ofthe 119864ang and 119872ang features to differentiate walking upstairsand walking downstairs Figure 2(e) shows that the triaxialacceleration signal can be classified into standing sitting andsleeping based on the 119864mag and 119872mag features

In this study we used 119872mag 119881mag 119864mag 119872ang 119881ang and119864ang as the best features for the classifiers in each layer [45]

4 Activity Estimation for Multiclassification

We presented an activity estimator for multiclassificationto estimate the human activity from the feature data Eachactivity estimator for the multiclassification included oneLS-SVM classifier and a maximum Act Label frequencyestimator (Figure 3)

We used the LS-SVM [34] method to cluster the fea-ture data After loading the testing data into Matlab webuilt an activity-recognizing model from the data After theparameters of the model were calculated we estimated theactivity by inputting some test feature data [46]The functiontrainlssvm() was used to train the support features of an LS-SVM for classification and the function simlssvm() was usedto evaluate the LS-SVM for some test feature data

Because 119872mag 119881mag 119864mag 119872ang 119881ang and 119864ang have(119899119879) elements the LS-SVM for the multiclassifier outputsan activity set which includes 119899119879 elements of Act LabelThe activity set may have different Act Labels and we mustestimate the Act Label maximum likelihood in this activityset We used the Naive Bayes algorithm to compute allAct Label likelihoods and obtained the human activity usingthe maximum Act Label likelihood The following describedhow to mathematically compute the maximum Act Labellikelihood

119860119888119905119894V119894119905119910 = [ 119886119888119905119894 ]

= LS SVM (119872mag 119881mag 119864mag119872ang 119881ang 119864ang)

Raw accelerometer data(sample)

Accelerometer datanormalization normalization

Feature extraction

Training sample

Raw accelerometer data

Accelerometer data

Feature extraction

Testing data

Estimator for classification

Result

Training stage Testing stage

Figure 4 Activity estimator working process

119901 (119860119888119905 119871119886119887119890119897119895| 119888119897119886119904119904119894119891119894119890119903

119896)

= 119873 119861119886119910119890119904 119860119888119905119894V119894119905119910 | 119888119897119886119904119904119894119891119894119890119903119896

119895 = 1 10 119896 = 1 5

119903119890119904119906119897119905119903119890119888

= max 119901 (119886119888119905 =119860119888119905 119871119886119887119890119897119895| 119888119897119886119904119904119894119891119894119890119903

119896)

119895 = 1 10 119896 = 1 5 119886119888119905 isin 119860119888119905119894V119894119905119910

(3)

Figure 4 shows the activity estimator working processwhich includes the training stage and testing stage (onlineactivity recognition) In the training stage the labeled dataof triaxial acceleration were normalized and the statisticalfeatures were extracted from those synthesized-accelerationdataThen themulticlassification estimator was used to buildthe classification model In the testing stage unlabeled rawdata of the triaxial accelerometer were processed with themethod that was used in the training stageThese synthesizeddata were classified using the multiclassification estimatorand the recognized result was obtained [47 48]

5 Experiment

Theactivity recognition dataset was the USCHumanActivityDataset The activity dataset included ten activities andcollected data from 14 subjects To capture the day-to-dayactivity variations each subject was asked to perform 5 trialsfor each activity on different days at various indoor andoutdoor locations Although the duration of each trial variesfor different activities it was sufficiently long to capture allinformation for each performed activity [37] In this sectionwe estimated the performances of the five activity classifiersin this activity recognition scheme Table 3 shows the results

Journal of Electrical and Computer Engineering 7

Table 3 Activity classifier accuracy test

Classifiers Activities recognition accuracy rate () Classifier averageaccuracy ()WF WL WR WU WD RU JU ST SI SL

Classifier 1 983 971 971 100 982Classifier 2 981 991 mdash mdash mdash mdash mdash 986Classifier 3 986 971 957 mdash mdash mdash mdash mdash mdash mdash 971Classifier 4 mdash mdash mdash 971 971 mdash mdash mdash mdash mdash 971Classifier 5 mdash mdash mdash mdash mdash mdash mdash 986 971 986 981

Table 4 Confusion matrix for average recognition accuracy for allactivities

InputAccuracy rate () 956

OutputWF WL WR WU WD RU JU ST SI SL

WF 957 12 31 0 0 0 0 0 0 0WL 28 929 43 0 0 0 0 0 0 0WR 34 52 914 0 0 0 0 0 0 0WU 0 0 0 929 71 0 0 0 0 0WD 0 0 0 57 943 0 0 0 0 0RU 08 0 0 0 0 971 21 0 0 0JU 0 0 0 0 0 29 971 0 0 0ST 0 0 0 0 0 0 0 986 08 06SI 0 0 0 0 0 0 0 21 971 08SL 0 0 0 0 0 0 0 04 10 986

of five activity recognition classifiersThese activity classifiershad over 95 accuracy [24] and were acceptable

The results of these folds are summarized in Table 4The average recognition accuracy of 956 indicates that ourproposed human activity recognition scheme can achievehigh recognition rates for a specific subject Because 2Dwalking and 3Dwalking are similar the recognition accuracyof the five walking activities is low We will attempt to obtainhigher recognition accuracy using an adequate amount oftraining data in future research

We compared the accuracy rate and running time forcommon multiclassification methods All algorithms wererun on a computer with CPU i7-2670QM 22G 8G ram andMatlab 2013a The LS-SVM performed notably well in thetestsThe average running time for the hierarchical classifica-tion framework with the LS-SVM recognizing activities was0021 seconds whichwas less than theANN (Artificial NeuralNetwork) DT (Decision Tree) and 119870NN (119896-Nearest Neigh-bor) algorithmsWe performed the ANNDT and119870NNclas-sifier tests with the built-in functions of MatlabThe LS-SVMmethod was also better than ANN DT and 119870NN in termsof the average recognition accuracy rate for the ten activitiesTable 5 shows the results

6 Conclusion and Future Work

This paper aims to provide an accurate and robust humanactivity recognition scheme The scheme used triaxial

acceleration data a hierarchical recognition scheme andactivity classifiers based on the LS-SVM and the NB algo-rithm The mean variance entropy of magnitude and angleof triaxial acceleration data were used as the features ofthe activity classifiers The scheme effectively recognizeda typical set of daily physical activities with an averageaccuracy of 956 It could distinguish walking (forward leftright upstairs and downstairs) running jumping standingsitting and sleeping activities using only a single triaxialaccelerometer The experimental results of the hierarchicalrecognition scheme show significant potential in its abilityto accurately differentiate activities using triaxial accelerationdata Although the scheme remains to be tested with USC-HAD datasets the core of this scheme is independent of thefeatures of other activity datasets therefore it is applicable toany dataset

The novelty of the proposed human activity recognitionscheme is the introduction of the LS-SVMmethod as the clas-sifier algorithm The LS-SVM is an advanced version of thestandard SVM and there are recently relatively few studiesusing LS-SVM to recognize activities with only one triaxialaccelerometer The human activity recognition scheme withLS-SVM classifiers simplifies the construction of the hierar-chical classification framework and has a lower running timethan other commonmulticlassification algorithms Accuracyis the basic element thatmust be consideredwhen any activityrecognition system is implemented and this recognitionscheme has a high success rate for which it can recognizeten different types of activities with an average accuracy of95

The next stage of our research has two parts First thealgorithms are improved to recognize these activities and theuser will not have to worry about placing the sensors at thecorrect positions to correctly detect the activities Second anunsupervised approach for automatic activity recognition isconsidered An unsupervised learning framework of humanactivity recognition will automatically cluster a large amountof unlabeled acceleration data into discrete groups of activitywhich implies that the human activity recognition can benaturally performed

Conflict of Interests

The author declares that there is no conflict of interestsregarding the publication of this paper

8 Journal of Electrical and Computer Engineering

Table 5 Accuracy rates and running times of the classification methods

Method Accuracy rate () Average rate () Running time (s)WF WL WR WU WD JU RU ST SI SL

ANN 961 914 902 905 854 775 986 967 952 991 921 0085DT 939 946 917 912 908 849 941 957 972 944 929 0411119870NN 935 921 901 882 867 886 938 961 957 938 919 0183LS-SVM 957 929 914 929 943 971 971 986 971 986 956 0021

Acknowledgments

This work was partially supported by AppropriativeResearching Fund for Professors and Doctors GuangdongUniversity of Education under Grant 11ARF04 and Guang-dong Provincial Department of Education under Grants2013LYM 0063 and 2014GXJK161

References

[1] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014

[2] J Hernandez R Cabido A S Montemayor and J J PantrigoldquoHuman activity recognition based on kinematic featuresrdquoExpert Systems vol 31 no 4 pp 345ndash353 2014

[3] J Yin G Tian Z Feng and J Li ldquoHuman activity recognitionbased on multiple order temporal informationrdquo Computers ampElectrical Engineering vol 40 no 5 pp 1538ndash1551 2014

[4] A M Khan Y-K Lee S Y Lee and T-S Kim ldquoA triaxialaccelerometer-based physical-activity recognition via aug-mented-signal features and a hierarchical recognizerrdquo IEEETransactions on Information Technology in Biomedicine vol 14no 5 pp 1166ndash1172 2010

[5] D Trabelsi S Mohammed F Chamroukhi L Oukhellou andY Amirat ldquoAn unsupervised approach for automatic activityrecognition based on hidden markov model regressionrdquo IEEETransactions onAutomation Science and Engineering vol 10 no3 pp 829ndash835 2013

[6] W L Tang and E S Sazonov ldquoHighly accurate recognitionof human postures and activities through classification withrejectionrdquo IEEE Journal of Biomedical and Health Informaticsvol 18 no 1 pp 309ndash315 2014

[7] M-W Lee A M Khan and T-S Kim ldquoA single tri-axialaccelerometer-based real-time personal life log system capableof human activity recognition and exercise information gener-ationrdquo Personal amp Ubiquitous Computing vol 15 no 8 pp 887ndash898 2011

[8] W-Y Deng Q-H Zheng and Z-M Wang ldquoCross-personactivity recognition using reduced kernel extreme learningmachinerdquo Neural Networks vol 53 pp 1ndash7 2014

[9] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014

[10] D P Tao L Jin Y Wang and X Li ldquoRank preserving discrim-inant analysis for human behavior recognition on wireless sen-sor networksrdquo IEEE Transactions on Industrial Informatics vol10 no 1 pp 813ndash823 2014

[11] N Alshurafa W Xu J J Liu et al ldquoDesigning a robust activityrecognition framework for health and exergaming using wear-able sensorsrdquo IEEE Journal of Biomedical and Health Informat-ics vol 18 no 5 pp 1636ndash1646 2014

[12] A Chang S Mota and H Lieberman ldquoGestureNet a commonsense approach to physical activity similarityrdquo in Proceedings ofthe Conference on Electronic Visualisation and the Arts LondonUK July 2014

[13] O Banos M Damas H Pomares F Rojas B Delgado-Marquez and O Valenzuela ldquoHuman activity recognitionbased on a sensor weighting hierarchical classifierrdquo Soft Com-puting vol 17 no 2 pp 333ndash343 2013

[14] J Cheng X Chen andM Shen ldquoA framework for daily activitymonitoring and fall detection based on surface electromyogra-phy and accelerometer signalsrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 1 pp 38ndash45 2013

[15] N Kern B Schiele and A Schmidt ldquoMulti-sensor activity con-text detection for wearable computingrdquo in Ambient Intelligencevol 2875 of Lecture Notes in Computer Science pp 220ndash232Springer Berlin Germany 2003

[16] L Gao A K Bourke and J Nelson ldquoSensor positioning foractivity recognition using multiple accelerometer-based sen-sorsrdquo in Proceedings of the 21st European Symposium on Arti-ficial Neural Networks Computational Intelligence and MachineLearning pp 425ndash430 April 2013

[17] L Gao A K Bourke and J Nelson ldquoEvaluation of accelerome-ter based multi-sensor versus single-sensor activity recognitionsystemsrdquo Medical Engineering and Physics vol 36 no 6 pp779ndash785 2014

[18] S Liu R X Gao D John J W Staudenmayer and P S Freed-son ldquoMultisensor data fusion for physical activity assessmentrdquoIEEE Transactions on Biomedical Engineering vol 59 no 3 pp687ndash696 2012

[19] J Wan M J OrsquoGrady and G M P OrsquoHare ldquoDynamic sensorevent segmentation for real-time activity recognition in a smarthome contextrdquo Personal amp Ubiquitous Computing vol 19 no 2pp 287ndash301 2015

[20] Y Zhan and T Kuroda ldquoWearable sensor-based human activityrecognition from environmental background soundsrdquo Journalof Ambient Intelligence amp Humanized Computing vol 5 no 1pp 77ndash89 2014

[21] httpenwikipediaorgwikiPhysical exercise[22] httpwwwnhlbinihgovhealthhealth-topicstopicsphys[23] httpenwikipediaorgwikiActivities of daily living[24] O D Lara and M A Labrador ldquoA survey on human activity

recognition using wearable sensorsrdquo IEEE CommunicationsSurveys amp Tutorials vol 15 no 3 pp 1192ndash1209 2013

[25] S Arora D Bhattacharjee M Nasipuri L Malik M Kunduand D K Basu ldquoPerformance comparison of SVM and ANNfor handwritten devnagari character recognitionrdquo InternationalJournal of Computer Science Issues vol 18 pp 63ndash72 2010

Journal of Electrical and Computer Engineering 9

[26] J Ren ldquoANN vs SVM which one performs better in classifi-cation of MCCs in mammogram imagingrdquo Knowledge-BasedSystems vol 26 pp 144ndash153 2012

[27] J S Raikwal and K Saxena ldquoPerformance evaluation of SVMand k-nearest neighbor algorithm over medical data setrdquoInternational Journal of Computer Applications vol 50 no 14pp 12ndash24 2012

[28] M Eastwood and B Gabrys ldquoA non-sequential representationof sequential data for churn predictionrdquo in Knowledge-Basedand Intelligent Information and Engineering Systems pp 209ndash218 Springer Berlin Germany 2009

[29] Y Nam and J W Park ldquoChild activity recognition based oncooperative fusion model of a triaxial accelerometer and abarometric pressure sensorrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 2 pp 420ndash426 2013

[30] J A Nasiri N MoghadamCharkari and K Mozafari ldquoEnergy-based model of least squares twin Support Vector Machines forhuman action recognitionrdquo Signal Processing vol 104 pp 248ndash257 2014

[31] K Altun B Barshan and O Tuncel ldquoComparative studyon classifying human activities with miniature inertial andmagnetic sensorsrdquo Pattern Recognition vol 43 no 10 pp 3605ndash3620 2010

[32] R Wang S Kwong D Chen and J Cao ldquoA vector-valued sup-port vectormachinemodel formulticlass problemrdquo InformationSciences vol 235 pp 174ndash194 2013

[33] N Zhang and C Williams ldquoWater quantity prediction usingleast squares support vector machines (LSSVM) methodrdquoJournal of Systemics Cybernetics and Informatics vol 2 no 4pp 53ndash58 2014

[34] K D Brabanter and P Karsmakers ldquoLS-SVMlab Toolbox UserrsquosGuiderdquo 2011 httpwwwesatkuleuvenbesistalssvmlabdown-loadstutorialv1 8pdf

[35] httpwwwmotionnodecom[36] M Zhang and A A Sawchuk ldquoA feature selection-based frame-

work for human activity recognition using wearable multi-modal sensorsrdquo inProceedings of the International Conference onBody Area Networks (BodyNets rsquo11) Beijing China November2011

[37] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearable sen-sorsrdquo in Proceedings of the ACM International Conference onUbiquitous Computing (UbiComp rsquo12) International Workshopon Situation Activity and Goal Awareness Pittsburgh Pa USASeptember 2012

[38] httpsipiusceduHAD[39] O D Incel M Kose and C Ersoy ldquoA review and taxonomy of

activity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013

[40] Y Liang X Zhou Z Yu and B Guo ldquoEnergy-efficient motionrelated activity recognition on mobile devices for pervasivehealthcarerdquoMobile Networks and Applications vol 19 no 3 pp303ndash317 2014

[41] R Cross ldquoStanding walking running and jumping on a forceplaterdquo American Journal of Physics vol 67 no 4 pp 304ndash3091999

[42] A L Hof J P Van Zandwijk and M F Bobbert ldquoMechanics ofhuman triceps surae muscle in walking running and jumpingrdquoActa Physiologica Scandinavica vol 174 no 1 pp 17ndash30 2002

[43] G Cola A Vecchio and M Avvenuti ldquoImproving the per-formance of fall detection systems through walk recognitionrdquo

Journal of Ambient Intelligence amp Humanized Computing vol 5no 6 pp 843ndash855 2014

[44] H-I Wu B-L Li T A Springer and W H Neill ldquoModellinganimal movement as a persistent random walk in two dimen-sions expected magnitude of net displacementrdquo EcologicalModelling vol 132 no 1-2 pp 115ndash124 2000

[45] C Li M Lin L T Yang and C Ding ldquoIntegrating the enrichedfeaturewithmachine learning algorithms for humanmovementand fall detectionrdquo Journal of Supercomputing vol 67 no 3 pp854ndash865 2014

[46] N Zhang C Williams E Ososanya and W MahmoudldquoStreamflow Prediction Based on Least Squares Support VectorMachinesrdquo 2013 httpwwwaseeorgdocumentssectionsmid-dle-atlanticfall-201311-ASEE2013 Final20Zhangpdf

[47] D Rodriguez-Martin A Sama C Perez-Lopez A Catala JCabestany and A Rodriguez-Molinero ldquoSVM-based postureidentificationwith a single waist-located triaxial accelerometerrdquoExpert Systems with Applications vol 40 no 18 pp 7203ndash72112013

[48] J P Varkey D Pompili and T AWalls ldquoHumanmotion recog-nition using a wireless sensor-based wearable systemrdquo Personalamp Ubiquitous Computing vol 16 no 7 pp 897ndash910 2012

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Electrical and Computer Engineering 3

Table 1 Classified states and activities recognized in this study

State Activities Act Label

Walking 2DWalking forward WFWalking left WLWalking right WR

Walking 3D Walking upstairs WUWalking downstairs WD

Plane motion Jumping JURunning RU

Static activityStanding STSitting SISleeping SL

Table 2 A preliminary investigation of 119899 selection

119899-classclassifier

Numberof classifiers

Average accuracyrate of eachclassifier

2 9 ge903 6 ge904 5 ge905 3 asymp80

downstairs) plane motion (jumping running) and staticactivities (standing sitting and sleeping) The division wasperformed using a single triaxial accelerometerThe activitiesare listed in Table 1

22 Hierarchical Classification Framework To achieve higherscalability than the single-layer framework a multilayer clas-sification framework was presented In the first layer becausethe walking-related activities (walking forward walking leftwalking right walking upstairs and walking downstairs)jumping running and static activities were differentiatedfrom one another we classified the activities into two subsets(walking and all static activities) and two activities (jumpingand running) based on feature selection In the second layerthe walking-related activities subset included plane motionand 3Dmotion In this layer the static activity subset could beclassified by standing sitting and sleeping In the third layerall detailed activities of 2D and 3D walking were recognized[39 40]

Figure 1 illustrates the structure of the hierarchical clas-sification framework The yellow boxes represent the activityset and the green boxes represent the ten types of activitiesto recognize Now the problem of recognizing ten activityclasses was broken down to 119899 distinct classification problemsand the red boxes represent the classifiers A preliminaryinvestigation of 119899 selection is reported in Table 2 The four-class classifier was the best selection in this hierarchicalclassification framework because of the small number ofclassifiers and high average accuracy rate of each classifierThe four-class classifier was used in this paper

In the hierarchical classification framework of the four-class classifier classifier 1 at the top layer distinguishes

walking-related activities jumping running and static activ-ities Walking-related activities include walking forwardwalking left walking right walking upstairs and walkingdownstairs Static activities include standing sitting andsleeping [37] Classifier 2 at the second layer distinguishesplane motions and 3Dmotions Classifier 3 recognizes activi-ties from plane motion and classifier 4 distinguishes walkingupstairs and downstairs from 3Dmotions Finally classifier 5focuses on recognizing different static activities

3 Feature Design and Selection

Recent related work in feature selection was performedin a filter-based approach using Relief-F and a wrapper-based approach using a variant of sequential forward floatingsearch Because different features were on different scalesall features were normalized to obtain the best results for119870NN or Naive Bayes classifiers which were used for errorestimation and ensure equal weight to all potential features[1ndash6 8ndash10 13 18 24 29]

In our approach according to the elementary mechanicsof walking running jumping and sleeping we used themeans and variances of magnitudes and angles as the activityfeatures and the magnitudes and angles that were producedby a triaxial acceleration vectorThe reasons for this approachare as follows First according to [41ndash43] the musclesproduce different forces when people walk run jump andsleep Normally the forces increase in the order of sleepingwalking running and jumping Based on Newtonrsquos secondlaw the resultant accelerations of these activities also increasein that order Second as in [44] a model of persistent2D random walks can be represented by drawing turningangles Detailed features are described belowThird Shannonentropy in the time domain can measure the accelerationsignal uncertainty and describe the information-relatedproperties for an accurate representation of a given acceler-ation signal

The triaxial acceleration vector (119905) is

(119905) = 119909119886(119905) 119890119909+119910119886(119905) 119890119910+ 119911119886(119905) 119890119911 119905 = 1 119899 (1)

where 119909119886(119905) 119910

119886(119905) and 119911

119886(119905) represent the 119905 acceleration

sample of the 119909 119910 and 119911 axes This feature is independentof the orientation of the sensing device and measures theinstantaneous intensity of human movements at index 119905

We computed the mean variance and entropy of magni-tude and of the angle of over the window and used them assix features 119872mag 119881mag 119864mag 119872ang 119881ang and 119864ang where 119879

is the window length 120579 is the angle between vectors (119905 minus 1)and (119905) as shown in the following Let 119894 = 1 2 119899119879 then

119872mag = [119872mag (1) 119872mag (2) ]

where 119872mag (119894) =1119879

119879

sum

119905=(119894minus1)119879+1

10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816

4 Journal of Electrical and Computer Engineering

7 8

9

RU JU

WF

WL WR

WU WD

SI ST

SL

WF WL WR WU WDRU JU ST SI SL

1

32

4 5 6

(a) Two-class classifier

1

2 4

5 6

3

RU JU

WF WL WR WU WD

SI ST SL

WF WL WR WU WDRU JU ST SI SL

(b) Three-class classifier

1

2 3

4 5

Activity setsClassifierActivity

RU JU

WF WL WR WU WD

SI ST SL

WF WL WR WU WDRU JU ST SI SL

(c) Four-class classifier

1

2 3

Activity setsClassifierActivity

RU JU

WF WL WR WU WD SI ST SL

WF WL WR WU WDRU JU ST SI SL

(d) Five-class classifier

Figure 1 Structure of the hierarchical classification framework

119881mag = [119881mag (1) 119881mag (2) ]

where 119881mag (119894) =1119879

119879

sum

119905=(119894minus1)119879+1[10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816minus 119872mag (119894)]

2

119864mag = [119864mag (1) 119864mag (2) ]

where 119864mag (119894) = minus

119879

sum

119905=(119894minus1)119879+1[10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816

2log2(10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816

2)]

119872ang = [119872ang (1) 119872ang (2) ]

where 119872ang (119894) =1

119879 minus 1

119879

sum

119905=(119894minus1)119879+2120579 (119905)

119881ang = [119881ang (1) 119881ang (2) ]

where 119881ang (119894) =1

119879 minus 1

119879

sum

119905=(119894minus1)119879+2[120579 (119905) minus 119872ang (119894)]

2

119864ang = [119864ang (1) 119864ang (2) ]

where 119864ang (119894) = minus

119879

sum

119905=(119894minus1)119879+2[120579 (119905)

2 log2120579 (119905)

2]

(119905 minus 1) sdot (119905) =10038161003816100381610038161003816 (119905 minus 1)10038161003816100381610038161003816

10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816cos 120579 (119905)

(2)

To explore the performance and correlation among thesesix features a series of scatter plots in a 2D feature space isshown in Figure 2The horizontal and vertical axes representtwo different featuresThe points in different colors representdifferent activities In Figure 2(a) the relationship between119872mag and 119881mag is described and the running jumpingwalking and static activities are clustered In Figure 2(b)the straight line between 2D walking (forward left andright) and 3D walking (upstairs and downstairs) implies that119872ang is an available feature Figure 2(c) illustrates that the119864mag and 119872mag features successfully partition the triaxialacceleration data samples fromwalking forward walking left

Journal of Electrical and Computer Engineering 5

08 1 12 14 16 18 20

05

1

15

2

25

3

AccMean

AccV

ar

RunningJumping

Static activitiesWalking

Centroids

(a) 119872mag versus 119881mag

14 15 16 17 18 19 20

002

004

006

008

01

012

AccAngleMean

AccA

ngle

Var

Walking forward left and rightWalking upstairs downstairs

Centroids

(b) 119872ang versus 119881ang

AccM

ean

minus80 minus70 minus60 minus50 minus40 minus30 minus20 minus10 0 10095

1

105

11

115

12

125

AccEntropy

Walking forwardWalking left

Walking rightCentroids

(c) 119864mag versus119872mag

AccM

ean

minus280 minus240 minus200 minus160 minus120 minus8008

085

09

095

1

105

11

115

12

125

AccAngleEntropy

Walking upstairsWalking downstairs

Centroids

(d) 119864ang versus119872ang

AccM

ean

minus6 minus5 minus4 minus3 minus2 minus1 0 1099

1

101

102

103

104

105

106

107

AccEntropy

SittingSleeping

StandingCentroids

(e) 119864mag versus119872mag

Figure 2 Scatter plots in the 2D feature space (119879 = 50)

6 Journal of Electrical and Computer Engineering

Least Squares Support Vector Machines classifier

Mmag Vmag Emag Mang Vang Eang

Maximum Act_Labellikelihood estimation

resultrec

Act_Label1 Act_Label2 Act_Label1Activity =

Figure 3 Activity estimator for multiclassification

and walking right into three isolated clusters where eachcluster contains data samples roughly from one single activityclass Figure 2(d) demonstrates the discrimination power ofthe 119864ang and 119872ang features to differentiate walking upstairsand walking downstairs Figure 2(e) shows that the triaxialacceleration signal can be classified into standing sitting andsleeping based on the 119864mag and 119872mag features

In this study we used 119872mag 119881mag 119864mag 119872ang 119881ang and119864ang as the best features for the classifiers in each layer [45]

4 Activity Estimation for Multiclassification

We presented an activity estimator for multiclassificationto estimate the human activity from the feature data Eachactivity estimator for the multiclassification included oneLS-SVM classifier and a maximum Act Label frequencyestimator (Figure 3)

We used the LS-SVM [34] method to cluster the fea-ture data After loading the testing data into Matlab webuilt an activity-recognizing model from the data After theparameters of the model were calculated we estimated theactivity by inputting some test feature data [46]The functiontrainlssvm() was used to train the support features of an LS-SVM for classification and the function simlssvm() was usedto evaluate the LS-SVM for some test feature data

Because 119872mag 119881mag 119864mag 119872ang 119881ang and 119864ang have(119899119879) elements the LS-SVM for the multiclassifier outputsan activity set which includes 119899119879 elements of Act LabelThe activity set may have different Act Labels and we mustestimate the Act Label maximum likelihood in this activityset We used the Naive Bayes algorithm to compute allAct Label likelihoods and obtained the human activity usingthe maximum Act Label likelihood The following describedhow to mathematically compute the maximum Act Labellikelihood

119860119888119905119894V119894119905119910 = [ 119886119888119905119894 ]

= LS SVM (119872mag 119881mag 119864mag119872ang 119881ang 119864ang)

Raw accelerometer data(sample)

Accelerometer datanormalization normalization

Feature extraction

Training sample

Raw accelerometer data

Accelerometer data

Feature extraction

Testing data

Estimator for classification

Result

Training stage Testing stage

Figure 4 Activity estimator working process

119901 (119860119888119905 119871119886119887119890119897119895| 119888119897119886119904119904119894119891119894119890119903

119896)

= 119873 119861119886119910119890119904 119860119888119905119894V119894119905119910 | 119888119897119886119904119904119894119891119894119890119903119896

119895 = 1 10 119896 = 1 5

119903119890119904119906119897119905119903119890119888

= max 119901 (119886119888119905 =119860119888119905 119871119886119887119890119897119895| 119888119897119886119904119904119894119891119894119890119903

119896)

119895 = 1 10 119896 = 1 5 119886119888119905 isin 119860119888119905119894V119894119905119910

(3)

Figure 4 shows the activity estimator working processwhich includes the training stage and testing stage (onlineactivity recognition) In the training stage the labeled dataof triaxial acceleration were normalized and the statisticalfeatures were extracted from those synthesized-accelerationdataThen themulticlassification estimator was used to buildthe classification model In the testing stage unlabeled rawdata of the triaxial accelerometer were processed with themethod that was used in the training stageThese synthesizeddata were classified using the multiclassification estimatorand the recognized result was obtained [47 48]

5 Experiment

Theactivity recognition dataset was the USCHumanActivityDataset The activity dataset included ten activities andcollected data from 14 subjects To capture the day-to-dayactivity variations each subject was asked to perform 5 trialsfor each activity on different days at various indoor andoutdoor locations Although the duration of each trial variesfor different activities it was sufficiently long to capture allinformation for each performed activity [37] In this sectionwe estimated the performances of the five activity classifiersin this activity recognition scheme Table 3 shows the results

Journal of Electrical and Computer Engineering 7

Table 3 Activity classifier accuracy test

Classifiers Activities recognition accuracy rate () Classifier averageaccuracy ()WF WL WR WU WD RU JU ST SI SL

Classifier 1 983 971 971 100 982Classifier 2 981 991 mdash mdash mdash mdash mdash 986Classifier 3 986 971 957 mdash mdash mdash mdash mdash mdash mdash 971Classifier 4 mdash mdash mdash 971 971 mdash mdash mdash mdash mdash 971Classifier 5 mdash mdash mdash mdash mdash mdash mdash 986 971 986 981

Table 4 Confusion matrix for average recognition accuracy for allactivities

InputAccuracy rate () 956

OutputWF WL WR WU WD RU JU ST SI SL

WF 957 12 31 0 0 0 0 0 0 0WL 28 929 43 0 0 0 0 0 0 0WR 34 52 914 0 0 0 0 0 0 0WU 0 0 0 929 71 0 0 0 0 0WD 0 0 0 57 943 0 0 0 0 0RU 08 0 0 0 0 971 21 0 0 0JU 0 0 0 0 0 29 971 0 0 0ST 0 0 0 0 0 0 0 986 08 06SI 0 0 0 0 0 0 0 21 971 08SL 0 0 0 0 0 0 0 04 10 986

of five activity recognition classifiersThese activity classifiershad over 95 accuracy [24] and were acceptable

The results of these folds are summarized in Table 4The average recognition accuracy of 956 indicates that ourproposed human activity recognition scheme can achievehigh recognition rates for a specific subject Because 2Dwalking and 3Dwalking are similar the recognition accuracyof the five walking activities is low We will attempt to obtainhigher recognition accuracy using an adequate amount oftraining data in future research

We compared the accuracy rate and running time forcommon multiclassification methods All algorithms wererun on a computer with CPU i7-2670QM 22G 8G ram andMatlab 2013a The LS-SVM performed notably well in thetestsThe average running time for the hierarchical classifica-tion framework with the LS-SVM recognizing activities was0021 seconds whichwas less than theANN (Artificial NeuralNetwork) DT (Decision Tree) and 119870NN (119896-Nearest Neigh-bor) algorithmsWe performed the ANNDT and119870NNclas-sifier tests with the built-in functions of MatlabThe LS-SVMmethod was also better than ANN DT and 119870NN in termsof the average recognition accuracy rate for the ten activitiesTable 5 shows the results

6 Conclusion and Future Work

This paper aims to provide an accurate and robust humanactivity recognition scheme The scheme used triaxial

acceleration data a hierarchical recognition scheme andactivity classifiers based on the LS-SVM and the NB algo-rithm The mean variance entropy of magnitude and angleof triaxial acceleration data were used as the features ofthe activity classifiers The scheme effectively recognizeda typical set of daily physical activities with an averageaccuracy of 956 It could distinguish walking (forward leftright upstairs and downstairs) running jumping standingsitting and sleeping activities using only a single triaxialaccelerometer The experimental results of the hierarchicalrecognition scheme show significant potential in its abilityto accurately differentiate activities using triaxial accelerationdata Although the scheme remains to be tested with USC-HAD datasets the core of this scheme is independent of thefeatures of other activity datasets therefore it is applicable toany dataset

The novelty of the proposed human activity recognitionscheme is the introduction of the LS-SVMmethod as the clas-sifier algorithm The LS-SVM is an advanced version of thestandard SVM and there are recently relatively few studiesusing LS-SVM to recognize activities with only one triaxialaccelerometer The human activity recognition scheme withLS-SVM classifiers simplifies the construction of the hierar-chical classification framework and has a lower running timethan other commonmulticlassification algorithms Accuracyis the basic element thatmust be consideredwhen any activityrecognition system is implemented and this recognitionscheme has a high success rate for which it can recognizeten different types of activities with an average accuracy of95

The next stage of our research has two parts First thealgorithms are improved to recognize these activities and theuser will not have to worry about placing the sensors at thecorrect positions to correctly detect the activities Second anunsupervised approach for automatic activity recognition isconsidered An unsupervised learning framework of humanactivity recognition will automatically cluster a large amountof unlabeled acceleration data into discrete groups of activitywhich implies that the human activity recognition can benaturally performed

Conflict of Interests

The author declares that there is no conflict of interestsregarding the publication of this paper

8 Journal of Electrical and Computer Engineering

Table 5 Accuracy rates and running times of the classification methods

Method Accuracy rate () Average rate () Running time (s)WF WL WR WU WD JU RU ST SI SL

ANN 961 914 902 905 854 775 986 967 952 991 921 0085DT 939 946 917 912 908 849 941 957 972 944 929 0411119870NN 935 921 901 882 867 886 938 961 957 938 919 0183LS-SVM 957 929 914 929 943 971 971 986 971 986 956 0021

Acknowledgments

This work was partially supported by AppropriativeResearching Fund for Professors and Doctors GuangdongUniversity of Education under Grant 11ARF04 and Guang-dong Provincial Department of Education under Grants2013LYM 0063 and 2014GXJK161

References

[1] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014

[2] J Hernandez R Cabido A S Montemayor and J J PantrigoldquoHuman activity recognition based on kinematic featuresrdquoExpert Systems vol 31 no 4 pp 345ndash353 2014

[3] J Yin G Tian Z Feng and J Li ldquoHuman activity recognitionbased on multiple order temporal informationrdquo Computers ampElectrical Engineering vol 40 no 5 pp 1538ndash1551 2014

[4] A M Khan Y-K Lee S Y Lee and T-S Kim ldquoA triaxialaccelerometer-based physical-activity recognition via aug-mented-signal features and a hierarchical recognizerrdquo IEEETransactions on Information Technology in Biomedicine vol 14no 5 pp 1166ndash1172 2010

[5] D Trabelsi S Mohammed F Chamroukhi L Oukhellou andY Amirat ldquoAn unsupervised approach for automatic activityrecognition based on hidden markov model regressionrdquo IEEETransactions onAutomation Science and Engineering vol 10 no3 pp 829ndash835 2013

[6] W L Tang and E S Sazonov ldquoHighly accurate recognitionof human postures and activities through classification withrejectionrdquo IEEE Journal of Biomedical and Health Informaticsvol 18 no 1 pp 309ndash315 2014

[7] M-W Lee A M Khan and T-S Kim ldquoA single tri-axialaccelerometer-based real-time personal life log system capableof human activity recognition and exercise information gener-ationrdquo Personal amp Ubiquitous Computing vol 15 no 8 pp 887ndash898 2011

[8] W-Y Deng Q-H Zheng and Z-M Wang ldquoCross-personactivity recognition using reduced kernel extreme learningmachinerdquo Neural Networks vol 53 pp 1ndash7 2014

[9] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014

[10] D P Tao L Jin Y Wang and X Li ldquoRank preserving discrim-inant analysis for human behavior recognition on wireless sen-sor networksrdquo IEEE Transactions on Industrial Informatics vol10 no 1 pp 813ndash823 2014

[11] N Alshurafa W Xu J J Liu et al ldquoDesigning a robust activityrecognition framework for health and exergaming using wear-able sensorsrdquo IEEE Journal of Biomedical and Health Informat-ics vol 18 no 5 pp 1636ndash1646 2014

[12] A Chang S Mota and H Lieberman ldquoGestureNet a commonsense approach to physical activity similarityrdquo in Proceedings ofthe Conference on Electronic Visualisation and the Arts LondonUK July 2014

[13] O Banos M Damas H Pomares F Rojas B Delgado-Marquez and O Valenzuela ldquoHuman activity recognitionbased on a sensor weighting hierarchical classifierrdquo Soft Com-puting vol 17 no 2 pp 333ndash343 2013

[14] J Cheng X Chen andM Shen ldquoA framework for daily activitymonitoring and fall detection based on surface electromyogra-phy and accelerometer signalsrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 1 pp 38ndash45 2013

[15] N Kern B Schiele and A Schmidt ldquoMulti-sensor activity con-text detection for wearable computingrdquo in Ambient Intelligencevol 2875 of Lecture Notes in Computer Science pp 220ndash232Springer Berlin Germany 2003

[16] L Gao A K Bourke and J Nelson ldquoSensor positioning foractivity recognition using multiple accelerometer-based sen-sorsrdquo in Proceedings of the 21st European Symposium on Arti-ficial Neural Networks Computational Intelligence and MachineLearning pp 425ndash430 April 2013

[17] L Gao A K Bourke and J Nelson ldquoEvaluation of accelerome-ter based multi-sensor versus single-sensor activity recognitionsystemsrdquo Medical Engineering and Physics vol 36 no 6 pp779ndash785 2014

[18] S Liu R X Gao D John J W Staudenmayer and P S Freed-son ldquoMultisensor data fusion for physical activity assessmentrdquoIEEE Transactions on Biomedical Engineering vol 59 no 3 pp687ndash696 2012

[19] J Wan M J OrsquoGrady and G M P OrsquoHare ldquoDynamic sensorevent segmentation for real-time activity recognition in a smarthome contextrdquo Personal amp Ubiquitous Computing vol 19 no 2pp 287ndash301 2015

[20] Y Zhan and T Kuroda ldquoWearable sensor-based human activityrecognition from environmental background soundsrdquo Journalof Ambient Intelligence amp Humanized Computing vol 5 no 1pp 77ndash89 2014

[21] httpenwikipediaorgwikiPhysical exercise[22] httpwwwnhlbinihgovhealthhealth-topicstopicsphys[23] httpenwikipediaorgwikiActivities of daily living[24] O D Lara and M A Labrador ldquoA survey on human activity

recognition using wearable sensorsrdquo IEEE CommunicationsSurveys amp Tutorials vol 15 no 3 pp 1192ndash1209 2013

[25] S Arora D Bhattacharjee M Nasipuri L Malik M Kunduand D K Basu ldquoPerformance comparison of SVM and ANNfor handwritten devnagari character recognitionrdquo InternationalJournal of Computer Science Issues vol 18 pp 63ndash72 2010

Journal of Electrical and Computer Engineering 9

[26] J Ren ldquoANN vs SVM which one performs better in classifi-cation of MCCs in mammogram imagingrdquo Knowledge-BasedSystems vol 26 pp 144ndash153 2012

[27] J S Raikwal and K Saxena ldquoPerformance evaluation of SVMand k-nearest neighbor algorithm over medical data setrdquoInternational Journal of Computer Applications vol 50 no 14pp 12ndash24 2012

[28] M Eastwood and B Gabrys ldquoA non-sequential representationof sequential data for churn predictionrdquo in Knowledge-Basedand Intelligent Information and Engineering Systems pp 209ndash218 Springer Berlin Germany 2009

[29] Y Nam and J W Park ldquoChild activity recognition based oncooperative fusion model of a triaxial accelerometer and abarometric pressure sensorrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 2 pp 420ndash426 2013

[30] J A Nasiri N MoghadamCharkari and K Mozafari ldquoEnergy-based model of least squares twin Support Vector Machines forhuman action recognitionrdquo Signal Processing vol 104 pp 248ndash257 2014

[31] K Altun B Barshan and O Tuncel ldquoComparative studyon classifying human activities with miniature inertial andmagnetic sensorsrdquo Pattern Recognition vol 43 no 10 pp 3605ndash3620 2010

[32] R Wang S Kwong D Chen and J Cao ldquoA vector-valued sup-port vectormachinemodel formulticlass problemrdquo InformationSciences vol 235 pp 174ndash194 2013

[33] N Zhang and C Williams ldquoWater quantity prediction usingleast squares support vector machines (LSSVM) methodrdquoJournal of Systemics Cybernetics and Informatics vol 2 no 4pp 53ndash58 2014

[34] K D Brabanter and P Karsmakers ldquoLS-SVMlab Toolbox UserrsquosGuiderdquo 2011 httpwwwesatkuleuvenbesistalssvmlabdown-loadstutorialv1 8pdf

[35] httpwwwmotionnodecom[36] M Zhang and A A Sawchuk ldquoA feature selection-based frame-

work for human activity recognition using wearable multi-modal sensorsrdquo inProceedings of the International Conference onBody Area Networks (BodyNets rsquo11) Beijing China November2011

[37] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearable sen-sorsrdquo in Proceedings of the ACM International Conference onUbiquitous Computing (UbiComp rsquo12) International Workshopon Situation Activity and Goal Awareness Pittsburgh Pa USASeptember 2012

[38] httpsipiusceduHAD[39] O D Incel M Kose and C Ersoy ldquoA review and taxonomy of

activity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013

[40] Y Liang X Zhou Z Yu and B Guo ldquoEnergy-efficient motionrelated activity recognition on mobile devices for pervasivehealthcarerdquoMobile Networks and Applications vol 19 no 3 pp303ndash317 2014

[41] R Cross ldquoStanding walking running and jumping on a forceplaterdquo American Journal of Physics vol 67 no 4 pp 304ndash3091999

[42] A L Hof J P Van Zandwijk and M F Bobbert ldquoMechanics ofhuman triceps surae muscle in walking running and jumpingrdquoActa Physiologica Scandinavica vol 174 no 1 pp 17ndash30 2002

[43] G Cola A Vecchio and M Avvenuti ldquoImproving the per-formance of fall detection systems through walk recognitionrdquo

Journal of Ambient Intelligence amp Humanized Computing vol 5no 6 pp 843ndash855 2014

[44] H-I Wu B-L Li T A Springer and W H Neill ldquoModellinganimal movement as a persistent random walk in two dimen-sions expected magnitude of net displacementrdquo EcologicalModelling vol 132 no 1-2 pp 115ndash124 2000

[45] C Li M Lin L T Yang and C Ding ldquoIntegrating the enrichedfeaturewithmachine learning algorithms for humanmovementand fall detectionrdquo Journal of Supercomputing vol 67 no 3 pp854ndash865 2014

[46] N Zhang C Williams E Ososanya and W MahmoudldquoStreamflow Prediction Based on Least Squares Support VectorMachinesrdquo 2013 httpwwwaseeorgdocumentssectionsmid-dle-atlanticfall-201311-ASEE2013 Final20Zhangpdf

[47] D Rodriguez-Martin A Sama C Perez-Lopez A Catala JCabestany and A Rodriguez-Molinero ldquoSVM-based postureidentificationwith a single waist-located triaxial accelerometerrdquoExpert Systems with Applications vol 40 no 18 pp 7203ndash72112013

[48] J P Varkey D Pompili and T AWalls ldquoHumanmotion recog-nition using a wireless sensor-based wearable systemrdquo Personalamp Ubiquitous Computing vol 16 no 7 pp 897ndash910 2012

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

4 Journal of Electrical and Computer Engineering

7 8

9

RU JU

WF

WL WR

WU WD

SI ST

SL

WF WL WR WU WDRU JU ST SI SL

1

32

4 5 6

(a) Two-class classifier

1

2 4

5 6

3

RU JU

WF WL WR WU WD

SI ST SL

WF WL WR WU WDRU JU ST SI SL

(b) Three-class classifier

1

2 3

4 5

Activity setsClassifierActivity

RU JU

WF WL WR WU WD

SI ST SL

WF WL WR WU WDRU JU ST SI SL

(c) Four-class classifier

1

2 3

Activity setsClassifierActivity

RU JU

WF WL WR WU WD SI ST SL

WF WL WR WU WDRU JU ST SI SL

(d) Five-class classifier

Figure 1 Structure of the hierarchical classification framework

119881mag = [119881mag (1) 119881mag (2) ]

where 119881mag (119894) =1119879

119879

sum

119905=(119894minus1)119879+1[10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816minus 119872mag (119894)]

2

119864mag = [119864mag (1) 119864mag (2) ]

where 119864mag (119894) = minus

119879

sum

119905=(119894minus1)119879+1[10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816

2log2(10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816

2)]

119872ang = [119872ang (1) 119872ang (2) ]

where 119872ang (119894) =1

119879 minus 1

119879

sum

119905=(119894minus1)119879+2120579 (119905)

119881ang = [119881ang (1) 119881ang (2) ]

where 119881ang (119894) =1

119879 minus 1

119879

sum

119905=(119894minus1)119879+2[120579 (119905) minus 119872ang (119894)]

2

119864ang = [119864ang (1) 119864ang (2) ]

where 119864ang (119894) = minus

119879

sum

119905=(119894minus1)119879+2[120579 (119905)

2 log2120579 (119905)

2]

(119905 minus 1) sdot (119905) =10038161003816100381610038161003816 (119905 minus 1)10038161003816100381610038161003816

10038161003816100381610038161003816 (119905)

10038161003816100381610038161003816cos 120579 (119905)

(2)

To explore the performance and correlation among thesesix features a series of scatter plots in a 2D feature space isshown in Figure 2The horizontal and vertical axes representtwo different featuresThe points in different colors representdifferent activities In Figure 2(a) the relationship between119872mag and 119881mag is described and the running jumpingwalking and static activities are clustered In Figure 2(b)the straight line between 2D walking (forward left andright) and 3D walking (upstairs and downstairs) implies that119872ang is an available feature Figure 2(c) illustrates that the119864mag and 119872mag features successfully partition the triaxialacceleration data samples fromwalking forward walking left

Journal of Electrical and Computer Engineering 5

08 1 12 14 16 18 20

05

1

15

2

25

3

AccMean

AccV

ar

RunningJumping

Static activitiesWalking

Centroids

(a) 119872mag versus 119881mag

14 15 16 17 18 19 20

002

004

006

008

01

012

AccAngleMean

AccA

ngle

Var

Walking forward left and rightWalking upstairs downstairs

Centroids

(b) 119872ang versus 119881ang

AccM

ean

minus80 minus70 minus60 minus50 minus40 minus30 minus20 minus10 0 10095

1

105

11

115

12

125

AccEntropy

Walking forwardWalking left

Walking rightCentroids

(c) 119864mag versus119872mag

AccM

ean

minus280 minus240 minus200 minus160 minus120 minus8008

085

09

095

1

105

11

115

12

125

AccAngleEntropy

Walking upstairsWalking downstairs

Centroids

(d) 119864ang versus119872ang

AccM

ean

minus6 minus5 minus4 minus3 minus2 minus1 0 1099

1

101

102

103

104

105

106

107

AccEntropy

SittingSleeping

StandingCentroids

(e) 119864mag versus119872mag

Figure 2 Scatter plots in the 2D feature space (119879 = 50)

6 Journal of Electrical and Computer Engineering

Least Squares Support Vector Machines classifier

Mmag Vmag Emag Mang Vang Eang

Maximum Act_Labellikelihood estimation

resultrec

Act_Label1 Act_Label2 Act_Label1Activity =

Figure 3 Activity estimator for multiclassification

and walking right into three isolated clusters where eachcluster contains data samples roughly from one single activityclass Figure 2(d) demonstrates the discrimination power ofthe 119864ang and 119872ang features to differentiate walking upstairsand walking downstairs Figure 2(e) shows that the triaxialacceleration signal can be classified into standing sitting andsleeping based on the 119864mag and 119872mag features

In this study we used 119872mag 119881mag 119864mag 119872ang 119881ang and119864ang as the best features for the classifiers in each layer [45]

4 Activity Estimation for Multiclassification

We presented an activity estimator for multiclassificationto estimate the human activity from the feature data Eachactivity estimator for the multiclassification included oneLS-SVM classifier and a maximum Act Label frequencyestimator (Figure 3)

We used the LS-SVM [34] method to cluster the fea-ture data After loading the testing data into Matlab webuilt an activity-recognizing model from the data After theparameters of the model were calculated we estimated theactivity by inputting some test feature data [46]The functiontrainlssvm() was used to train the support features of an LS-SVM for classification and the function simlssvm() was usedto evaluate the LS-SVM for some test feature data

Because 119872mag 119881mag 119864mag 119872ang 119881ang and 119864ang have(119899119879) elements the LS-SVM for the multiclassifier outputsan activity set which includes 119899119879 elements of Act LabelThe activity set may have different Act Labels and we mustestimate the Act Label maximum likelihood in this activityset We used the Naive Bayes algorithm to compute allAct Label likelihoods and obtained the human activity usingthe maximum Act Label likelihood The following describedhow to mathematically compute the maximum Act Labellikelihood

119860119888119905119894V119894119905119910 = [ 119886119888119905119894 ]

= LS SVM (119872mag 119881mag 119864mag119872ang 119881ang 119864ang)

Raw accelerometer data(sample)

Accelerometer datanormalization normalization

Feature extraction

Training sample

Raw accelerometer data

Accelerometer data

Feature extraction

Testing data

Estimator for classification

Result

Training stage Testing stage

Figure 4 Activity estimator working process

119901 (119860119888119905 119871119886119887119890119897119895| 119888119897119886119904119904119894119891119894119890119903

119896)

= 119873 119861119886119910119890119904 119860119888119905119894V119894119905119910 | 119888119897119886119904119904119894119891119894119890119903119896

119895 = 1 10 119896 = 1 5

119903119890119904119906119897119905119903119890119888

= max 119901 (119886119888119905 =119860119888119905 119871119886119887119890119897119895| 119888119897119886119904119904119894119891119894119890119903

119896)

119895 = 1 10 119896 = 1 5 119886119888119905 isin 119860119888119905119894V119894119905119910

(3)

Figure 4 shows the activity estimator working processwhich includes the training stage and testing stage (onlineactivity recognition) In the training stage the labeled dataof triaxial acceleration were normalized and the statisticalfeatures were extracted from those synthesized-accelerationdataThen themulticlassification estimator was used to buildthe classification model In the testing stage unlabeled rawdata of the triaxial accelerometer were processed with themethod that was used in the training stageThese synthesizeddata were classified using the multiclassification estimatorand the recognized result was obtained [47 48]

5 Experiment

Theactivity recognition dataset was the USCHumanActivityDataset The activity dataset included ten activities andcollected data from 14 subjects To capture the day-to-dayactivity variations each subject was asked to perform 5 trialsfor each activity on different days at various indoor andoutdoor locations Although the duration of each trial variesfor different activities it was sufficiently long to capture allinformation for each performed activity [37] In this sectionwe estimated the performances of the five activity classifiersin this activity recognition scheme Table 3 shows the results

Journal of Electrical and Computer Engineering 7

Table 3 Activity classifier accuracy test

Classifiers Activities recognition accuracy rate () Classifier averageaccuracy ()WF WL WR WU WD RU JU ST SI SL

Classifier 1 983 971 971 100 982Classifier 2 981 991 mdash mdash mdash mdash mdash 986Classifier 3 986 971 957 mdash mdash mdash mdash mdash mdash mdash 971Classifier 4 mdash mdash mdash 971 971 mdash mdash mdash mdash mdash 971Classifier 5 mdash mdash mdash mdash mdash mdash mdash 986 971 986 981

Table 4 Confusion matrix for average recognition accuracy for allactivities

InputAccuracy rate () 956

OutputWF WL WR WU WD RU JU ST SI SL

WF 957 12 31 0 0 0 0 0 0 0WL 28 929 43 0 0 0 0 0 0 0WR 34 52 914 0 0 0 0 0 0 0WU 0 0 0 929 71 0 0 0 0 0WD 0 0 0 57 943 0 0 0 0 0RU 08 0 0 0 0 971 21 0 0 0JU 0 0 0 0 0 29 971 0 0 0ST 0 0 0 0 0 0 0 986 08 06SI 0 0 0 0 0 0 0 21 971 08SL 0 0 0 0 0 0 0 04 10 986

of five activity recognition classifiersThese activity classifiershad over 95 accuracy [24] and were acceptable

The results of these folds are summarized in Table 4The average recognition accuracy of 956 indicates that ourproposed human activity recognition scheme can achievehigh recognition rates for a specific subject Because 2Dwalking and 3Dwalking are similar the recognition accuracyof the five walking activities is low We will attempt to obtainhigher recognition accuracy using an adequate amount oftraining data in future research

We compared the accuracy rate and running time forcommon multiclassification methods All algorithms wererun on a computer with CPU i7-2670QM 22G 8G ram andMatlab 2013a The LS-SVM performed notably well in thetestsThe average running time for the hierarchical classifica-tion framework with the LS-SVM recognizing activities was0021 seconds whichwas less than theANN (Artificial NeuralNetwork) DT (Decision Tree) and 119870NN (119896-Nearest Neigh-bor) algorithmsWe performed the ANNDT and119870NNclas-sifier tests with the built-in functions of MatlabThe LS-SVMmethod was also better than ANN DT and 119870NN in termsof the average recognition accuracy rate for the ten activitiesTable 5 shows the results

6 Conclusion and Future Work

This paper aims to provide an accurate and robust humanactivity recognition scheme The scheme used triaxial

acceleration data a hierarchical recognition scheme andactivity classifiers based on the LS-SVM and the NB algo-rithm The mean variance entropy of magnitude and angleof triaxial acceleration data were used as the features ofthe activity classifiers The scheme effectively recognizeda typical set of daily physical activities with an averageaccuracy of 956 It could distinguish walking (forward leftright upstairs and downstairs) running jumping standingsitting and sleeping activities using only a single triaxialaccelerometer The experimental results of the hierarchicalrecognition scheme show significant potential in its abilityto accurately differentiate activities using triaxial accelerationdata Although the scheme remains to be tested with USC-HAD datasets the core of this scheme is independent of thefeatures of other activity datasets therefore it is applicable toany dataset

The novelty of the proposed human activity recognitionscheme is the introduction of the LS-SVMmethod as the clas-sifier algorithm The LS-SVM is an advanced version of thestandard SVM and there are recently relatively few studiesusing LS-SVM to recognize activities with only one triaxialaccelerometer The human activity recognition scheme withLS-SVM classifiers simplifies the construction of the hierar-chical classification framework and has a lower running timethan other commonmulticlassification algorithms Accuracyis the basic element thatmust be consideredwhen any activityrecognition system is implemented and this recognitionscheme has a high success rate for which it can recognizeten different types of activities with an average accuracy of95

The next stage of our research has two parts First thealgorithms are improved to recognize these activities and theuser will not have to worry about placing the sensors at thecorrect positions to correctly detect the activities Second anunsupervised approach for automatic activity recognition isconsidered An unsupervised learning framework of humanactivity recognition will automatically cluster a large amountof unlabeled acceleration data into discrete groups of activitywhich implies that the human activity recognition can benaturally performed

Conflict of Interests

The author declares that there is no conflict of interestsregarding the publication of this paper

8 Journal of Electrical and Computer Engineering

Table 5 Accuracy rates and running times of the classification methods

Method Accuracy rate () Average rate () Running time (s)WF WL WR WU WD JU RU ST SI SL

ANN 961 914 902 905 854 775 986 967 952 991 921 0085DT 939 946 917 912 908 849 941 957 972 944 929 0411119870NN 935 921 901 882 867 886 938 961 957 938 919 0183LS-SVM 957 929 914 929 943 971 971 986 971 986 956 0021

Acknowledgments

This work was partially supported by AppropriativeResearching Fund for Professors and Doctors GuangdongUniversity of Education under Grant 11ARF04 and Guang-dong Provincial Department of Education under Grants2013LYM 0063 and 2014GXJK161

References

[1] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014

[2] J Hernandez R Cabido A S Montemayor and J J PantrigoldquoHuman activity recognition based on kinematic featuresrdquoExpert Systems vol 31 no 4 pp 345ndash353 2014

[3] J Yin G Tian Z Feng and J Li ldquoHuman activity recognitionbased on multiple order temporal informationrdquo Computers ampElectrical Engineering vol 40 no 5 pp 1538ndash1551 2014

[4] A M Khan Y-K Lee S Y Lee and T-S Kim ldquoA triaxialaccelerometer-based physical-activity recognition via aug-mented-signal features and a hierarchical recognizerrdquo IEEETransactions on Information Technology in Biomedicine vol 14no 5 pp 1166ndash1172 2010

[5] D Trabelsi S Mohammed F Chamroukhi L Oukhellou andY Amirat ldquoAn unsupervised approach for automatic activityrecognition based on hidden markov model regressionrdquo IEEETransactions onAutomation Science and Engineering vol 10 no3 pp 829ndash835 2013

[6] W L Tang and E S Sazonov ldquoHighly accurate recognitionof human postures and activities through classification withrejectionrdquo IEEE Journal of Biomedical and Health Informaticsvol 18 no 1 pp 309ndash315 2014

[7] M-W Lee A M Khan and T-S Kim ldquoA single tri-axialaccelerometer-based real-time personal life log system capableof human activity recognition and exercise information gener-ationrdquo Personal amp Ubiquitous Computing vol 15 no 8 pp 887ndash898 2011

[8] W-Y Deng Q-H Zheng and Z-M Wang ldquoCross-personactivity recognition using reduced kernel extreme learningmachinerdquo Neural Networks vol 53 pp 1ndash7 2014

[9] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014

[10] D P Tao L Jin Y Wang and X Li ldquoRank preserving discrim-inant analysis for human behavior recognition on wireless sen-sor networksrdquo IEEE Transactions on Industrial Informatics vol10 no 1 pp 813ndash823 2014

[11] N Alshurafa W Xu J J Liu et al ldquoDesigning a robust activityrecognition framework for health and exergaming using wear-able sensorsrdquo IEEE Journal of Biomedical and Health Informat-ics vol 18 no 5 pp 1636ndash1646 2014

[12] A Chang S Mota and H Lieberman ldquoGestureNet a commonsense approach to physical activity similarityrdquo in Proceedings ofthe Conference on Electronic Visualisation and the Arts LondonUK July 2014

[13] O Banos M Damas H Pomares F Rojas B Delgado-Marquez and O Valenzuela ldquoHuman activity recognitionbased on a sensor weighting hierarchical classifierrdquo Soft Com-puting vol 17 no 2 pp 333ndash343 2013

[14] J Cheng X Chen andM Shen ldquoA framework for daily activitymonitoring and fall detection based on surface electromyogra-phy and accelerometer signalsrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 1 pp 38ndash45 2013

[15] N Kern B Schiele and A Schmidt ldquoMulti-sensor activity con-text detection for wearable computingrdquo in Ambient Intelligencevol 2875 of Lecture Notes in Computer Science pp 220ndash232Springer Berlin Germany 2003

[16] L Gao A K Bourke and J Nelson ldquoSensor positioning foractivity recognition using multiple accelerometer-based sen-sorsrdquo in Proceedings of the 21st European Symposium on Arti-ficial Neural Networks Computational Intelligence and MachineLearning pp 425ndash430 April 2013

[17] L Gao A K Bourke and J Nelson ldquoEvaluation of accelerome-ter based multi-sensor versus single-sensor activity recognitionsystemsrdquo Medical Engineering and Physics vol 36 no 6 pp779ndash785 2014

[18] S Liu R X Gao D John J W Staudenmayer and P S Freed-son ldquoMultisensor data fusion for physical activity assessmentrdquoIEEE Transactions on Biomedical Engineering vol 59 no 3 pp687ndash696 2012

[19] J Wan M J OrsquoGrady and G M P OrsquoHare ldquoDynamic sensorevent segmentation for real-time activity recognition in a smarthome contextrdquo Personal amp Ubiquitous Computing vol 19 no 2pp 287ndash301 2015

[20] Y Zhan and T Kuroda ldquoWearable sensor-based human activityrecognition from environmental background soundsrdquo Journalof Ambient Intelligence amp Humanized Computing vol 5 no 1pp 77ndash89 2014

[21] httpenwikipediaorgwikiPhysical exercise[22] httpwwwnhlbinihgovhealthhealth-topicstopicsphys[23] httpenwikipediaorgwikiActivities of daily living[24] O D Lara and M A Labrador ldquoA survey on human activity

recognition using wearable sensorsrdquo IEEE CommunicationsSurveys amp Tutorials vol 15 no 3 pp 1192ndash1209 2013

[25] S Arora D Bhattacharjee M Nasipuri L Malik M Kunduand D K Basu ldquoPerformance comparison of SVM and ANNfor handwritten devnagari character recognitionrdquo InternationalJournal of Computer Science Issues vol 18 pp 63ndash72 2010

Journal of Electrical and Computer Engineering 9

[26] J Ren ldquoANN vs SVM which one performs better in classifi-cation of MCCs in mammogram imagingrdquo Knowledge-BasedSystems vol 26 pp 144ndash153 2012

[27] J S Raikwal and K Saxena ldquoPerformance evaluation of SVMand k-nearest neighbor algorithm over medical data setrdquoInternational Journal of Computer Applications vol 50 no 14pp 12ndash24 2012

[28] M Eastwood and B Gabrys ldquoA non-sequential representationof sequential data for churn predictionrdquo in Knowledge-Basedand Intelligent Information and Engineering Systems pp 209ndash218 Springer Berlin Germany 2009

[29] Y Nam and J W Park ldquoChild activity recognition based oncooperative fusion model of a triaxial accelerometer and abarometric pressure sensorrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 2 pp 420ndash426 2013

[30] J A Nasiri N MoghadamCharkari and K Mozafari ldquoEnergy-based model of least squares twin Support Vector Machines forhuman action recognitionrdquo Signal Processing vol 104 pp 248ndash257 2014

[31] K Altun B Barshan and O Tuncel ldquoComparative studyon classifying human activities with miniature inertial andmagnetic sensorsrdquo Pattern Recognition vol 43 no 10 pp 3605ndash3620 2010

[32] R Wang S Kwong D Chen and J Cao ldquoA vector-valued sup-port vectormachinemodel formulticlass problemrdquo InformationSciences vol 235 pp 174ndash194 2013

[33] N Zhang and C Williams ldquoWater quantity prediction usingleast squares support vector machines (LSSVM) methodrdquoJournal of Systemics Cybernetics and Informatics vol 2 no 4pp 53ndash58 2014

[34] K D Brabanter and P Karsmakers ldquoLS-SVMlab Toolbox UserrsquosGuiderdquo 2011 httpwwwesatkuleuvenbesistalssvmlabdown-loadstutorialv1 8pdf

[35] httpwwwmotionnodecom[36] M Zhang and A A Sawchuk ldquoA feature selection-based frame-

work for human activity recognition using wearable multi-modal sensorsrdquo inProceedings of the International Conference onBody Area Networks (BodyNets rsquo11) Beijing China November2011

[37] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearable sen-sorsrdquo in Proceedings of the ACM International Conference onUbiquitous Computing (UbiComp rsquo12) International Workshopon Situation Activity and Goal Awareness Pittsburgh Pa USASeptember 2012

[38] httpsipiusceduHAD[39] O D Incel M Kose and C Ersoy ldquoA review and taxonomy of

activity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013

[40] Y Liang X Zhou Z Yu and B Guo ldquoEnergy-efficient motionrelated activity recognition on mobile devices for pervasivehealthcarerdquoMobile Networks and Applications vol 19 no 3 pp303ndash317 2014

[41] R Cross ldquoStanding walking running and jumping on a forceplaterdquo American Journal of Physics vol 67 no 4 pp 304ndash3091999

[42] A L Hof J P Van Zandwijk and M F Bobbert ldquoMechanics ofhuman triceps surae muscle in walking running and jumpingrdquoActa Physiologica Scandinavica vol 174 no 1 pp 17ndash30 2002

[43] G Cola A Vecchio and M Avvenuti ldquoImproving the per-formance of fall detection systems through walk recognitionrdquo

Journal of Ambient Intelligence amp Humanized Computing vol 5no 6 pp 843ndash855 2014

[44] H-I Wu B-L Li T A Springer and W H Neill ldquoModellinganimal movement as a persistent random walk in two dimen-sions expected magnitude of net displacementrdquo EcologicalModelling vol 132 no 1-2 pp 115ndash124 2000

[45] C Li M Lin L T Yang and C Ding ldquoIntegrating the enrichedfeaturewithmachine learning algorithms for humanmovementand fall detectionrdquo Journal of Supercomputing vol 67 no 3 pp854ndash865 2014

[46] N Zhang C Williams E Ososanya and W MahmoudldquoStreamflow Prediction Based on Least Squares Support VectorMachinesrdquo 2013 httpwwwaseeorgdocumentssectionsmid-dle-atlanticfall-201311-ASEE2013 Final20Zhangpdf

[47] D Rodriguez-Martin A Sama C Perez-Lopez A Catala JCabestany and A Rodriguez-Molinero ldquoSVM-based postureidentificationwith a single waist-located triaxial accelerometerrdquoExpert Systems with Applications vol 40 no 18 pp 7203ndash72112013

[48] J P Varkey D Pompili and T AWalls ldquoHumanmotion recog-nition using a wireless sensor-based wearable systemrdquo Personalamp Ubiquitous Computing vol 16 no 7 pp 897ndash910 2012

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Electrical and Computer Engineering 5

08 1 12 14 16 18 20

05

1

15

2

25

3

AccMean

AccV

ar

RunningJumping

Static activitiesWalking

Centroids

(a) 119872mag versus 119881mag

14 15 16 17 18 19 20

002

004

006

008

01

012

AccAngleMean

AccA

ngle

Var

Walking forward left and rightWalking upstairs downstairs

Centroids

(b) 119872ang versus 119881ang

AccM

ean

minus80 minus70 minus60 minus50 minus40 minus30 minus20 minus10 0 10095

1

105

11

115

12

125

AccEntropy

Walking forwardWalking left

Walking rightCentroids

(c) 119864mag versus119872mag

AccM

ean

minus280 minus240 minus200 minus160 minus120 minus8008

085

09

095

1

105

11

115

12

125

AccAngleEntropy

Walking upstairsWalking downstairs

Centroids

(d) 119864ang versus119872ang

AccM

ean

minus6 minus5 minus4 minus3 minus2 minus1 0 1099

1

101

102

103

104

105

106

107

AccEntropy

SittingSleeping

StandingCentroids

(e) 119864mag versus119872mag

Figure 2 Scatter plots in the 2D feature space (119879 = 50)

6 Journal of Electrical and Computer Engineering

Least Squares Support Vector Machines classifier

Mmag Vmag Emag Mang Vang Eang

Maximum Act_Labellikelihood estimation

resultrec

Act_Label1 Act_Label2 Act_Label1Activity =

Figure 3 Activity estimator for multiclassification

and walking right into three isolated clusters where eachcluster contains data samples roughly from one single activityclass Figure 2(d) demonstrates the discrimination power ofthe 119864ang and 119872ang features to differentiate walking upstairsand walking downstairs Figure 2(e) shows that the triaxialacceleration signal can be classified into standing sitting andsleeping based on the 119864mag and 119872mag features

In this study we used 119872mag 119881mag 119864mag 119872ang 119881ang and119864ang as the best features for the classifiers in each layer [45]

4 Activity Estimation for Multiclassification

We presented an activity estimator for multiclassificationto estimate the human activity from the feature data Eachactivity estimator for the multiclassification included oneLS-SVM classifier and a maximum Act Label frequencyestimator (Figure 3)

We used the LS-SVM [34] method to cluster the fea-ture data After loading the testing data into Matlab webuilt an activity-recognizing model from the data After theparameters of the model were calculated we estimated theactivity by inputting some test feature data [46]The functiontrainlssvm() was used to train the support features of an LS-SVM for classification and the function simlssvm() was usedto evaluate the LS-SVM for some test feature data

Because 119872mag 119881mag 119864mag 119872ang 119881ang and 119864ang have(119899119879) elements the LS-SVM for the multiclassifier outputsan activity set which includes 119899119879 elements of Act LabelThe activity set may have different Act Labels and we mustestimate the Act Label maximum likelihood in this activityset We used the Naive Bayes algorithm to compute allAct Label likelihoods and obtained the human activity usingthe maximum Act Label likelihood The following describedhow to mathematically compute the maximum Act Labellikelihood

119860119888119905119894V119894119905119910 = [ 119886119888119905119894 ]

= LS SVM (119872mag 119881mag 119864mag119872ang 119881ang 119864ang)

Raw accelerometer data(sample)

Accelerometer datanormalization normalization

Feature extraction

Training sample

Raw accelerometer data

Accelerometer data

Feature extraction

Testing data

Estimator for classification

Result

Training stage Testing stage

Figure 4 Activity estimator working process

119901 (119860119888119905 119871119886119887119890119897119895| 119888119897119886119904119904119894119891119894119890119903

119896)

= 119873 119861119886119910119890119904 119860119888119905119894V119894119905119910 | 119888119897119886119904119904119894119891119894119890119903119896

119895 = 1 10 119896 = 1 5

119903119890119904119906119897119905119903119890119888

= max 119901 (119886119888119905 =119860119888119905 119871119886119887119890119897119895| 119888119897119886119904119904119894119891119894119890119903

119896)

119895 = 1 10 119896 = 1 5 119886119888119905 isin 119860119888119905119894V119894119905119910

(3)

Figure 4 shows the activity estimator working processwhich includes the training stage and testing stage (onlineactivity recognition) In the training stage the labeled dataof triaxial acceleration were normalized and the statisticalfeatures were extracted from those synthesized-accelerationdataThen themulticlassification estimator was used to buildthe classification model In the testing stage unlabeled rawdata of the triaxial accelerometer were processed with themethod that was used in the training stageThese synthesizeddata were classified using the multiclassification estimatorand the recognized result was obtained [47 48]

5 Experiment

Theactivity recognition dataset was the USCHumanActivityDataset The activity dataset included ten activities andcollected data from 14 subjects To capture the day-to-dayactivity variations each subject was asked to perform 5 trialsfor each activity on different days at various indoor andoutdoor locations Although the duration of each trial variesfor different activities it was sufficiently long to capture allinformation for each performed activity [37] In this sectionwe estimated the performances of the five activity classifiersin this activity recognition scheme Table 3 shows the results

Journal of Electrical and Computer Engineering 7

Table 3 Activity classifier accuracy test

Classifiers Activities recognition accuracy rate () Classifier averageaccuracy ()WF WL WR WU WD RU JU ST SI SL

Classifier 1 983 971 971 100 982Classifier 2 981 991 mdash mdash mdash mdash mdash 986Classifier 3 986 971 957 mdash mdash mdash mdash mdash mdash mdash 971Classifier 4 mdash mdash mdash 971 971 mdash mdash mdash mdash mdash 971Classifier 5 mdash mdash mdash mdash mdash mdash mdash 986 971 986 981

Table 4 Confusion matrix for average recognition accuracy for allactivities

InputAccuracy rate () 956

OutputWF WL WR WU WD RU JU ST SI SL

WF 957 12 31 0 0 0 0 0 0 0WL 28 929 43 0 0 0 0 0 0 0WR 34 52 914 0 0 0 0 0 0 0WU 0 0 0 929 71 0 0 0 0 0WD 0 0 0 57 943 0 0 0 0 0RU 08 0 0 0 0 971 21 0 0 0JU 0 0 0 0 0 29 971 0 0 0ST 0 0 0 0 0 0 0 986 08 06SI 0 0 0 0 0 0 0 21 971 08SL 0 0 0 0 0 0 0 04 10 986

of five activity recognition classifiersThese activity classifiershad over 95 accuracy [24] and were acceptable

The results of these folds are summarized in Table 4The average recognition accuracy of 956 indicates that ourproposed human activity recognition scheme can achievehigh recognition rates for a specific subject Because 2Dwalking and 3Dwalking are similar the recognition accuracyof the five walking activities is low We will attempt to obtainhigher recognition accuracy using an adequate amount oftraining data in future research

We compared the accuracy rate and running time forcommon multiclassification methods All algorithms wererun on a computer with CPU i7-2670QM 22G 8G ram andMatlab 2013a The LS-SVM performed notably well in thetestsThe average running time for the hierarchical classifica-tion framework with the LS-SVM recognizing activities was0021 seconds whichwas less than theANN (Artificial NeuralNetwork) DT (Decision Tree) and 119870NN (119896-Nearest Neigh-bor) algorithmsWe performed the ANNDT and119870NNclas-sifier tests with the built-in functions of MatlabThe LS-SVMmethod was also better than ANN DT and 119870NN in termsof the average recognition accuracy rate for the ten activitiesTable 5 shows the results

6 Conclusion and Future Work

This paper aims to provide an accurate and robust humanactivity recognition scheme The scheme used triaxial

acceleration data a hierarchical recognition scheme andactivity classifiers based on the LS-SVM and the NB algo-rithm The mean variance entropy of magnitude and angleof triaxial acceleration data were used as the features ofthe activity classifiers The scheme effectively recognizeda typical set of daily physical activities with an averageaccuracy of 956 It could distinguish walking (forward leftright upstairs and downstairs) running jumping standingsitting and sleeping activities using only a single triaxialaccelerometer The experimental results of the hierarchicalrecognition scheme show significant potential in its abilityto accurately differentiate activities using triaxial accelerationdata Although the scheme remains to be tested with USC-HAD datasets the core of this scheme is independent of thefeatures of other activity datasets therefore it is applicable toany dataset

The novelty of the proposed human activity recognitionscheme is the introduction of the LS-SVMmethod as the clas-sifier algorithm The LS-SVM is an advanced version of thestandard SVM and there are recently relatively few studiesusing LS-SVM to recognize activities with only one triaxialaccelerometer The human activity recognition scheme withLS-SVM classifiers simplifies the construction of the hierar-chical classification framework and has a lower running timethan other commonmulticlassification algorithms Accuracyis the basic element thatmust be consideredwhen any activityrecognition system is implemented and this recognitionscheme has a high success rate for which it can recognizeten different types of activities with an average accuracy of95

The next stage of our research has two parts First thealgorithms are improved to recognize these activities and theuser will not have to worry about placing the sensors at thecorrect positions to correctly detect the activities Second anunsupervised approach for automatic activity recognition isconsidered An unsupervised learning framework of humanactivity recognition will automatically cluster a large amountof unlabeled acceleration data into discrete groups of activitywhich implies that the human activity recognition can benaturally performed

Conflict of Interests

The author declares that there is no conflict of interestsregarding the publication of this paper

8 Journal of Electrical and Computer Engineering

Table 5 Accuracy rates and running times of the classification methods

Method Accuracy rate () Average rate () Running time (s)WF WL WR WU WD JU RU ST SI SL

ANN 961 914 902 905 854 775 986 967 952 991 921 0085DT 939 946 917 912 908 849 941 957 972 944 929 0411119870NN 935 921 901 882 867 886 938 961 957 938 919 0183LS-SVM 957 929 914 929 943 971 971 986 971 986 956 0021

Acknowledgments

This work was partially supported by AppropriativeResearching Fund for Professors and Doctors GuangdongUniversity of Education under Grant 11ARF04 and Guang-dong Provincial Department of Education under Grants2013LYM 0063 and 2014GXJK161

References

[1] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014

[2] J Hernandez R Cabido A S Montemayor and J J PantrigoldquoHuman activity recognition based on kinematic featuresrdquoExpert Systems vol 31 no 4 pp 345ndash353 2014

[3] J Yin G Tian Z Feng and J Li ldquoHuman activity recognitionbased on multiple order temporal informationrdquo Computers ampElectrical Engineering vol 40 no 5 pp 1538ndash1551 2014

[4] A M Khan Y-K Lee S Y Lee and T-S Kim ldquoA triaxialaccelerometer-based physical-activity recognition via aug-mented-signal features and a hierarchical recognizerrdquo IEEETransactions on Information Technology in Biomedicine vol 14no 5 pp 1166ndash1172 2010

[5] D Trabelsi S Mohammed F Chamroukhi L Oukhellou andY Amirat ldquoAn unsupervised approach for automatic activityrecognition based on hidden markov model regressionrdquo IEEETransactions onAutomation Science and Engineering vol 10 no3 pp 829ndash835 2013

[6] W L Tang and E S Sazonov ldquoHighly accurate recognitionof human postures and activities through classification withrejectionrdquo IEEE Journal of Biomedical and Health Informaticsvol 18 no 1 pp 309ndash315 2014

[7] M-W Lee A M Khan and T-S Kim ldquoA single tri-axialaccelerometer-based real-time personal life log system capableof human activity recognition and exercise information gener-ationrdquo Personal amp Ubiquitous Computing vol 15 no 8 pp 887ndash898 2011

[8] W-Y Deng Q-H Zheng and Z-M Wang ldquoCross-personactivity recognition using reduced kernel extreme learningmachinerdquo Neural Networks vol 53 pp 1ndash7 2014

[9] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014

[10] D P Tao L Jin Y Wang and X Li ldquoRank preserving discrim-inant analysis for human behavior recognition on wireless sen-sor networksrdquo IEEE Transactions on Industrial Informatics vol10 no 1 pp 813ndash823 2014

[11] N Alshurafa W Xu J J Liu et al ldquoDesigning a robust activityrecognition framework for health and exergaming using wear-able sensorsrdquo IEEE Journal of Biomedical and Health Informat-ics vol 18 no 5 pp 1636ndash1646 2014

[12] A Chang S Mota and H Lieberman ldquoGestureNet a commonsense approach to physical activity similarityrdquo in Proceedings ofthe Conference on Electronic Visualisation and the Arts LondonUK July 2014

[13] O Banos M Damas H Pomares F Rojas B Delgado-Marquez and O Valenzuela ldquoHuman activity recognitionbased on a sensor weighting hierarchical classifierrdquo Soft Com-puting vol 17 no 2 pp 333ndash343 2013

[14] J Cheng X Chen andM Shen ldquoA framework for daily activitymonitoring and fall detection based on surface electromyogra-phy and accelerometer signalsrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 1 pp 38ndash45 2013

[15] N Kern B Schiele and A Schmidt ldquoMulti-sensor activity con-text detection for wearable computingrdquo in Ambient Intelligencevol 2875 of Lecture Notes in Computer Science pp 220ndash232Springer Berlin Germany 2003

[16] L Gao A K Bourke and J Nelson ldquoSensor positioning foractivity recognition using multiple accelerometer-based sen-sorsrdquo in Proceedings of the 21st European Symposium on Arti-ficial Neural Networks Computational Intelligence and MachineLearning pp 425ndash430 April 2013

[17] L Gao A K Bourke and J Nelson ldquoEvaluation of accelerome-ter based multi-sensor versus single-sensor activity recognitionsystemsrdquo Medical Engineering and Physics vol 36 no 6 pp779ndash785 2014

[18] S Liu R X Gao D John J W Staudenmayer and P S Freed-son ldquoMultisensor data fusion for physical activity assessmentrdquoIEEE Transactions on Biomedical Engineering vol 59 no 3 pp687ndash696 2012

[19] J Wan M J OrsquoGrady and G M P OrsquoHare ldquoDynamic sensorevent segmentation for real-time activity recognition in a smarthome contextrdquo Personal amp Ubiquitous Computing vol 19 no 2pp 287ndash301 2015

[20] Y Zhan and T Kuroda ldquoWearable sensor-based human activityrecognition from environmental background soundsrdquo Journalof Ambient Intelligence amp Humanized Computing vol 5 no 1pp 77ndash89 2014

[21] httpenwikipediaorgwikiPhysical exercise[22] httpwwwnhlbinihgovhealthhealth-topicstopicsphys[23] httpenwikipediaorgwikiActivities of daily living[24] O D Lara and M A Labrador ldquoA survey on human activity

recognition using wearable sensorsrdquo IEEE CommunicationsSurveys amp Tutorials vol 15 no 3 pp 1192ndash1209 2013

[25] S Arora D Bhattacharjee M Nasipuri L Malik M Kunduand D K Basu ldquoPerformance comparison of SVM and ANNfor handwritten devnagari character recognitionrdquo InternationalJournal of Computer Science Issues vol 18 pp 63ndash72 2010

Journal of Electrical and Computer Engineering 9

[26] J Ren ldquoANN vs SVM which one performs better in classifi-cation of MCCs in mammogram imagingrdquo Knowledge-BasedSystems vol 26 pp 144ndash153 2012

[27] J S Raikwal and K Saxena ldquoPerformance evaluation of SVMand k-nearest neighbor algorithm over medical data setrdquoInternational Journal of Computer Applications vol 50 no 14pp 12ndash24 2012

[28] M Eastwood and B Gabrys ldquoA non-sequential representationof sequential data for churn predictionrdquo in Knowledge-Basedand Intelligent Information and Engineering Systems pp 209ndash218 Springer Berlin Germany 2009

[29] Y Nam and J W Park ldquoChild activity recognition based oncooperative fusion model of a triaxial accelerometer and abarometric pressure sensorrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 2 pp 420ndash426 2013

[30] J A Nasiri N MoghadamCharkari and K Mozafari ldquoEnergy-based model of least squares twin Support Vector Machines forhuman action recognitionrdquo Signal Processing vol 104 pp 248ndash257 2014

[31] K Altun B Barshan and O Tuncel ldquoComparative studyon classifying human activities with miniature inertial andmagnetic sensorsrdquo Pattern Recognition vol 43 no 10 pp 3605ndash3620 2010

[32] R Wang S Kwong D Chen and J Cao ldquoA vector-valued sup-port vectormachinemodel formulticlass problemrdquo InformationSciences vol 235 pp 174ndash194 2013

[33] N Zhang and C Williams ldquoWater quantity prediction usingleast squares support vector machines (LSSVM) methodrdquoJournal of Systemics Cybernetics and Informatics vol 2 no 4pp 53ndash58 2014

[34] K D Brabanter and P Karsmakers ldquoLS-SVMlab Toolbox UserrsquosGuiderdquo 2011 httpwwwesatkuleuvenbesistalssvmlabdown-loadstutorialv1 8pdf

[35] httpwwwmotionnodecom[36] M Zhang and A A Sawchuk ldquoA feature selection-based frame-

work for human activity recognition using wearable multi-modal sensorsrdquo inProceedings of the International Conference onBody Area Networks (BodyNets rsquo11) Beijing China November2011

[37] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearable sen-sorsrdquo in Proceedings of the ACM International Conference onUbiquitous Computing (UbiComp rsquo12) International Workshopon Situation Activity and Goal Awareness Pittsburgh Pa USASeptember 2012

[38] httpsipiusceduHAD[39] O D Incel M Kose and C Ersoy ldquoA review and taxonomy of

activity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013

[40] Y Liang X Zhou Z Yu and B Guo ldquoEnergy-efficient motionrelated activity recognition on mobile devices for pervasivehealthcarerdquoMobile Networks and Applications vol 19 no 3 pp303ndash317 2014

[41] R Cross ldquoStanding walking running and jumping on a forceplaterdquo American Journal of Physics vol 67 no 4 pp 304ndash3091999

[42] A L Hof J P Van Zandwijk and M F Bobbert ldquoMechanics ofhuman triceps surae muscle in walking running and jumpingrdquoActa Physiologica Scandinavica vol 174 no 1 pp 17ndash30 2002

[43] G Cola A Vecchio and M Avvenuti ldquoImproving the per-formance of fall detection systems through walk recognitionrdquo

Journal of Ambient Intelligence amp Humanized Computing vol 5no 6 pp 843ndash855 2014

[44] H-I Wu B-L Li T A Springer and W H Neill ldquoModellinganimal movement as a persistent random walk in two dimen-sions expected magnitude of net displacementrdquo EcologicalModelling vol 132 no 1-2 pp 115ndash124 2000

[45] C Li M Lin L T Yang and C Ding ldquoIntegrating the enrichedfeaturewithmachine learning algorithms for humanmovementand fall detectionrdquo Journal of Supercomputing vol 67 no 3 pp854ndash865 2014

[46] N Zhang C Williams E Ososanya and W MahmoudldquoStreamflow Prediction Based on Least Squares Support VectorMachinesrdquo 2013 httpwwwaseeorgdocumentssectionsmid-dle-atlanticfall-201311-ASEE2013 Final20Zhangpdf

[47] D Rodriguez-Martin A Sama C Perez-Lopez A Catala JCabestany and A Rodriguez-Molinero ldquoSVM-based postureidentificationwith a single waist-located triaxial accelerometerrdquoExpert Systems with Applications vol 40 no 18 pp 7203ndash72112013

[48] J P Varkey D Pompili and T AWalls ldquoHumanmotion recog-nition using a wireless sensor-based wearable systemrdquo Personalamp Ubiquitous Computing vol 16 no 7 pp 897ndash910 2012

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

6 Journal of Electrical and Computer Engineering

Least Squares Support Vector Machines classifier

Mmag Vmag Emag Mang Vang Eang

Maximum Act_Labellikelihood estimation

resultrec

Act_Label1 Act_Label2 Act_Label1Activity =

Figure 3 Activity estimator for multiclassification

and walking right into three isolated clusters where eachcluster contains data samples roughly from one single activityclass Figure 2(d) demonstrates the discrimination power ofthe 119864ang and 119872ang features to differentiate walking upstairsand walking downstairs Figure 2(e) shows that the triaxialacceleration signal can be classified into standing sitting andsleeping based on the 119864mag and 119872mag features

In this study we used 119872mag 119881mag 119864mag 119872ang 119881ang and119864ang as the best features for the classifiers in each layer [45]

4 Activity Estimation for Multiclassification

We presented an activity estimator for multiclassificationto estimate the human activity from the feature data Eachactivity estimator for the multiclassification included oneLS-SVM classifier and a maximum Act Label frequencyestimator (Figure 3)

We used the LS-SVM [34] method to cluster the fea-ture data After loading the testing data into Matlab webuilt an activity-recognizing model from the data After theparameters of the model were calculated we estimated theactivity by inputting some test feature data [46]The functiontrainlssvm() was used to train the support features of an LS-SVM for classification and the function simlssvm() was usedto evaluate the LS-SVM for some test feature data

Because 119872mag 119881mag 119864mag 119872ang 119881ang and 119864ang have(119899119879) elements the LS-SVM for the multiclassifier outputsan activity set which includes 119899119879 elements of Act LabelThe activity set may have different Act Labels and we mustestimate the Act Label maximum likelihood in this activityset We used the Naive Bayes algorithm to compute allAct Label likelihoods and obtained the human activity usingthe maximum Act Label likelihood The following describedhow to mathematically compute the maximum Act Labellikelihood

119860119888119905119894V119894119905119910 = [ 119886119888119905119894 ]

= LS SVM (119872mag 119881mag 119864mag119872ang 119881ang 119864ang)

Raw accelerometer data(sample)

Accelerometer datanormalization normalization

Feature extraction

Training sample

Raw accelerometer data

Accelerometer data

Feature extraction

Testing data

Estimator for classification

Result

Training stage Testing stage

Figure 4 Activity estimator working process

119901 (119860119888119905 119871119886119887119890119897119895| 119888119897119886119904119904119894119891119894119890119903

119896)

= 119873 119861119886119910119890119904 119860119888119905119894V119894119905119910 | 119888119897119886119904119904119894119891119894119890119903119896

119895 = 1 10 119896 = 1 5

119903119890119904119906119897119905119903119890119888

= max 119901 (119886119888119905 =119860119888119905 119871119886119887119890119897119895| 119888119897119886119904119904119894119891119894119890119903

119896)

119895 = 1 10 119896 = 1 5 119886119888119905 isin 119860119888119905119894V119894119905119910

(3)

Figure 4 shows the activity estimator working processwhich includes the training stage and testing stage (onlineactivity recognition) In the training stage the labeled dataof triaxial acceleration were normalized and the statisticalfeatures were extracted from those synthesized-accelerationdataThen themulticlassification estimator was used to buildthe classification model In the testing stage unlabeled rawdata of the triaxial accelerometer were processed with themethod that was used in the training stageThese synthesizeddata were classified using the multiclassification estimatorand the recognized result was obtained [47 48]

5 Experiment

Theactivity recognition dataset was the USCHumanActivityDataset The activity dataset included ten activities andcollected data from 14 subjects To capture the day-to-dayactivity variations each subject was asked to perform 5 trialsfor each activity on different days at various indoor andoutdoor locations Although the duration of each trial variesfor different activities it was sufficiently long to capture allinformation for each performed activity [37] In this sectionwe estimated the performances of the five activity classifiersin this activity recognition scheme Table 3 shows the results

Journal of Electrical and Computer Engineering 7

Table 3 Activity classifier accuracy test

Classifiers Activities recognition accuracy rate () Classifier averageaccuracy ()WF WL WR WU WD RU JU ST SI SL

Classifier 1 983 971 971 100 982Classifier 2 981 991 mdash mdash mdash mdash mdash 986Classifier 3 986 971 957 mdash mdash mdash mdash mdash mdash mdash 971Classifier 4 mdash mdash mdash 971 971 mdash mdash mdash mdash mdash 971Classifier 5 mdash mdash mdash mdash mdash mdash mdash 986 971 986 981

Table 4 Confusion matrix for average recognition accuracy for allactivities

InputAccuracy rate () 956

OutputWF WL WR WU WD RU JU ST SI SL

WF 957 12 31 0 0 0 0 0 0 0WL 28 929 43 0 0 0 0 0 0 0WR 34 52 914 0 0 0 0 0 0 0WU 0 0 0 929 71 0 0 0 0 0WD 0 0 0 57 943 0 0 0 0 0RU 08 0 0 0 0 971 21 0 0 0JU 0 0 0 0 0 29 971 0 0 0ST 0 0 0 0 0 0 0 986 08 06SI 0 0 0 0 0 0 0 21 971 08SL 0 0 0 0 0 0 0 04 10 986

of five activity recognition classifiersThese activity classifiershad over 95 accuracy [24] and were acceptable

The results of these folds are summarized in Table 4The average recognition accuracy of 956 indicates that ourproposed human activity recognition scheme can achievehigh recognition rates for a specific subject Because 2Dwalking and 3Dwalking are similar the recognition accuracyof the five walking activities is low We will attempt to obtainhigher recognition accuracy using an adequate amount oftraining data in future research

We compared the accuracy rate and running time forcommon multiclassification methods All algorithms wererun on a computer with CPU i7-2670QM 22G 8G ram andMatlab 2013a The LS-SVM performed notably well in thetestsThe average running time for the hierarchical classifica-tion framework with the LS-SVM recognizing activities was0021 seconds whichwas less than theANN (Artificial NeuralNetwork) DT (Decision Tree) and 119870NN (119896-Nearest Neigh-bor) algorithmsWe performed the ANNDT and119870NNclas-sifier tests with the built-in functions of MatlabThe LS-SVMmethod was also better than ANN DT and 119870NN in termsof the average recognition accuracy rate for the ten activitiesTable 5 shows the results

6 Conclusion and Future Work

This paper aims to provide an accurate and robust humanactivity recognition scheme The scheme used triaxial

acceleration data a hierarchical recognition scheme andactivity classifiers based on the LS-SVM and the NB algo-rithm The mean variance entropy of magnitude and angleof triaxial acceleration data were used as the features ofthe activity classifiers The scheme effectively recognizeda typical set of daily physical activities with an averageaccuracy of 956 It could distinguish walking (forward leftright upstairs and downstairs) running jumping standingsitting and sleeping activities using only a single triaxialaccelerometer The experimental results of the hierarchicalrecognition scheme show significant potential in its abilityto accurately differentiate activities using triaxial accelerationdata Although the scheme remains to be tested with USC-HAD datasets the core of this scheme is independent of thefeatures of other activity datasets therefore it is applicable toany dataset

The novelty of the proposed human activity recognitionscheme is the introduction of the LS-SVMmethod as the clas-sifier algorithm The LS-SVM is an advanced version of thestandard SVM and there are recently relatively few studiesusing LS-SVM to recognize activities with only one triaxialaccelerometer The human activity recognition scheme withLS-SVM classifiers simplifies the construction of the hierar-chical classification framework and has a lower running timethan other commonmulticlassification algorithms Accuracyis the basic element thatmust be consideredwhen any activityrecognition system is implemented and this recognitionscheme has a high success rate for which it can recognizeten different types of activities with an average accuracy of95

The next stage of our research has two parts First thealgorithms are improved to recognize these activities and theuser will not have to worry about placing the sensors at thecorrect positions to correctly detect the activities Second anunsupervised approach for automatic activity recognition isconsidered An unsupervised learning framework of humanactivity recognition will automatically cluster a large amountof unlabeled acceleration data into discrete groups of activitywhich implies that the human activity recognition can benaturally performed

Conflict of Interests

The author declares that there is no conflict of interestsregarding the publication of this paper

8 Journal of Electrical and Computer Engineering

Table 5 Accuracy rates and running times of the classification methods

Method Accuracy rate () Average rate () Running time (s)WF WL WR WU WD JU RU ST SI SL

ANN 961 914 902 905 854 775 986 967 952 991 921 0085DT 939 946 917 912 908 849 941 957 972 944 929 0411119870NN 935 921 901 882 867 886 938 961 957 938 919 0183LS-SVM 957 929 914 929 943 971 971 986 971 986 956 0021

Acknowledgments

This work was partially supported by AppropriativeResearching Fund for Professors and Doctors GuangdongUniversity of Education under Grant 11ARF04 and Guang-dong Provincial Department of Education under Grants2013LYM 0063 and 2014GXJK161

References

[1] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014

[2] J Hernandez R Cabido A S Montemayor and J J PantrigoldquoHuman activity recognition based on kinematic featuresrdquoExpert Systems vol 31 no 4 pp 345ndash353 2014

[3] J Yin G Tian Z Feng and J Li ldquoHuman activity recognitionbased on multiple order temporal informationrdquo Computers ampElectrical Engineering vol 40 no 5 pp 1538ndash1551 2014

[4] A M Khan Y-K Lee S Y Lee and T-S Kim ldquoA triaxialaccelerometer-based physical-activity recognition via aug-mented-signal features and a hierarchical recognizerrdquo IEEETransactions on Information Technology in Biomedicine vol 14no 5 pp 1166ndash1172 2010

[5] D Trabelsi S Mohammed F Chamroukhi L Oukhellou andY Amirat ldquoAn unsupervised approach for automatic activityrecognition based on hidden markov model regressionrdquo IEEETransactions onAutomation Science and Engineering vol 10 no3 pp 829ndash835 2013

[6] W L Tang and E S Sazonov ldquoHighly accurate recognitionof human postures and activities through classification withrejectionrdquo IEEE Journal of Biomedical and Health Informaticsvol 18 no 1 pp 309ndash315 2014

[7] M-W Lee A M Khan and T-S Kim ldquoA single tri-axialaccelerometer-based real-time personal life log system capableof human activity recognition and exercise information gener-ationrdquo Personal amp Ubiquitous Computing vol 15 no 8 pp 887ndash898 2011

[8] W-Y Deng Q-H Zheng and Z-M Wang ldquoCross-personactivity recognition using reduced kernel extreme learningmachinerdquo Neural Networks vol 53 pp 1ndash7 2014

[9] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014

[10] D P Tao L Jin Y Wang and X Li ldquoRank preserving discrim-inant analysis for human behavior recognition on wireless sen-sor networksrdquo IEEE Transactions on Industrial Informatics vol10 no 1 pp 813ndash823 2014

[11] N Alshurafa W Xu J J Liu et al ldquoDesigning a robust activityrecognition framework for health and exergaming using wear-able sensorsrdquo IEEE Journal of Biomedical and Health Informat-ics vol 18 no 5 pp 1636ndash1646 2014

[12] A Chang S Mota and H Lieberman ldquoGestureNet a commonsense approach to physical activity similarityrdquo in Proceedings ofthe Conference on Electronic Visualisation and the Arts LondonUK July 2014

[13] O Banos M Damas H Pomares F Rojas B Delgado-Marquez and O Valenzuela ldquoHuman activity recognitionbased on a sensor weighting hierarchical classifierrdquo Soft Com-puting vol 17 no 2 pp 333ndash343 2013

[14] J Cheng X Chen andM Shen ldquoA framework for daily activitymonitoring and fall detection based on surface electromyogra-phy and accelerometer signalsrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 1 pp 38ndash45 2013

[15] N Kern B Schiele and A Schmidt ldquoMulti-sensor activity con-text detection for wearable computingrdquo in Ambient Intelligencevol 2875 of Lecture Notes in Computer Science pp 220ndash232Springer Berlin Germany 2003

[16] L Gao A K Bourke and J Nelson ldquoSensor positioning foractivity recognition using multiple accelerometer-based sen-sorsrdquo in Proceedings of the 21st European Symposium on Arti-ficial Neural Networks Computational Intelligence and MachineLearning pp 425ndash430 April 2013

[17] L Gao A K Bourke and J Nelson ldquoEvaluation of accelerome-ter based multi-sensor versus single-sensor activity recognitionsystemsrdquo Medical Engineering and Physics vol 36 no 6 pp779ndash785 2014

[18] S Liu R X Gao D John J W Staudenmayer and P S Freed-son ldquoMultisensor data fusion for physical activity assessmentrdquoIEEE Transactions on Biomedical Engineering vol 59 no 3 pp687ndash696 2012

[19] J Wan M J OrsquoGrady and G M P OrsquoHare ldquoDynamic sensorevent segmentation for real-time activity recognition in a smarthome contextrdquo Personal amp Ubiquitous Computing vol 19 no 2pp 287ndash301 2015

[20] Y Zhan and T Kuroda ldquoWearable sensor-based human activityrecognition from environmental background soundsrdquo Journalof Ambient Intelligence amp Humanized Computing vol 5 no 1pp 77ndash89 2014

[21] httpenwikipediaorgwikiPhysical exercise[22] httpwwwnhlbinihgovhealthhealth-topicstopicsphys[23] httpenwikipediaorgwikiActivities of daily living[24] O D Lara and M A Labrador ldquoA survey on human activity

recognition using wearable sensorsrdquo IEEE CommunicationsSurveys amp Tutorials vol 15 no 3 pp 1192ndash1209 2013

[25] S Arora D Bhattacharjee M Nasipuri L Malik M Kunduand D K Basu ldquoPerformance comparison of SVM and ANNfor handwritten devnagari character recognitionrdquo InternationalJournal of Computer Science Issues vol 18 pp 63ndash72 2010

Journal of Electrical and Computer Engineering 9

[26] J Ren ldquoANN vs SVM which one performs better in classifi-cation of MCCs in mammogram imagingrdquo Knowledge-BasedSystems vol 26 pp 144ndash153 2012

[27] J S Raikwal and K Saxena ldquoPerformance evaluation of SVMand k-nearest neighbor algorithm over medical data setrdquoInternational Journal of Computer Applications vol 50 no 14pp 12ndash24 2012

[28] M Eastwood and B Gabrys ldquoA non-sequential representationof sequential data for churn predictionrdquo in Knowledge-Basedand Intelligent Information and Engineering Systems pp 209ndash218 Springer Berlin Germany 2009

[29] Y Nam and J W Park ldquoChild activity recognition based oncooperative fusion model of a triaxial accelerometer and abarometric pressure sensorrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 2 pp 420ndash426 2013

[30] J A Nasiri N MoghadamCharkari and K Mozafari ldquoEnergy-based model of least squares twin Support Vector Machines forhuman action recognitionrdquo Signal Processing vol 104 pp 248ndash257 2014

[31] K Altun B Barshan and O Tuncel ldquoComparative studyon classifying human activities with miniature inertial andmagnetic sensorsrdquo Pattern Recognition vol 43 no 10 pp 3605ndash3620 2010

[32] R Wang S Kwong D Chen and J Cao ldquoA vector-valued sup-port vectormachinemodel formulticlass problemrdquo InformationSciences vol 235 pp 174ndash194 2013

[33] N Zhang and C Williams ldquoWater quantity prediction usingleast squares support vector machines (LSSVM) methodrdquoJournal of Systemics Cybernetics and Informatics vol 2 no 4pp 53ndash58 2014

[34] K D Brabanter and P Karsmakers ldquoLS-SVMlab Toolbox UserrsquosGuiderdquo 2011 httpwwwesatkuleuvenbesistalssvmlabdown-loadstutorialv1 8pdf

[35] httpwwwmotionnodecom[36] M Zhang and A A Sawchuk ldquoA feature selection-based frame-

work for human activity recognition using wearable multi-modal sensorsrdquo inProceedings of the International Conference onBody Area Networks (BodyNets rsquo11) Beijing China November2011

[37] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearable sen-sorsrdquo in Proceedings of the ACM International Conference onUbiquitous Computing (UbiComp rsquo12) International Workshopon Situation Activity and Goal Awareness Pittsburgh Pa USASeptember 2012

[38] httpsipiusceduHAD[39] O D Incel M Kose and C Ersoy ldquoA review and taxonomy of

activity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013

[40] Y Liang X Zhou Z Yu and B Guo ldquoEnergy-efficient motionrelated activity recognition on mobile devices for pervasivehealthcarerdquoMobile Networks and Applications vol 19 no 3 pp303ndash317 2014

[41] R Cross ldquoStanding walking running and jumping on a forceplaterdquo American Journal of Physics vol 67 no 4 pp 304ndash3091999

[42] A L Hof J P Van Zandwijk and M F Bobbert ldquoMechanics ofhuman triceps surae muscle in walking running and jumpingrdquoActa Physiologica Scandinavica vol 174 no 1 pp 17ndash30 2002

[43] G Cola A Vecchio and M Avvenuti ldquoImproving the per-formance of fall detection systems through walk recognitionrdquo

Journal of Ambient Intelligence amp Humanized Computing vol 5no 6 pp 843ndash855 2014

[44] H-I Wu B-L Li T A Springer and W H Neill ldquoModellinganimal movement as a persistent random walk in two dimen-sions expected magnitude of net displacementrdquo EcologicalModelling vol 132 no 1-2 pp 115ndash124 2000

[45] C Li M Lin L T Yang and C Ding ldquoIntegrating the enrichedfeaturewithmachine learning algorithms for humanmovementand fall detectionrdquo Journal of Supercomputing vol 67 no 3 pp854ndash865 2014

[46] N Zhang C Williams E Ososanya and W MahmoudldquoStreamflow Prediction Based on Least Squares Support VectorMachinesrdquo 2013 httpwwwaseeorgdocumentssectionsmid-dle-atlanticfall-201311-ASEE2013 Final20Zhangpdf

[47] D Rodriguez-Martin A Sama C Perez-Lopez A Catala JCabestany and A Rodriguez-Molinero ldquoSVM-based postureidentificationwith a single waist-located triaxial accelerometerrdquoExpert Systems with Applications vol 40 no 18 pp 7203ndash72112013

[48] J P Varkey D Pompili and T AWalls ldquoHumanmotion recog-nition using a wireless sensor-based wearable systemrdquo Personalamp Ubiquitous Computing vol 16 no 7 pp 897ndash910 2012

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Electrical and Computer Engineering 7

Table 3 Activity classifier accuracy test

Classifiers Activities recognition accuracy rate () Classifier averageaccuracy ()WF WL WR WU WD RU JU ST SI SL

Classifier 1 983 971 971 100 982Classifier 2 981 991 mdash mdash mdash mdash mdash 986Classifier 3 986 971 957 mdash mdash mdash mdash mdash mdash mdash 971Classifier 4 mdash mdash mdash 971 971 mdash mdash mdash mdash mdash 971Classifier 5 mdash mdash mdash mdash mdash mdash mdash 986 971 986 981

Table 4 Confusion matrix for average recognition accuracy for allactivities

InputAccuracy rate () 956

OutputWF WL WR WU WD RU JU ST SI SL

WF 957 12 31 0 0 0 0 0 0 0WL 28 929 43 0 0 0 0 0 0 0WR 34 52 914 0 0 0 0 0 0 0WU 0 0 0 929 71 0 0 0 0 0WD 0 0 0 57 943 0 0 0 0 0RU 08 0 0 0 0 971 21 0 0 0JU 0 0 0 0 0 29 971 0 0 0ST 0 0 0 0 0 0 0 986 08 06SI 0 0 0 0 0 0 0 21 971 08SL 0 0 0 0 0 0 0 04 10 986

of five activity recognition classifiersThese activity classifiershad over 95 accuracy [24] and were acceptable

The results of these folds are summarized in Table 4The average recognition accuracy of 956 indicates that ourproposed human activity recognition scheme can achievehigh recognition rates for a specific subject Because 2Dwalking and 3Dwalking are similar the recognition accuracyof the five walking activities is low We will attempt to obtainhigher recognition accuracy using an adequate amount oftraining data in future research

We compared the accuracy rate and running time forcommon multiclassification methods All algorithms wererun on a computer with CPU i7-2670QM 22G 8G ram andMatlab 2013a The LS-SVM performed notably well in thetestsThe average running time for the hierarchical classifica-tion framework with the LS-SVM recognizing activities was0021 seconds whichwas less than theANN (Artificial NeuralNetwork) DT (Decision Tree) and 119870NN (119896-Nearest Neigh-bor) algorithmsWe performed the ANNDT and119870NNclas-sifier tests with the built-in functions of MatlabThe LS-SVMmethod was also better than ANN DT and 119870NN in termsof the average recognition accuracy rate for the ten activitiesTable 5 shows the results

6 Conclusion and Future Work

This paper aims to provide an accurate and robust humanactivity recognition scheme The scheme used triaxial

acceleration data a hierarchical recognition scheme andactivity classifiers based on the LS-SVM and the NB algo-rithm The mean variance entropy of magnitude and angleof triaxial acceleration data were used as the features ofthe activity classifiers The scheme effectively recognizeda typical set of daily physical activities with an averageaccuracy of 956 It could distinguish walking (forward leftright upstairs and downstairs) running jumping standingsitting and sleeping activities using only a single triaxialaccelerometer The experimental results of the hierarchicalrecognition scheme show significant potential in its abilityto accurately differentiate activities using triaxial accelerationdata Although the scheme remains to be tested with USC-HAD datasets the core of this scheme is independent of thefeatures of other activity datasets therefore it is applicable toany dataset

The novelty of the proposed human activity recognitionscheme is the introduction of the LS-SVMmethod as the clas-sifier algorithm The LS-SVM is an advanced version of thestandard SVM and there are recently relatively few studiesusing LS-SVM to recognize activities with only one triaxialaccelerometer The human activity recognition scheme withLS-SVM classifiers simplifies the construction of the hierar-chical classification framework and has a lower running timethan other commonmulticlassification algorithms Accuracyis the basic element thatmust be consideredwhen any activityrecognition system is implemented and this recognitionscheme has a high success rate for which it can recognizeten different types of activities with an average accuracy of95

The next stage of our research has two parts First thealgorithms are improved to recognize these activities and theuser will not have to worry about placing the sensors at thecorrect positions to correctly detect the activities Second anunsupervised approach for automatic activity recognition isconsidered An unsupervised learning framework of humanactivity recognition will automatically cluster a large amountof unlabeled acceleration data into discrete groups of activitywhich implies that the human activity recognition can benaturally performed

Conflict of Interests

The author declares that there is no conflict of interestsregarding the publication of this paper

8 Journal of Electrical and Computer Engineering

Table 5 Accuracy rates and running times of the classification methods

Method Accuracy rate () Average rate () Running time (s)WF WL WR WU WD JU RU ST SI SL

ANN 961 914 902 905 854 775 986 967 952 991 921 0085DT 939 946 917 912 908 849 941 957 972 944 929 0411119870NN 935 921 901 882 867 886 938 961 957 938 919 0183LS-SVM 957 929 914 929 943 971 971 986 971 986 956 0021

Acknowledgments

This work was partially supported by AppropriativeResearching Fund for Professors and Doctors GuangdongUniversity of Education under Grant 11ARF04 and Guang-dong Provincial Department of Education under Grants2013LYM 0063 and 2014GXJK161

References

[1] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014

[2] J Hernandez R Cabido A S Montemayor and J J PantrigoldquoHuman activity recognition based on kinematic featuresrdquoExpert Systems vol 31 no 4 pp 345ndash353 2014

[3] J Yin G Tian Z Feng and J Li ldquoHuman activity recognitionbased on multiple order temporal informationrdquo Computers ampElectrical Engineering vol 40 no 5 pp 1538ndash1551 2014

[4] A M Khan Y-K Lee S Y Lee and T-S Kim ldquoA triaxialaccelerometer-based physical-activity recognition via aug-mented-signal features and a hierarchical recognizerrdquo IEEETransactions on Information Technology in Biomedicine vol 14no 5 pp 1166ndash1172 2010

[5] D Trabelsi S Mohammed F Chamroukhi L Oukhellou andY Amirat ldquoAn unsupervised approach for automatic activityrecognition based on hidden markov model regressionrdquo IEEETransactions onAutomation Science and Engineering vol 10 no3 pp 829ndash835 2013

[6] W L Tang and E S Sazonov ldquoHighly accurate recognitionof human postures and activities through classification withrejectionrdquo IEEE Journal of Biomedical and Health Informaticsvol 18 no 1 pp 309ndash315 2014

[7] M-W Lee A M Khan and T-S Kim ldquoA single tri-axialaccelerometer-based real-time personal life log system capableof human activity recognition and exercise information gener-ationrdquo Personal amp Ubiquitous Computing vol 15 no 8 pp 887ndash898 2011

[8] W-Y Deng Q-H Zheng and Z-M Wang ldquoCross-personactivity recognition using reduced kernel extreme learningmachinerdquo Neural Networks vol 53 pp 1ndash7 2014

[9] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014

[10] D P Tao L Jin Y Wang and X Li ldquoRank preserving discrim-inant analysis for human behavior recognition on wireless sen-sor networksrdquo IEEE Transactions on Industrial Informatics vol10 no 1 pp 813ndash823 2014

[11] N Alshurafa W Xu J J Liu et al ldquoDesigning a robust activityrecognition framework for health and exergaming using wear-able sensorsrdquo IEEE Journal of Biomedical and Health Informat-ics vol 18 no 5 pp 1636ndash1646 2014

[12] A Chang S Mota and H Lieberman ldquoGestureNet a commonsense approach to physical activity similarityrdquo in Proceedings ofthe Conference on Electronic Visualisation and the Arts LondonUK July 2014

[13] O Banos M Damas H Pomares F Rojas B Delgado-Marquez and O Valenzuela ldquoHuman activity recognitionbased on a sensor weighting hierarchical classifierrdquo Soft Com-puting vol 17 no 2 pp 333ndash343 2013

[14] J Cheng X Chen andM Shen ldquoA framework for daily activitymonitoring and fall detection based on surface electromyogra-phy and accelerometer signalsrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 1 pp 38ndash45 2013

[15] N Kern B Schiele and A Schmidt ldquoMulti-sensor activity con-text detection for wearable computingrdquo in Ambient Intelligencevol 2875 of Lecture Notes in Computer Science pp 220ndash232Springer Berlin Germany 2003

[16] L Gao A K Bourke and J Nelson ldquoSensor positioning foractivity recognition using multiple accelerometer-based sen-sorsrdquo in Proceedings of the 21st European Symposium on Arti-ficial Neural Networks Computational Intelligence and MachineLearning pp 425ndash430 April 2013

[17] L Gao A K Bourke and J Nelson ldquoEvaluation of accelerome-ter based multi-sensor versus single-sensor activity recognitionsystemsrdquo Medical Engineering and Physics vol 36 no 6 pp779ndash785 2014

[18] S Liu R X Gao D John J W Staudenmayer and P S Freed-son ldquoMultisensor data fusion for physical activity assessmentrdquoIEEE Transactions on Biomedical Engineering vol 59 no 3 pp687ndash696 2012

[19] J Wan M J OrsquoGrady and G M P OrsquoHare ldquoDynamic sensorevent segmentation for real-time activity recognition in a smarthome contextrdquo Personal amp Ubiquitous Computing vol 19 no 2pp 287ndash301 2015

[20] Y Zhan and T Kuroda ldquoWearable sensor-based human activityrecognition from environmental background soundsrdquo Journalof Ambient Intelligence amp Humanized Computing vol 5 no 1pp 77ndash89 2014

[21] httpenwikipediaorgwikiPhysical exercise[22] httpwwwnhlbinihgovhealthhealth-topicstopicsphys[23] httpenwikipediaorgwikiActivities of daily living[24] O D Lara and M A Labrador ldquoA survey on human activity

recognition using wearable sensorsrdquo IEEE CommunicationsSurveys amp Tutorials vol 15 no 3 pp 1192ndash1209 2013

[25] S Arora D Bhattacharjee M Nasipuri L Malik M Kunduand D K Basu ldquoPerformance comparison of SVM and ANNfor handwritten devnagari character recognitionrdquo InternationalJournal of Computer Science Issues vol 18 pp 63ndash72 2010

Journal of Electrical and Computer Engineering 9

[26] J Ren ldquoANN vs SVM which one performs better in classifi-cation of MCCs in mammogram imagingrdquo Knowledge-BasedSystems vol 26 pp 144ndash153 2012

[27] J S Raikwal and K Saxena ldquoPerformance evaluation of SVMand k-nearest neighbor algorithm over medical data setrdquoInternational Journal of Computer Applications vol 50 no 14pp 12ndash24 2012

[28] M Eastwood and B Gabrys ldquoA non-sequential representationof sequential data for churn predictionrdquo in Knowledge-Basedand Intelligent Information and Engineering Systems pp 209ndash218 Springer Berlin Germany 2009

[29] Y Nam and J W Park ldquoChild activity recognition based oncooperative fusion model of a triaxial accelerometer and abarometric pressure sensorrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 2 pp 420ndash426 2013

[30] J A Nasiri N MoghadamCharkari and K Mozafari ldquoEnergy-based model of least squares twin Support Vector Machines forhuman action recognitionrdquo Signal Processing vol 104 pp 248ndash257 2014

[31] K Altun B Barshan and O Tuncel ldquoComparative studyon classifying human activities with miniature inertial andmagnetic sensorsrdquo Pattern Recognition vol 43 no 10 pp 3605ndash3620 2010

[32] R Wang S Kwong D Chen and J Cao ldquoA vector-valued sup-port vectormachinemodel formulticlass problemrdquo InformationSciences vol 235 pp 174ndash194 2013

[33] N Zhang and C Williams ldquoWater quantity prediction usingleast squares support vector machines (LSSVM) methodrdquoJournal of Systemics Cybernetics and Informatics vol 2 no 4pp 53ndash58 2014

[34] K D Brabanter and P Karsmakers ldquoLS-SVMlab Toolbox UserrsquosGuiderdquo 2011 httpwwwesatkuleuvenbesistalssvmlabdown-loadstutorialv1 8pdf

[35] httpwwwmotionnodecom[36] M Zhang and A A Sawchuk ldquoA feature selection-based frame-

work for human activity recognition using wearable multi-modal sensorsrdquo inProceedings of the International Conference onBody Area Networks (BodyNets rsquo11) Beijing China November2011

[37] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearable sen-sorsrdquo in Proceedings of the ACM International Conference onUbiquitous Computing (UbiComp rsquo12) International Workshopon Situation Activity and Goal Awareness Pittsburgh Pa USASeptember 2012

[38] httpsipiusceduHAD[39] O D Incel M Kose and C Ersoy ldquoA review and taxonomy of

activity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013

[40] Y Liang X Zhou Z Yu and B Guo ldquoEnergy-efficient motionrelated activity recognition on mobile devices for pervasivehealthcarerdquoMobile Networks and Applications vol 19 no 3 pp303ndash317 2014

[41] R Cross ldquoStanding walking running and jumping on a forceplaterdquo American Journal of Physics vol 67 no 4 pp 304ndash3091999

[42] A L Hof J P Van Zandwijk and M F Bobbert ldquoMechanics ofhuman triceps surae muscle in walking running and jumpingrdquoActa Physiologica Scandinavica vol 174 no 1 pp 17ndash30 2002

[43] G Cola A Vecchio and M Avvenuti ldquoImproving the per-formance of fall detection systems through walk recognitionrdquo

Journal of Ambient Intelligence amp Humanized Computing vol 5no 6 pp 843ndash855 2014

[44] H-I Wu B-L Li T A Springer and W H Neill ldquoModellinganimal movement as a persistent random walk in two dimen-sions expected magnitude of net displacementrdquo EcologicalModelling vol 132 no 1-2 pp 115ndash124 2000

[45] C Li M Lin L T Yang and C Ding ldquoIntegrating the enrichedfeaturewithmachine learning algorithms for humanmovementand fall detectionrdquo Journal of Supercomputing vol 67 no 3 pp854ndash865 2014

[46] N Zhang C Williams E Ososanya and W MahmoudldquoStreamflow Prediction Based on Least Squares Support VectorMachinesrdquo 2013 httpwwwaseeorgdocumentssectionsmid-dle-atlanticfall-201311-ASEE2013 Final20Zhangpdf

[47] D Rodriguez-Martin A Sama C Perez-Lopez A Catala JCabestany and A Rodriguez-Molinero ldquoSVM-based postureidentificationwith a single waist-located triaxial accelerometerrdquoExpert Systems with Applications vol 40 no 18 pp 7203ndash72112013

[48] J P Varkey D Pompili and T AWalls ldquoHumanmotion recog-nition using a wireless sensor-based wearable systemrdquo Personalamp Ubiquitous Computing vol 16 no 7 pp 897ndash910 2012

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

8 Journal of Electrical and Computer Engineering

Table 5 Accuracy rates and running times of the classification methods

Method Accuracy rate () Average rate () Running time (s)WF WL WR WU WD JU RU ST SI SL

ANN 961 914 902 905 854 775 986 967 952 991 921 0085DT 939 946 917 912 908 849 941 957 972 944 929 0411119870NN 935 921 901 882 867 886 938 961 957 938 919 0183LS-SVM 957 929 914 929 943 971 971 986 971 986 956 0021

Acknowledgments

This work was partially supported by AppropriativeResearching Fund for Professors and Doctors GuangdongUniversity of Education under Grant 11ARF04 and Guang-dong Provincial Department of Education under Grants2013LYM 0063 and 2014GXJK161

References

[1] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014

[2] J Hernandez R Cabido A S Montemayor and J J PantrigoldquoHuman activity recognition based on kinematic featuresrdquoExpert Systems vol 31 no 4 pp 345ndash353 2014

[3] J Yin G Tian Z Feng and J Li ldquoHuman activity recognitionbased on multiple order temporal informationrdquo Computers ampElectrical Engineering vol 40 no 5 pp 1538ndash1551 2014

[4] A M Khan Y-K Lee S Y Lee and T-S Kim ldquoA triaxialaccelerometer-based physical-activity recognition via aug-mented-signal features and a hierarchical recognizerrdquo IEEETransactions on Information Technology in Biomedicine vol 14no 5 pp 1166ndash1172 2010

[5] D Trabelsi S Mohammed F Chamroukhi L Oukhellou andY Amirat ldquoAn unsupervised approach for automatic activityrecognition based on hidden markov model regressionrdquo IEEETransactions onAutomation Science and Engineering vol 10 no3 pp 829ndash835 2013

[6] W L Tang and E S Sazonov ldquoHighly accurate recognitionof human postures and activities through classification withrejectionrdquo IEEE Journal of Biomedical and Health Informaticsvol 18 no 1 pp 309ndash315 2014

[7] M-W Lee A M Khan and T-S Kim ldquoA single tri-axialaccelerometer-based real-time personal life log system capableof human activity recognition and exercise information gener-ationrdquo Personal amp Ubiquitous Computing vol 15 no 8 pp 887ndash898 2011

[8] W-Y Deng Q-H Zheng and Z-M Wang ldquoCross-personactivity recognition using reduced kernel extreme learningmachinerdquo Neural Networks vol 53 pp 1ndash7 2014

[9] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014

[10] D P Tao L Jin Y Wang and X Li ldquoRank preserving discrim-inant analysis for human behavior recognition on wireless sen-sor networksrdquo IEEE Transactions on Industrial Informatics vol10 no 1 pp 813ndash823 2014

[11] N Alshurafa W Xu J J Liu et al ldquoDesigning a robust activityrecognition framework for health and exergaming using wear-able sensorsrdquo IEEE Journal of Biomedical and Health Informat-ics vol 18 no 5 pp 1636ndash1646 2014

[12] A Chang S Mota and H Lieberman ldquoGestureNet a commonsense approach to physical activity similarityrdquo in Proceedings ofthe Conference on Electronic Visualisation and the Arts LondonUK July 2014

[13] O Banos M Damas H Pomares F Rojas B Delgado-Marquez and O Valenzuela ldquoHuman activity recognitionbased on a sensor weighting hierarchical classifierrdquo Soft Com-puting vol 17 no 2 pp 333ndash343 2013

[14] J Cheng X Chen andM Shen ldquoA framework for daily activitymonitoring and fall detection based on surface electromyogra-phy and accelerometer signalsrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 1 pp 38ndash45 2013

[15] N Kern B Schiele and A Schmidt ldquoMulti-sensor activity con-text detection for wearable computingrdquo in Ambient Intelligencevol 2875 of Lecture Notes in Computer Science pp 220ndash232Springer Berlin Germany 2003

[16] L Gao A K Bourke and J Nelson ldquoSensor positioning foractivity recognition using multiple accelerometer-based sen-sorsrdquo in Proceedings of the 21st European Symposium on Arti-ficial Neural Networks Computational Intelligence and MachineLearning pp 425ndash430 April 2013

[17] L Gao A K Bourke and J Nelson ldquoEvaluation of accelerome-ter based multi-sensor versus single-sensor activity recognitionsystemsrdquo Medical Engineering and Physics vol 36 no 6 pp779ndash785 2014

[18] S Liu R X Gao D John J W Staudenmayer and P S Freed-son ldquoMultisensor data fusion for physical activity assessmentrdquoIEEE Transactions on Biomedical Engineering vol 59 no 3 pp687ndash696 2012

[19] J Wan M J OrsquoGrady and G M P OrsquoHare ldquoDynamic sensorevent segmentation for real-time activity recognition in a smarthome contextrdquo Personal amp Ubiquitous Computing vol 19 no 2pp 287ndash301 2015

[20] Y Zhan and T Kuroda ldquoWearable sensor-based human activityrecognition from environmental background soundsrdquo Journalof Ambient Intelligence amp Humanized Computing vol 5 no 1pp 77ndash89 2014

[21] httpenwikipediaorgwikiPhysical exercise[22] httpwwwnhlbinihgovhealthhealth-topicstopicsphys[23] httpenwikipediaorgwikiActivities of daily living[24] O D Lara and M A Labrador ldquoA survey on human activity

recognition using wearable sensorsrdquo IEEE CommunicationsSurveys amp Tutorials vol 15 no 3 pp 1192ndash1209 2013

[25] S Arora D Bhattacharjee M Nasipuri L Malik M Kunduand D K Basu ldquoPerformance comparison of SVM and ANNfor handwritten devnagari character recognitionrdquo InternationalJournal of Computer Science Issues vol 18 pp 63ndash72 2010

Journal of Electrical and Computer Engineering 9

[26] J Ren ldquoANN vs SVM which one performs better in classifi-cation of MCCs in mammogram imagingrdquo Knowledge-BasedSystems vol 26 pp 144ndash153 2012

[27] J S Raikwal and K Saxena ldquoPerformance evaluation of SVMand k-nearest neighbor algorithm over medical data setrdquoInternational Journal of Computer Applications vol 50 no 14pp 12ndash24 2012

[28] M Eastwood and B Gabrys ldquoA non-sequential representationof sequential data for churn predictionrdquo in Knowledge-Basedand Intelligent Information and Engineering Systems pp 209ndash218 Springer Berlin Germany 2009

[29] Y Nam and J W Park ldquoChild activity recognition based oncooperative fusion model of a triaxial accelerometer and abarometric pressure sensorrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 2 pp 420ndash426 2013

[30] J A Nasiri N MoghadamCharkari and K Mozafari ldquoEnergy-based model of least squares twin Support Vector Machines forhuman action recognitionrdquo Signal Processing vol 104 pp 248ndash257 2014

[31] K Altun B Barshan and O Tuncel ldquoComparative studyon classifying human activities with miniature inertial andmagnetic sensorsrdquo Pattern Recognition vol 43 no 10 pp 3605ndash3620 2010

[32] R Wang S Kwong D Chen and J Cao ldquoA vector-valued sup-port vectormachinemodel formulticlass problemrdquo InformationSciences vol 235 pp 174ndash194 2013

[33] N Zhang and C Williams ldquoWater quantity prediction usingleast squares support vector machines (LSSVM) methodrdquoJournal of Systemics Cybernetics and Informatics vol 2 no 4pp 53ndash58 2014

[34] K D Brabanter and P Karsmakers ldquoLS-SVMlab Toolbox UserrsquosGuiderdquo 2011 httpwwwesatkuleuvenbesistalssvmlabdown-loadstutorialv1 8pdf

[35] httpwwwmotionnodecom[36] M Zhang and A A Sawchuk ldquoA feature selection-based frame-

work for human activity recognition using wearable multi-modal sensorsrdquo inProceedings of the International Conference onBody Area Networks (BodyNets rsquo11) Beijing China November2011

[37] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearable sen-sorsrdquo in Proceedings of the ACM International Conference onUbiquitous Computing (UbiComp rsquo12) International Workshopon Situation Activity and Goal Awareness Pittsburgh Pa USASeptember 2012

[38] httpsipiusceduHAD[39] O D Incel M Kose and C Ersoy ldquoA review and taxonomy of

activity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013

[40] Y Liang X Zhou Z Yu and B Guo ldquoEnergy-efficient motionrelated activity recognition on mobile devices for pervasivehealthcarerdquoMobile Networks and Applications vol 19 no 3 pp303ndash317 2014

[41] R Cross ldquoStanding walking running and jumping on a forceplaterdquo American Journal of Physics vol 67 no 4 pp 304ndash3091999

[42] A L Hof J P Van Zandwijk and M F Bobbert ldquoMechanics ofhuman triceps surae muscle in walking running and jumpingrdquoActa Physiologica Scandinavica vol 174 no 1 pp 17ndash30 2002

[43] G Cola A Vecchio and M Avvenuti ldquoImproving the per-formance of fall detection systems through walk recognitionrdquo

Journal of Ambient Intelligence amp Humanized Computing vol 5no 6 pp 843ndash855 2014

[44] H-I Wu B-L Li T A Springer and W H Neill ldquoModellinganimal movement as a persistent random walk in two dimen-sions expected magnitude of net displacementrdquo EcologicalModelling vol 132 no 1-2 pp 115ndash124 2000

[45] C Li M Lin L T Yang and C Ding ldquoIntegrating the enrichedfeaturewithmachine learning algorithms for humanmovementand fall detectionrdquo Journal of Supercomputing vol 67 no 3 pp854ndash865 2014

[46] N Zhang C Williams E Ososanya and W MahmoudldquoStreamflow Prediction Based on Least Squares Support VectorMachinesrdquo 2013 httpwwwaseeorgdocumentssectionsmid-dle-atlanticfall-201311-ASEE2013 Final20Zhangpdf

[47] D Rodriguez-Martin A Sama C Perez-Lopez A Catala JCabestany and A Rodriguez-Molinero ldquoSVM-based postureidentificationwith a single waist-located triaxial accelerometerrdquoExpert Systems with Applications vol 40 no 18 pp 7203ndash72112013

[48] J P Varkey D Pompili and T AWalls ldquoHumanmotion recog-nition using a wireless sensor-based wearable systemrdquo Personalamp Ubiquitous Computing vol 16 no 7 pp 897ndash910 2012

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Electrical and Computer Engineering 9

[26] J Ren ldquoANN vs SVM which one performs better in classifi-cation of MCCs in mammogram imagingrdquo Knowledge-BasedSystems vol 26 pp 144ndash153 2012

[27] J S Raikwal and K Saxena ldquoPerformance evaluation of SVMand k-nearest neighbor algorithm over medical data setrdquoInternational Journal of Computer Applications vol 50 no 14pp 12ndash24 2012

[28] M Eastwood and B Gabrys ldquoA non-sequential representationof sequential data for churn predictionrdquo in Knowledge-Basedand Intelligent Information and Engineering Systems pp 209ndash218 Springer Berlin Germany 2009

[29] Y Nam and J W Park ldquoChild activity recognition based oncooperative fusion model of a triaxial accelerometer and abarometric pressure sensorrdquo IEEE Journal of Biomedical andHealth Informatics vol 17 no 2 pp 420ndash426 2013

[30] J A Nasiri N MoghadamCharkari and K Mozafari ldquoEnergy-based model of least squares twin Support Vector Machines forhuman action recognitionrdquo Signal Processing vol 104 pp 248ndash257 2014

[31] K Altun B Barshan and O Tuncel ldquoComparative studyon classifying human activities with miniature inertial andmagnetic sensorsrdquo Pattern Recognition vol 43 no 10 pp 3605ndash3620 2010

[32] R Wang S Kwong D Chen and J Cao ldquoA vector-valued sup-port vectormachinemodel formulticlass problemrdquo InformationSciences vol 235 pp 174ndash194 2013

[33] N Zhang and C Williams ldquoWater quantity prediction usingleast squares support vector machines (LSSVM) methodrdquoJournal of Systemics Cybernetics and Informatics vol 2 no 4pp 53ndash58 2014

[34] K D Brabanter and P Karsmakers ldquoLS-SVMlab Toolbox UserrsquosGuiderdquo 2011 httpwwwesatkuleuvenbesistalssvmlabdown-loadstutorialv1 8pdf

[35] httpwwwmotionnodecom[36] M Zhang and A A Sawchuk ldquoA feature selection-based frame-

work for human activity recognition using wearable multi-modal sensorsrdquo inProceedings of the International Conference onBody Area Networks (BodyNets rsquo11) Beijing China November2011

[37] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearable sen-sorsrdquo in Proceedings of the ACM International Conference onUbiquitous Computing (UbiComp rsquo12) International Workshopon Situation Activity and Goal Awareness Pittsburgh Pa USASeptember 2012

[38] httpsipiusceduHAD[39] O D Incel M Kose and C Ersoy ldquoA review and taxonomy of

activity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013

[40] Y Liang X Zhou Z Yu and B Guo ldquoEnergy-efficient motionrelated activity recognition on mobile devices for pervasivehealthcarerdquoMobile Networks and Applications vol 19 no 3 pp303ndash317 2014

[41] R Cross ldquoStanding walking running and jumping on a forceplaterdquo American Journal of Physics vol 67 no 4 pp 304ndash3091999

[42] A L Hof J P Van Zandwijk and M F Bobbert ldquoMechanics ofhuman triceps surae muscle in walking running and jumpingrdquoActa Physiologica Scandinavica vol 174 no 1 pp 17ndash30 2002

[43] G Cola A Vecchio and M Avvenuti ldquoImproving the per-formance of fall detection systems through walk recognitionrdquo

Journal of Ambient Intelligence amp Humanized Computing vol 5no 6 pp 843ndash855 2014

[44] H-I Wu B-L Li T A Springer and W H Neill ldquoModellinganimal movement as a persistent random walk in two dimen-sions expected magnitude of net displacementrdquo EcologicalModelling vol 132 no 1-2 pp 115ndash124 2000

[45] C Li M Lin L T Yang and C Ding ldquoIntegrating the enrichedfeaturewithmachine learning algorithms for humanmovementand fall detectionrdquo Journal of Supercomputing vol 67 no 3 pp854ndash865 2014

[46] N Zhang C Williams E Ososanya and W MahmoudldquoStreamflow Prediction Based on Least Squares Support VectorMachinesrdquo 2013 httpwwwaseeorgdocumentssectionsmid-dle-atlanticfall-201311-ASEE2013 Final20Zhangpdf

[47] D Rodriguez-Martin A Sama C Perez-Lopez A Catala JCabestany and A Rodriguez-Molinero ldquoSVM-based postureidentificationwith a single waist-located triaxial accelerometerrdquoExpert Systems with Applications vol 40 no 18 pp 7203ndash72112013

[48] J P Varkey D Pompili and T AWalls ldquoHumanmotion recog-nition using a wireless sensor-based wearable systemrdquo Personalamp Ubiquitous Computing vol 16 no 7 pp 897ndash910 2012

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of