This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2519679, IEEE Sensors
Journal
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2015 1

A Triaxial Accelerometer-Based Human Activity

Recognition via EEMD-Based Features and
Game-Theory-Based Feature Selection
Zhelong Wang, Donghui Wu, Jianming Chen, Ahmed Ghoneim and M. Anwar Hossain

Abstract—In recent years, sensor-based human activity recog- based researches with different placements and single sensor
nition has attracted lots of researches. This paper presents mounted researches. Multiple sensors attached to different
a single wearable triaxial accelerometer based human activity parts of subjects have been investigated by lots of systems
recognition system which can be used in the real life of activity
monitoring. The sensor is attached around different parts of [5]–[9]. Eight foot force sensors and smart phone GPS placed
the body: waist and left ankle, respectively. In order to im- on the two insole positions have been utilized to detect five
prove the accuracy and reduce the computational complexity, different mobility activities: walking, cycling, bus-passenger,
Ensemble empirical mode decomposition (EEMD) based features car-passenger, and car-driver [9]. Five accelerometers have
and feature selection (FS) method are introduced, respectively. been utilized by Bao and Intille [10] to classify 20 different
Considering the feature interaction, game theory based feature
selection method is proposed to evaluate the features. Relevant activities. The sensors are attached to five different parts of
and distinguished features that are robust to the placement of the body: right hip, dominant wrist, non-dominant upper arm,
sensors are selected. In the experiment, the data acquired from dominant ankle, and non-dominant thigh. Four features: mean,
the two different parts of the bodywaist and ankle is utilized energy, frequency-domain entropy, and correlation between
to evaluate the proposed FS method. To verify the effectiveness axes, have been extracted for recognition by different classi-
of the proposed method, k-NN and SVM are used to recognize
the human activities from waist and ankle. Experiment results fiers. Zhu and Sheng [11] introduce two accelerometers placed
demonstrate the effectiveness of the introduced EEMD based on the wrist and waist of subject, respectively, to classify 11
features for human activity recognition. Comparing with the activities by the proposed hierarchical classifier. Though the
representative FS methods, including Relief-F and mRMR, the multi-sensor system have made outstanding achievement, this
proposed feature selection approach selects fewer features and approach is not feasible for the long-term activity monitoring
provides higher accuracy. The results also show that the triaxial
accelerometer around the waist produces optimal results. that is because of the limitation of the number of sensors with
predicted orientation and the cable connections.
Index Terms—Human Activity Recognition, Game Theory, Comparatively, single sensor based systems mounted at
Feature Selection, Wearable Triaxial Accelerometer, EEMD.
different sites of subjects have also been studied by a number
of researches [12]–[17]. The data from 3-axis accelerometer
I. I NTRODUCTION of a cell phone attached to the front plants leg pocket has
been utilized by [17] to recognize the six specific activities:
W ITH the development of microelectronics, human ac-
tivity recognition using wearable inertial sensors has
gained tremendous attentions in the world among a range of
walking, jogging,upstairs, downstairs, sitting, and standing.
Khan et al. [12] applied a single triaxial accelerometer attached
health-related areas. By attaching different types of sensors to chest with particular orientation to identify fifteen activities
on different locations and objects, human activities can be by hierarchical model using the linear-discriminant analysis
tracked and monitored. A major goal of wearable sensor-based and ANNs, and an average accuracy of 97.9% has been
health-related research is to long-term monitor daily activities achieved. The features of AR-coefficients, signal-magnitude
of patients. Daily activities may provide additional information area (SMA), and tilt angle (TA) are extracted for classification
to medical staff to accurately diagnose chronic diseases, as [12]. In order to detect activities irrespective of the position
well as design the care plan for the patients. The quality of the sensor, Khan et al. [15] chose the data acquired from
of healthcare provided to the elderly and children could be five positions: chest pocket, front left trousers pocket, front
effectively improved by human activity recognition [1]–[4]. right trousers pocket, rear trousers pocket, and inner jacket
Previous works about human activity recognition systems pocket, respectively, to classify 7 activities. However, the
are mainly forced on two main researches: multiple sensors transition states are all excluded from the experiments.The
extracted features are Spectral entropy(SE), Autoregressive
Zhelong Wang and Donghui Wu are with the School of Control Science (AR) coefficients, Signal magnitude area (SMA). A single,
and Engineering,Dalian University of Technology, Dalian 116024, Liaoning, triaxial accelerometer has been utilized by Cheng and Jhan
China (e-mail:w donghui@163.com).
Jianming Chen: School of Electric Power, North China University of Water [18] for fall detection with sensors placed on left ankle, right
Conservancy and Electric Power, Zhengzhou 450046, China. ankle, chest and waist, respectively. Raw triaxial data, signal
Ahmed Ghoneim and M. Anwar Hossain are with the College of Computer magnitude vector (SMV) and signal magnitude area (SMA)
and Information Sciences, King Saud University, Riyadh 11543, Saudi Arabia.
Dr. Ahmed is also with the Department of Computer Science, College of are extracted as features.
Science, Menoufia University, Egypt. Though most of the aforementioned researches have inves-

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2519679, IEEE Sensors
Journal
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2015 2

tigated fall detection and daily activity recognition, most of TABLE I

systems require sensors in a predicted orientation and position T HE ACQUISITION FEATURES OF DIFFERENT ACTIVITIES FROM THE FOUR
BODY PARTS
on human body or else exclude transition states, etc. In this
paper, we focused on a single accelerometer based human
activity recognition and selecting optimum number of features Spectral entropy
energy [10] Root Mean Square (RMS) [32]
(SE) [10]
robust to the position around waist and left ankle. kurtosis [32] correlation between axes [10] tilt angle(TA) [12]
High quality features are essential to improve the classi- Mean [10] signal-magnitude area (SMA) [12] variance [10]
fication accuracy of pattern recognition. In human activity EEMD∗ Mean-crossing rate (MCR) [32] AR-coefficients [12]
recognition, some researches pool all available features into
one vector as the input of classifiers. However, some features
may be irrelevant or redundant, and provide no contribution evaluated and analysed. The conclusion is finally stated in
to the classification accuracy. Furthermore, the computational section V.
cost of this approach is large. Some features might even
confuse the classifier and reduce the classification accuracy II. F EATURE S ELECTION
[19], [20].Therefore, it is necessary to extract optimum feature A large set of features have been extracted from the raw
subsets to achieve the best classification performance and re- acceleration signal during the activity. Some of features are
duce the computational cost. In addition,the selected optimum effective for activity recognition detailed in the previously
feature subsets are also required to be robust to the orientation published researches [18], [23], [32], [33]. However, according
and position on human body. to our best knowledge, some features rarely used for activity
Up to recent, several approaches are employed in feature recognition previously will be detailed in this paper. All these
selection, such as SVM [21], genetic algorithm [22], kNN [23] features have be listed in Table.I.
and so on. All of these feature selection approaches are usually
classified into three main groups: wrapper, embedded, and
filter methods [24]. Wrapper method [25] applied the induced A. Ensemble empirical mode decomposition
learning algorithm as a black box to evaluate each feature Ensemble empirical mode decomposition (EEMD), as a
subsets. Embedded methods [26] incorporate knowledge about time-analysis method, has been proposed to deal with climate
the specific structure of a given machine learning approach data and a high resolution digitalized sound record [34]–
to evaluate feature subsets. Although these methods may [36]. Comparing with empirical mode decomposition (EMD),
achieve a good results, the computational cost are expensive. EEMD overcomes the mode mixing problem. In the following,
In addition, many classifiers are prone to overfitting and show EMD will be firstly introduced. EEMD will be described
sensitiveness to initialization. following.
Comparatively, filter methods have no reliance with the 1) empirical mode decomposition: EMD has been widely
learning method, and have much lower computational com- applied to fault detection of rotating machinery [37]–[40],
plexity than the former two approaches. Meanwhile, they which is able to decompose a complicated signal into some
achieve comparable classification accuracy for most classifiers. intrinsic mode functions (IMF). IMFs as simple oscillatory
To date, virous efficient filter methods have been proposed functions, which are irrelevant with each other, have two
[27]–[31]. However, most of these methods tend to ignore fea- properties:
tures which as a group have strong discriminatory power but
are weak as individuals. To tackle this problem, cooperative (1) Throughout the data set, the amount of extrema and
game theory based feature selection method is proposed to the amount of zero-crossings should be either equal or differ
select optimum feature subsets which is robust to the position at most by one;
around waist, chest, right ankle and left ankle. The main
contributions of this paper are listed as follows: (2) At any data point, the mean value of the envelope
1) Applying multiple sensors may increase the cost, the energy defined by local maxima and the envelope defined by the
cost and discomfort, so this paper is focused on a single local minima is zero.
accelerometer based human activity recognition. According to
our best knowledge, EEMD based features are presented in The procedure of EMD decomposition is shown as follows,
this paper which are rarely mentioned previously for activity which is also briefly described in Figs. 1:
recognition; (1) collect the local maxima and the minima of the initial
2) Cooperative game theory based feature selection approach signal x(t);
is proposed to select optimum number of features which are (2) connect the local extrema using cubic spline to form the
relevance and robust to the position around waist and left upper and lower envelopes(emax (t) and emin (t)), respectively;
ankle. (3) using the equation
The detailed structure of this paper is arranged as follows:
m(t) = [emax (t) + emin (t)]/2
after the introduction, features used in this paper are summa-
rized in section II. Cooperative game theory based feature to get the local mean of the two envelopes;
selection approach is presented in section III. In section (4) calculate h(t) = x(t) − m(t);
IV, the proposed features and the method are experimentally (5) if h(t) satisfy the two properties of IMF, then h(t) will be

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2519679, IEEE Sensors
Journal
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2015 3

treated as an IMF, else set x(t) = h(t), and go back to step(1); i = 1, 2, · · ·, N; j = 1, 2, · · ·, M

(4) Report the mean c− mean j , j = 1, 2, · · ·, M of each of all

X(t) the IMFs as the final IMFs.
3) The extracted useful EEMD-based feature: IMFs
(c− mean1 , c− mean2 , · · ·, c− meanM ) decomposed by EEMD
H1 C1 consist of signal components of different frequency, which
reflect the trend of the signal data. Different actions contains
different IMFs, and the energy of each IMF vary significantly.
H2 C2
Therefore, it is necessary to extract the energy of each IMF
as the features to identify different actions. Generally, the
. previous m IMFs are the most distinctive features. So, the
.
.
energies of the previous m IMFs are integrated as a feature
set. The procedure of the feature extraction is shown as
follows:
HN CN (1) Decompose the action signal using EEMD into IMFs,
extract the previous m IMFs;
Fig. 1. The schematic of decomposition of EMD (2) Calculate the energy E j of each IMF:
Z +∞
At last, the signal x(t) is decomposed in terms of IMFs:
Ej = |c− mean j (t)|2 dt, j = 1, 2, · · ·, m
n=N −∞
x(t) = ∑ cn + rN (1) (3) Obtain the feature set:
n=1

where cn is the nth IMF, rN is the residue of signal x(t), and E = [E1 , E2 , · · ·, Em ]
N is the total number of extracted IMFs.
2) Ensemble empirical mode decomposition: EEMD as the The normalization procedure of the feature set is carried
improvement approach EMD has been presented to overcome out: Firstqcalculate the total energy of IMFs by the equation
the mode mixing problem of EMD. To improve the accuracy Etotal = ∑mj=1 |E j |2 ; Then we got the normalized feature set:
of measurements, white noise as the possible random noise is
0 E1 E2 Em
added to the signal data [34]. Comparing with EMD, EEMD E =[ , , · · ·, ]
as the improment of EMD is mainly developed as follows: Etotal Etotal Etotal
0
(1) The different white noise is added to the target signal,and E is utilized as the EEMD-based feature set for human activity
a new signal data is composed each time. recognition. In the experiment, eight IMFs are selected as a
(2) Using the EMD method to decompose the new signal feature set, which means that m is set as 8 in this paper for
IMFs. the experiment.

Based on the principle and EMD, EEMD approach can be

implemented as following procedure: III. G AME -T HEORY- BASED F EATURE S ELECTION
(1) Initialize the number of ensemble M, the amplitude of the Information theory is originally used within the context
added white noise, and i = 1. of communication theory about the data compression and so
(2) Performance of each ith (i ∈ [1, 2, · · ·, N]) trial on the new on [41], [42]. Entropy and mutual information are basically
signal with the white noise. defined in information theory. Game theory as a mathematical
(a) Add a white noise series with the same amplitude of the method describes the phenomenon of conflict and cooperation
signal x(t) to the targeted signal: between intelligent rational decision-makers, which has been
utilized to deal with the camera assignment and handoff in a
xi (t) = x(t) + ψi (t)
video network based on a set of user-supplied criteria [43] and
where ψi (t) denotes the ith added white noise and xi (t) also introduced in wireless sensor networks for the application
represents the ith trail signal with added noise. of power control and energy saving [44], routing protocol
(b) Apply EMD method to decompose the signal xi (t) into design, topology control [45], data collection and so on [46].
IMFs ci, j ( j = 1, 2, · · ·, M), where ci, j denotes the jth IMF of In this paper, game theory is utilized for feature selection,
ith trail, and M means the the total number of IMFs. and the proposed framework of game theory based feature
(c) Repeat step (a) and (b) until i ≥ N, and each time different selection will be detailed in this section.
white noise series is added into the signal xi (t).
(3) Calculate the ensemble mean c− mean j of the N trials for
A. Basic Concepts about Information Theory
each IMF:
Entropy in statistics is a logarithmic measure of uncertainty
1 N of a random variable X which is referred to the probability of
c− mean j = ∑ ci, j (t)
N i=1 the number of states p(x(l)), x(l) ∈ X. The more such states

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2519679, IEEE Sensors
Journal
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2015 4

available with appreciable probability, the greater the entropy. full feature set with M features as F = { f1 , f2 , · · ·, fM }, fi ∈ F
Formally, the entropy of the variable X can be described: , F0 = F − fi , and C means the variable of instance class.
m
Definition 1. (strongly relevance)
H(x) = − ∑ p(x(l)) log p(x(l)) (2)
l=1
If and only if I( fi ;C|F0 ) > 0, feature fi is strongly relevance.
Let X and Y be the two random variable discrete variables Definition 2. (weakly relevance)
with probability mass function p(x(l)) and p(y(i)), x(l) ∈ X If and only if I( fi ;C|F0 ) = 0, and ∃S ∈ F0 , such that
and y( j) ∈ Y . The joint entropy of the two variables with I( fi ;C|S) > 0 , feature fi is identified weakly relevance.
joint function mass function p(x(l), y(i)):
Definition 3. (irrelevance)
m m The feature fi is said to be irrelevance, if and only if
H(x, y) = − ∑ ∑ p(x(l), y(i)) log p(x(l), y(i)) (3) I( fi ;C|S) = 0.
l=1 i=1

where 2) redundancy: If two features are completely correlated,

the two features are consider redundant to each other. Features
max(H(x), H(y) ≤ H(x, y) ≤ H(x) + H(y) (4) labeled redundancy should be distinguished from weakly rele-
vance but not redundant features. Therefore, MI and CMI has
When the variable X and Y are completely independent,
been utilized to distinguish the redundant features.
the joint entropy H(x, y) achieves the maximum value. The
minimum entropy is obtained when the variable X and Y are Definition 4. (redundancy)
completely interdependent on each other. If the features fi and f j are said to be redundancy, the
following formulation should be satisfied I( fi ;C| f j ) ≤ I( fi ;C).
The conditional entropy of the two random variables X and
Y is defined as follows:
3) interdependence: Interdependence implies that the im-
m pact of features which can not be apart from each other on
H(x | y) = H(x, y) − H(y) = ∑ p(y(i))H(x | Y = y(i)) (5) the recognition. If the features fi and f j are interdependence,
i=1
the relationship among the features fi , f j and the variable of
where instance class C should satisfy the following formulation:
0 ≤ H(x | y) ≤ H(x) (6)
I( fi ;C| f j ) > I( fi ;C)
Mutual information (MI) as the measure of the variables’
(X and Y ) mutual dependence is given as follows:
C. The Proposed Framework of Feature Selection
p(x, y)
I(x, y) = ∑ ∑ p(x, y) log p(x)p(y) (7) Based on the mutual information, minimum-Redundancy
x∈X y∈Y
Maximum-Relevance (mRMR) has been proposed to deal
The relationship between MI and entropy of random vari- with feature selection [27]. Though mRMR has been
ables is shown: widely applied in different areas, mRMR is equivalent to
I(x, y) = H(x) − H(x|y) Max-dependency for firs-order incremental search without
I(x, y) = H(y) − H(y|x) (8) considering the performance of the interaction of the features.
Based on the criterion of mRMR and considering the
I(x, y) = H(y) + H(x) − H(x, y) interaction of features, the criterion of Game-theory based
In information theory, conditional mutual information feature selection is proposed:
(CMI) means the expected value of the MI of the two variables
{X,Y } given a discrete random variable Z, which measures the 1
J = (1 + ω)(I( fi ;C) − ∑ I( fi ; f j )) (9)
interaction of the three variables and can be formally defined |S| f j ∈S
by:
I(X,Y |Z) = H(X|Z) − H(X|Y, Z) where fi ∈ F − S, and the parameter ω, which is integrated
p(x, y | z) with the influence of the interaction of features and obtained
= ∑ ∑ ∑ log p(x | z)p(y | z) through Game-theory based method, regulates the relative
x∈X y∈Y z∈Z
importance of the MI between the candidate feature and the
already-selected features with respect to the MI with the output
B. Definitions of relevance, redundancy and interdependence class.
1) relevance: In information theory, relevance, which im- In game theory, the powers of game players is measured by
plies the correlation between a feature and the predictive of Shapley value proposed by Shapley [50], which is introduced
the class, is widely measured by mutual information. Under in this method to evaluate each feature weight. In feature
the framework of MI, features can be classified into three selection, each coalition can be selected as a candidate subset
levels of relevance: strongly relevance, weakly relevance and for the final optimal feature subset. Considering the relevance,
irrelevance [47]–[49], which are defined following.Denote a redundance and interdependence of features, the Shapley value

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2519679, IEEE Sensors
Journal
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2015 5

Algorithm 1 The proposed Gamed-theory based feature selection

Input: A training sample Data with feature set F = {F1 , F2 , · ·
·, FN } ; Class label Label and a predicted selected feature
number K.
Output: The selected feature subset FS.
Initialize: FS = φ , k = 0,initial Shapley value based feature
interaction weight wi = 0. (a) (b)
while k < K do
for i = 1 to N Accelerameter
Creat all coalitions set L = {L1 , L2 , · · ·, Lτ } from feature
set L \ i limited by σ ;
for each coalition L j ∈ L

e
be
calculate the function ∆i (L j )

Zig
end
using equation (12) to calculate the Shapley value wi of feature
(c)
fi
calculate the elements of equation (9): I( fi ;C) and Fig. 2. The inertial sensor based experimental plateform(a)collection node
1 containing the triaxial accelerometer;(b) The receiver node;(c)The graphical
|S| ∑ f j ∈S I( f i ; f j );
depiction of experimental setup;
using equation (9) to calculate the criterion GT (i)
end
selected the feature fi with largest GT (i)
FS = FS ∪ { fi } IV. E XPERIMENTAL R ESULTS AND A NALYSIS
L = L \ { fi }
k = k + 1; A. Experimental Setup and Data Sets
end while
The experimental dataset has been acquired in our labo-
ratory using the wearable platform as shown in Fig. 2.The
as an efficient approach is utilized to evaluate the contribution inertial measurement units (IMU) performed as the major
of features, which is formulated as following: component of the experiment platform is mainly composed
of a three-axis accelerometer. The collection node and the
|L|!(m − |L| − 1)! receiver node are shown in Fig. 2(a) and Fig. 2(b), respectively.
Ψi (ψ) = ∑ ∆i (L) (10)
L⊂M\{i}
m! Fig. 2(c) presents the complete system implementation. In sys-
tem, ZigBee protocol has been utilized by the accelerometers
and to send the signal data to the receiver node. Once received by
∆i (L) = ψ(L ∪ i) − ψ(L) (11) the receiver node, the data packets are sent to the computer
side. Fig. 3(a) and Fig. 4(a) show the placements of the
where m means the number of players and the sum extends accelerometers, and the accordingly action data of ”walking”
over all subsets L of M not including player i. from left ankle and waist is presented in Fig. 3(b) and
Fig. 4(b). Previous researches have demonstrated that human
The interdependence of features can be defined following : activity can be recognized by signals at and below 18Hz [13],
[23]. Therefore, a sampling rate of 100Hz is considered to be
(
1, i f I( f j ,C| fi ) ≥ I( f j ,C) more than sufficient. The data sampling rate is 100Hz. Five
Φ(i, j) = (12) subjects take part in this experiment. Each subject is asked
0, otherwise
to tie the IMU to the two different body parts: waist and left
ankle. The actions performed by each subject are listed in the
Based on the aforementioned information, the function ∆i (L)
first column of Table II. After preprocessing including the
can be redefined associated with feature information:
approaches of outlier removal, the sliding window technique
(
|L|
is adopted to cut the data into windows with a same length.
1, i f I(L,C|I( fi ) ≥ 0 and ∑ f j ∈L Φ(i, j) ≥ 2 100 samples are chosen as the length of each window. A
∆i (L) =
0, otherwise 50% overlap between adjacent windows is also adopted to
(13) avoid information loss at the boundary of a single window.
This equation means that only if the feature is relevance with In order to extract the feature of AR-coefficients, AR model
the coalition and interdependent with at least half of the L with 3-order has been utilized in each window. Therefore,
features, the Shapley value of feature fi will be positive. the features extracted from each window come up to 58
dimensions. The features are mentioned above. All features
The framework of the proposed game theory based feature have been normalized to the interval [0, 1] [51] . The quantities
selection using Shapley value to evaluate the features will be of the features in different classes are shown in the second
detailed in Algorithm 1. columns of Table II.

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2519679, IEEE Sensors
Journal
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2015 6

1 TABLE III

Accerleration amplitude
0 ACCURACY OF CLASSIFICATION FOR INDIVIDUAL ACTIVITIES FROM LEFT
ANKLE AND WAIST USING KNN

Waist
−1

−2 the left ankle the waist

Activity
FS FS+proposed FS FS+proposed
−3
0 200 400 600 800 1000 Sitting(ST) 90.20% 93.88% 99.59% 99.59%
10
Accerleration samples Lying(LY) 97.96% 98.37% 99.18% 99.59%

Accerleration amplitude
x−axis standing up
5 y−axis 73.36% 77.46% 88.93% 90.57%
from lying(SFL)
z−axis
Ankle 0 Standing(SD) 86.46% 87.27% 92.12% 93.13%
Walking(W) 51.78% 56.20% 62.03% 63.20%
−5
Running(R) 95.75% 97.16% 90.47% 93.10%
−10 Bicycling
0 200 400 600 800 1000 62.93% 66.33% 70.54% 71.54%
Accerleration samples (50 watt)(B5)
Bicycling
(a) (b) 47.69% 49.90% 55.65% 56.65%
(100 watt)(B10)
Jumping(JP) 54.34% 55.49% 56.07% 58.38%
Fig. 3. The placements of sensor nodes and the triaxial accelerometer data of total 72.73% 75.11% 78.41% 79.70%
“walking” from ankle and waist;(a) the placements of sensor nodes;(b) The
triaxial accelerometer data of “walking” from ankle and waist

TABLE IV
1
Accerleration amplitude

ACCURACY OF CLASSIFICATION FOR INDIVIDUAL ACTIVITIES FROM LEFT

ANKLE AND WAIST USING SVM
0
Waist

the left ankle the waist

−1 Activity
FS FS+proposed FS FS+proposed
−2
Sitting(ST) 64.89% 70.61% 100% 100%
0 200 400 600 800 1000 Lying(LY) 98.78% 98.78% 99.18% 99.18%
Accerleration samples
10 standing up
Accerleration amplitude

x−axis 0.41% 23.77% 71.72% 72.13%

from lying(SFL)
5 y−axis
z−axis Standing(SD) 89.70% 89.70% 89.90% 90.10%
Ankle

0 Walking(W) 76.90% 78.48% 75.13% 75.13%

Running(R) 99.80% 99.80% 98.98% 98.99%
−5
Bicycling
65.33% 65.73% 61.52% 64.33%
−10 (50 watt)(B5)
0 200 400 600 800 1000
Accerleration samples Bicycling
65.99% 70.22% 56.25% 57.66%
(100 watt)(B10)
(a) (b) Jumping(JP) 97.11% 98.84% 92.49% 95.95%
total 74.98% 78.12% 80.32% 81.21%
Fig. 4. The placements of sensor nodes and the triaxial accelerometer data of
“walking” from ankle and waist;(a) the placements of sensor nodes;(b) The
triaxial accelerometer data of “walking” from ankle and waist
TABLE V
C ONFUSION MATRIX FOR ACTIVITY RECOGNITION ON THE DATA FROM
TABLE II WAIST USING SVM
T HE SAME ACQUISITION FEATURES OF DIFFERENT ACTIONS FROM THE
FOUR BODY PARTS ST LY SFL SD W R B5 B10 JP
ST 244 0 0 0 0 0 0 0 0
Actions Feature Samples LY 0 243 0 0 0 0 0 0 2
sitting 1223 SFL 0 0 175 61 6 0 2 0 0
lying 1227 SD 0 0 24 445 22 0 4 0 0
standing up from lying 1221 W 0 0 4 32 296 7 17 36 2
Standing 2473 R 0 0 0 0 2 488 0 0 3
Walking 1973 B5 0 0 0 1 16 0 307 175 0
running 2466 B10 0 0 0 1 39 0 177 279 0
Bicycling (50 watt) 2495 JP 0 0 0 0 1 12 0 0 160
Bicycling (100 watt) 2483
jumping 863

TABLE VI
C ONFUSION MATRIX FOR ACTIVITY RECOGNITION ON THE DATA FROM
WAIST USING SVM AND EEMD BASED FEATURE
B. Experimental Results
ST LY SFL SD W R B5 B10 JP
After extracting the features from the five subjects, clas- ST 245 0 0 0 0 0 0 0 0
sifiers of SVM and kNN have been utilized to evaluate the LY 0 243 0 0 0 0 0 0 2
SFL 0 0 176 57 8 0 3 0 0
proposed features and the proposed feature selection method. SD 0 0 18 446 24 0 4 2 1
Since data was acquired from five subjects, the leave-one- W 0 0 4 32 296 7 17 36 2
out (LOO) strategy is introduced to train and classify the R 0 0 0 0 2 488 0 0 3
activities.The validation was repeated five times with each B5 0 0 0 3 9 0 321 166 0
B10 0 0 0 3 28 0 179 286 0
union used exactly once for testing [51]. The final results are JP 0 0 0 0 0 7 0 0 166
their average values.

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2519679, IEEE Sensors
Journal
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2015 7

KNN SVM
80 TABLE VII
80
C ONFUSION MATRIX FOR ACTIVITY RECOGNITION ON THE DATA FROM
70
70
ANKLE USING SVM AND EEMD BASED FEATURES

60
60 ST LY SFL SD W R B5 B10 JP
FS ST 173 0 9 59 3 1 0 0 0
50 FS+EEMD
50
LY 0 242 0 0 2 0 1 0 0
40 40
SFL 0 0 58 180 6 0 0 0 0
SD 2 0 25 444 24 0 0 0 0
30 30 W 1 0 12 62 310 5 1 2 2
R 0 0 0 0 0 492 0 0 1
20 20
B5 0 0 0 0 0 0 328 171 0
10 10
B10 0 0 0 0 5 0 143 349 0
JP 0 0 0 0 0 2 0 0 171
0 0
ankle waist ankle waist

TABLE VIII
Fig. 5. The total accuracy of classification for activities from left ankle and C ONFUSION MATRIX FOR ACTIVITY RECOGNITION ON THE DATA FROM
waist using k-NN and SVM, respectively ANKLE USING SVM

ST LY SFL SD W R B5 B10 JP
ST 159 0 0 84 2 0 0 0 0
1) The influence of features with or without EEMD: So LY 0 242 0 0 2 0 1 0 0
far though lots of features have been extracted for human SFL 0 0 1 238 5 0 0 0 0
activity recognition [10], [12], [32], [51], features of identify- SD 2 0 25 444 24 0 0 0 0
W 1 0 0 72 303 12 2 3 1
ing activities are still an important research in human activity R 0 0 0 0 0 492 0 0 1
recognition. Fig. 5 gives the total accuracy of classification B5 0 0 0 0 1 0 326 172 0
for activities from waist and left ankle using k-NN and B10 0 0 0 0 1 0 168 328 0
SVM, respectively. From Fig. 5, it can be observed that, after JP 0 0 0 0 1 1 2 1 168
applying the features of EEMD, the total accuracy of classifiers
is significantly increased, which reflects the effectiveness of
the proposed features. Table III and Table IV present the Though the comparison of classifiers is not focused on in
accuracy of classification for individual activities from waist this paper, different accuracies have been obtained by the
and left ankle using SVM and k-NN, respectively. Observed classifiers of k-NN and SVM which are mechanically different
from Table III and Table IV, it can be found that after on the activity data of ”standing up from lying” acquired from
applying the features of EEMD, the accuracy of the activities left ankle.
”sitting” ,”standing up from lying”,”walking”, ”Bicycling (50 2) The influence of the positions : Observed from Fig. 5,
watt)”,Bicycling (100 watt)”and ”Jumping” from left ankle it can be seen that better results have been achieved when the
and the activities ”standing up from lying”,”standing”, ”Bi- sensor node is attached to the waist. Table. V- VIII show the
cycling (50 watt)”,Bicycling (100 watt)”and ”Jumping” from confusion matrix of activity classification with sensor nodes
waist are significantly increased, respectively. attached to different positions using SVM. From Table. III- VI,
That may be because that different activity can be de- it can be seen that the samples from waist is much easier to
composed into different IMFs with different energy, which distinct the actions ”sitting” and ”standing up from lying” than
increase the discrimination of the activities.Table. V- VIII from ankle, and actions ”walking”, ”running”,”Bicycling(100
show the confusion matrix of activity classification using watt)” and ”jumping” are easier to be recognized when ac-
SVM with dataset of EEMD based features or without EEMD celerometers are attached to the left ankle.
based features, which also demonstrate the effectiveness of 3) Feature selection and classification results: In order
the EEMD based features. Observed from Table. V- VI, to test the proposed method empirically, the previous used
after applying the EEMD based features, the discrimination features and the introduced features are integrated as a full
of activities is increased,such as activity ”standing up from feature set, and several experiments have been carried out to
lying” to” standing”, activity ”standing” to ”standing up from evaluate the proposed approach comparing with the original
lying”, activity ”Bicycling(50 watt)” to ”Bicycling (100 watt)” feature selection methods Relief and mRMR. Relief, which is
and ”walking”, activity ”Bicycling(100 watt)” to ”walking”and previously proved as one of the most successful filter feature
activity ”Jumping” to ”walking” and ”running”. From Table. selection method, is chosen as the baseline algorithm for
VII- VIII ,it can be seen that, after applying the EEMD based comparison. In addition to the feature selection approaches,
features,it is much easier to distinguish the activity ”sitting” SVM and k-NN with different machine learning approaches
from ”standing”, activity ”standing up from lying” from are employed to measure the effectiveness of a feature se-
”standing”, and activity ”walking”from ”standing”, ”running”, lection algorithm. For the sake of convenience, the selected
”Bicycling(50 watt)” and ”Bicycling (100 watt)” , activity features are ranged in a descending order. Fig. 6 shows the the
”Bicycling (50 watt)” from ”Bicycling (100 watt)” , activity comparison of the classification accuracy against the number
”Bicycling (100 watt)” from ”Bicycling (50watt)”, activity of selected features. In Fig. 6, the number x on the X-axis
”jumping” from ”walking”, ”Bicycling (100 watt)”, ”Bicycling refers to the top x features with the selected descending order
(50watt)”. using different approaches, and accordingly the performance

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2519679, IEEE Sensors
Journal
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2015 8

KNN for activities from waist KNN for activities from ankle
80 80 TABLE IX
T HE COST TIME ( SECONDS ) OF EACH FEATURE SELECTION ALGORITHM
70 70
ON THE DATASETS FROM WAIST AND ANKLE

accuracy(%)
accuracy(%)

60 60
GTFS Relief-F mRMR GTFS
50 mRMR 50
Relief−F
Waist 2.157 3.211 13.160
40 40 Ankle 2.754 3.467 13.056
0 10 20 30 40 50 0 10 20 30 40 50
the number of features the number of features
SVM for activities from waist SVM for activities from ankle
100 80

70
80
the impact of parameter σ on the recognition accuracy. Fig. 7
accuracy(%)

accuracy(%)
60
60
50
shows the recognition accuracy of SVM against different value
40
40 of σ using the proposed GTFS and mRMR is chosen as the
20
0 10 20 30 40 50 30
0 10 20 30 40 50
base method for comparison. The range of the parameter σ
the number of features the number of features
is set as{1, 3, 5}. It can be seen from Fig. 7 that when the
parameter is set to 3, the accuracy of SVM achieves its highest
Fig. 6. Accuracies vs different number of features
value. It can also be found from the Fig. 7 that the accuracy of
SVM does not drop down dramatically with the parameter σ
90 increasing and generally better than the performance of SVM
using mRMR algorism, which validate the effectiveness of the
80
proposed GTFS.
70 All experiments are conducted on a computer, with an Intel(R)
Core(TM) i5, 2.55GHZ and 2GB RAM. The algorithms are
accuracy (%)

60
all implemented in Matlab. Observed from Table IX, it can
50
be seen that the proposed method can select features with
40 acceptable computational complexity.
GTFS(w=1)
30 GTFS(w=3)
GTFS(w=5) V. C ONCLUSION
mRMR
20
0 10 20 30
the number of features
40 50 In order to reduce the influence of sensor placement,
EEMD-based features have been introduced and game theory
Fig. 7. classification performance of svm against different value of w on
based feature selection algorithm has been proposed for human
activities data from waist activity recognition in this paper. Classifiers of k-NN and
SVM have been utilized to evaluate the introduced features
and the proposed method. Comparing with the previously used
of classifiers is represented on Y -axis. The parmeter σ is set features, the effectiveness of the introduced features have been
as 3. validated for HAR. Comparing with Relief-F and mRMR in
Feature selection is to select a smaller number of optional the datasets, results show that better classification has been
features from the full huge feature set. However, too small obtained by the proposed method. In future, more subjects
number of features can greatly reduce the recognition accu- and more activities will be required to test the EEMD based
racy, which is not expected. If the number of the selected features and the proposed algorithm.
top features less than 9, the accuracies are generally less than
70%, as shown in Fig. 6. In fact, the accuracy can reach VI. ACKNOWLEDGMENT
76.42% by GTFS with 21 features, 76.44% by mRMR with This work was supported by National Natural Science Foun-
45 features and 75.68% by Relief-F with 47 features, which is dation of China under Grant No.61473058 and No.61174027,
also greatly smaller than the number of the original features and Fundamental Research Funds for the Central Universities
in the dataset of ankle using kNN. On the other hand, too (DUT15ZD114). The authors would also like to extend their
many features selected provide no meaning for classification sincere appreciation to the Deanship of Scientific Research
accuracy and computational cost. Therefore, the number of at King Saud University for its funding of this International
selected features should stay in an accepted range. It can be Research Group (IRG14-28).
observed from Fig. 6 that the proposed GTFS outperform
mRMR and Relief-F within an accepted range of selected R EFERENCES
features, which demonstrates that the effectiveness of the pro-
[1] A. Avci, S. Bosch, M. Marin-Perianu, R. Marin-Perianu, and P. Havinga,
posed GTFS for human activity recognition. Feature selection “Activity recognition using inertial sensing for healthcare, wellbeing and
is aimed at extracting a smaller number of feature subset for sports applications: A survey,” in Architecture of computing systems
recognition. In the algorithm, the parameter σ as the size of (ARCS), 2010 23rd international conference on. VDE, 2010, pp. 1–10.
[2] P. Rashidi and A. Mihailidis, “A survey on ambient-assisted living tools
coalition set has been utilized to reduce the computational for older adults,” IEEE J Biomed Health Inform, vol. 17, no. 3, pp.
complexity. In fact, the size of the interdependent groups is 579–90, 2013.
generally not a constant according to various realistic human [3] Z. L. Wang, H. Y. Zhao, S. Qiu, and Q. Gao, “Stance phase detection
for zupt-aided foot-mounted pedestrian navigation system,” MECHA-
activity datasets or even within the same dataset. To better TRONICS, IEEE/ASME Transactions on, vol. 20, no. 6, pp. 4170–3181,
illustrate the algorithm, SVM has been utilized to evaluate 2015.

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2519679, IEEE Sensors
Journal
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2015 9

[4] N. K. Suryadevara and S. C. Mukhopadhyay, “Wireless sensor network [27] H. Peng and C. Ding, “Minimum redundancy and maximum relevance
based home monitoring system for wellness determination of elderly,” feature selection and recent advances in cancer classification,” Feature
Sensors Journal, IEEE, vol. 12, no. 6, pp. 1965–1972, 2012. Selection for Data Mining, p. 52, 2005.
[5] U. Maurer, A. Smailagic, D. P. Siewiorek, and M. Deisher, “Activity [28] H. Liu and H. Motoda, Computational methods of feature selection.
recognition and monitoring using multiple sensors on different body CRC Press, 2007.
positions,” in Wearable and Implantable Body Sensor Networks, 2006. [29] X. Jin, E. W. Ma, L. L. Cheng, and M. Pecht, “Health monitoring
BSN 2006. International Workshop on. IEEE, 2006, pp. 113–116. of cooling fans based on mahalanobis distance with mrmr feature
[6] M. Ermes, J. Parkka, J. Mantyjarvi, and I. Korhonen, “Detection of daily selection,” Instrumentation and Measurement, IEEE Transactions on,
activities and sports with wearable sensors in controlled and uncontrolled vol. 61, no. 8, pp. 2222–2229, 2012.
conditions,” Information Technology in Biomedicine, IEEE Transactions [30] M. Jiang, H. Shang, Z. Wang, H. Li, and Y. Wang, “A method to deal
on, vol. 12, no. 1, pp. 20–26, 2008. with installation errors of wearable accelerometers for human activity
[7] H. Z. Sen Qiu, Zhelong Wang and H. Hu, “Using distributed wear- recognition,” Physiological measurement, vol. 32, no. 3, p. 347, 2011.
able sensors to measure and evaluate human lower limb motions,” [31] U. Kurokawa, B. I. Choi, and C.-C. Chang, “Filter-based miniature
Instrumentation and Measurement, IEEE Transactions on, In Press. spectrometers: spectrum reconstruction using adaptive regularization,”
DOI:10.1109/TIM.2015.2504078, 2016. Sensors Journal, IEEE, vol. 11, no. 7, pp. 1556–1563, 2011.
[8] S. C. Mukhopadhyay, “Wearable sensors for human activity monitoring: [32] M. Zhang and A. A. Sawchuk, “Human daily activity recognition with
A review,” Sensors Journal, IEEE, vol. 15, no. 3, pp. 1321–1330, 2015. sparse representation using wearable sensors,” Biomedical and Health
[9] Z. Zhang and S. Poslad, “Improved use of foot force sensors and mobile Informatics, IEEE Journal of, vol. 17, no. 3, pp. 553–560, 2013.
phone gps for mobility activity recognition,” Sensors Journal, IEEE, [33] Z. Wang, C. Zhao, and S. Qiu, “A system of human vital signs
vol. 14, no. 12, pp. 4340–4347, 2014. monitoring and activity recognition based on body sensor network,”
[10] L. Bao and S. S. Intille, “Activity recognition from user-annotated Sensor Review, vol. 34, no. 1, pp. 42–50, 2014.
acceleration data,” in Pervasive computing. Springer, 2004, pp. 1–17. [34] Z. Wu and N. E. Huang, “Ensemble empirical mode decomposition:
[11] C. Zhu and W. Sheng, “Human daily activity recognition in robot- a noise-assisted data analysis method,” Advances in adaptive data
assisted living using multi-sensor fusion,” in Robotics and Automation, analysis, vol. 1, no. 01, pp. 1–41, 2009.
2009. ICRA’09. IEEE International Conference on. IEEE, 2009, pp. [35] Y.-H. Wang, C.-H. Yeh, H.-W. V. Young, K. Hu, and M.-T. Lo, “On
2154–2159. the computational complexity of the empirical mode decomposition
[12] A. M. Khan, Y.-K. Lee, S. Y. Lee, and T.-S. Kim, “A triaxial algorithm,” Physica A: Statistical Mechanics and its Applications, vol.
accelerometer-based physical-activity recognition via augmented-signal 400, pp. 159–167, 2014.
features and a hierarchical recognizer,” Information Technology in [36] H. Wang, J. Chen, and G. Dong, “Feature extraction of rolling bearing’s
Biomedicine, IEEE Transactions on, vol. 14, no. 5, pp. 1166–1172, 2010. early weak fault based on eemd and tunable q-factor wavelet transform,”
[13] P. Gupta and T. Dallas, “Feature selection and activity recognition Mechanical Systems and Signal Processing, vol. 48, no. 1, pp. 103–119,
system using a single triaxial accelerometer,” Biomedical Engineering, 2014.
IEEE Transactions on, vol. 61, no. 6, pp. 1780–1786, 2014. [37] B. Liu, S. Riemenschneider, and Y. Xu, “Gearbox fault diagnosis
using empirical mode decomposition and hilbert spectrum,” Mechanical
[14] Z. W. Donghui Wu and et al, “Mixed-kernel based weighted ex-
Systems and Signal Processing, vol. 20, no. 3, pp. 718–734, 2006.
treme learning machine for inertial sensor based human activity
recognition with imbalanced dataset,” Neurocomputing, In Press. [38] C. Junsheng, Y. Dejie, and Y. Yu, “The application of energy operator
DOI:10.1016/j.neucom.2015.11.095, 2016. demodulation approach based on emd in machinery fault diagnosis,”
Mechanical Systems and Signal Processing, vol. 21, no. 2, pp. 668–
[15] A. M. Khan, Y.-K. Lee, S. Lee, and T.-S. Kim, “Accelerometers
677, 2007.
position independent physical activity recognition system for long-term
[39] J. Cheng, D. Yu, J. Tang, and Y. Yang, “Application of frequency
activity monitoring in the elderly,” Medical & biological engineering &
family separation method based upon emd and local hilbert energy
computing, vol. 48, no. 12, pp. 1271–1279, 2010.
spectrum method to gear fault diagnosis,” Mechanism and Machine
[16] M. Shoaib, S. Bosch, O. D. Incel, H. Scholten, and P. J. Havinga, “A Theory, vol. 43, no. 6, pp. 712–723, 2008.
survey of online activity recognition using mobile phones,” Sensors,
[40] L. Du, B. Wang, Y. Li, and H. Liu, “Robust classification scheme for
vol. 15, no. 1, pp. 2059–2085, 2015.
airplane targets with low resolution radar based on emd-clean feature
[17] C. Catal, S. Tufekci, E. Pirmit, and G. Kocabag, “On the use of ensemble extraction method,” Sensors Journal, IEEE, vol. 13, no. 12, pp. 4648–
of classifiers for accelerometer-based activity recognition,” Applied Soft 4662, 2013.
Computing, 2015. [41] J. A. Thomas and T. Cover, Elements of information theory. Wiley
[18] W.-C. Cheng and D.-M. Jhan, “Triaxial accelerometer-based fall detec- New York, 2006, vol. 2.
tion method using a self-constructing cascade-adaboost-svm classifier,” [42] C. E. Shannon, “A mathematical theory of communication,” ACM
Biomedical and Health Informatics, IEEE Journal of, vol. 17, no. 2, pp. SIGMOBILE Mobile Computing and Communications Review, vol. 5,
411–419, 2013. no. 1, pp. 3–55, 2001.
[19] Ø. D. Trier, A. K. Jain, and T. Taxt, “Feature extraction methods for [43] Y. Li and B. Bhanu, “Utility-based camera assignment in a video
character recognition-a survey,” Pattern recognition, vol. 29, no. 4, pp. network: A game theoretic framework,” Sensors Journal, IEEE, vol. 11,
641–662, 1996. no. 3, pp. 676–687, 2011.
[20] J. Nishimura and T. Kuroda, “Versatile recognition using haar-like [44] L. D. H. Sampaio, A. R. C. E. Souza, T. Abrao, and P. J. E. Jeszen-
feature and cascaded classifier,” Sensors Journal, IEEE, vol. 10, no. 5, sky, “Game theoretic energy efficiency design in mc-cdma cooperative
pp. 942–951, 2010. networks,” Sensors Journal, IEEE, vol. 14, no. 9, pp. 3065–3075, 2014.
[21] J. C. Ang, H. Haron, and H. N. A. Hamed, “Semi-supervised svm- [45] M. Abbasi and N. Fisal, “Noncooperative game-based energy welfare
based feature selection for cancer classification using microarray gene topology control for wireless sensor networks,” Sensors Journal, IEEE,
expression data,” in Current Approaches in Applied Artificial Intelli- vol. 15, no. 4, pp. 2344–2355, 2015.
gence. Springer, 2015, pp. 468–477. [46] H.-Y. Shi, W.-L. Wang, N.-M. Kwok, and S.-Y. Chen, “Game theory
[22] P. Ghamisi and J. A. Benediktsson, “Feature selection based on hy- for wireless sensor networks: a survey,” Sensors, vol. 12, no. 7, pp.
bridization of genetic algorithm and particle swarm optimization,” 9055–9097, 2012.
Geoscience and Remote Sensing Letters, IEEE, vol. 12, no. 2, pp. 309– [47] D. A. Bell and H. Wang, “A formalism for relevance and its application
313, 2015. in feature subset selection,” Machine learning, vol. 41, no. 2, pp. 175–
[23] A. Wang, N. An, G. Chen, L. Li, and G. Alterovitz, “Accelerating 195, 2000.
wrapper-based feature selection with k-nearest-neighbor,” Knowledge- [48] L. Yu and H. Liu, “Efficient feature selection via analysis of relevance
Based Systems, vol. 83, pp. 81–91, 2015. and redundancy,” The Journal of Machine Learning Research, vol. 5,
[24] J. R. Vergara and P. A. Estévez, “A review of feature selection methods pp. 1205–1224, 2004.
based on mutual information,” Neural Computing and Applications, [49] G. Brown, A. Pocock, M.-J. Zhao, and M. Luján, “Conditional likelihood
vol. 24, no. 1, pp. 175–186, 2014. maximisation: a unifying framework for information theoretic feature
[25] R. Kohavi and G. H. John, “Wrappers for feature subset selection,” selection,” The Journal of Machine Learning Research, vol. 13, no. 1,
Artificial intelligence, vol. 97, no. 1, pp. 273–324, 1997. pp. 27–66, 2012.
[26] T. N. Lal, O. Chapelle, J. Weston, and A. Elisseeff, “Embedded [50] J. Eatwell, M. Milgate, and P. Newman, The New Palgrave: Game
methods,” in Feature extraction. Springer, 2006, pp. 137–165. Theory. Macmillan Reference Books, 1989.

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2519679, IEEE Sensors
Journal
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2015 10

[51] Z. Wang, M. Jiang, Y. Hu, and H. Li, “An incremental learning method M. Anwar Hossain M. Anwar Hossain is an As-
based on probabilistic neural networks and adjustable fuzzy clustering sociate Professor in the Department of Software
for human activity recognition by using wearable sensors,” Information Engineering, College of Computer and Information
Technology in Biomedicine, IEEE Transactions on, vol. 16, no. 4, pp. Sciences (CCIS) at King Saud University (KSU),
691–699, 2012. Riyadh. He completed his Master and Ph.D. in
Electrical and Computer Engineering from the Uni-
versity of Ottawa, Canada. At KSU, Dr. Hossain
received IBM faculty award. He has co-organized
more than ten IEEE/ACM workshops and served as
guest editors of some reputed journals. His current
Zhelong Wang Zhelong Wang received the B.S. research interests include multimedia cloud, multi-
and M.S.degrees in automatic control from Dalian media surveillance and privacy, Internet of Things, smart cities and ambient
University of Technology, Dalian, China, in 1996 intelligence.
and 1999, respectively, and the Ph.D. degree in
robotics from University of Durham, U.K. in 2003.
In 2004, he joined School of Electronic and Infor-
mation Engineering, Dalian University of Technol-
ogy, where he is currently a Professor and a Ph.D.
supervisor. He has received a research award from
the Audi Foundation, first prize of Liaoning Province
Natural Science Award (Academic Paper Program)
in 2008 and 2009 and Dalian Excellent IT Teacher Award in 2008. He has been
elected into Liaoning Province Hundred, Thousand and Ten Thousand Talent
Project in 2009. His research interests include robotics, intelligent control,
mechatronics, body sensor networks, and machine learning.

Donghui Wu Donghui Wu received the B.S. and

M.S.degrees in automatic control from North China
University of Water Conservancy and Hydropower,
Zhengzhou, China, in 2006 and 2009, respectively.
He is currently a Ph.D. student in control theory
and control engineering with the School of Control
Science and Engineering, Dalian University of Tech-
nology, Dalian, China. His research interests include
body sensor networks, bioinformatics, pattern recog-
nition and machine learning.

Jianming Chen Jianming Chen received the

B.S.degrees in Zhejiang University, China, in 1985.
In 2001, he joined School of North China University
of Water Conservancy and Hydropower, Zhengzhou,
China, where he is currently a Professor. His re-
search interests include human activity recognition,
wireless sensor networks, and intelligent control.

Ahmed Ghoneim Ahmed Ghoneim received his

M.Sc. degree in Software Modeling from University
of Menoufia, Egypt, and the Ph.D. degree from
the University of Magdeburg (Germany) in the
area of software engineering, in 1999 and 2007
respectively. He is currently an assistant professor at
the department of software engineering, king Saud
University. His research activities address software
evolution; service oriented engineering, software de-
velopment methodologies, Quality of Services, Net-
Centric Computing, and Human Computer Interac-
tion (HCI)

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.