Centralized Class Specific Dictionary Learning for Wearable Sensors based physical activity recognition

Centralized Class Specific Dictionary Learning for
wearable sensors based physical activity recognition
Sherin M Mathews
Department of Electrical
and Computer Engineering
University of Delaware
Newark, Delaware
sherinm@udel.edu
Chandra Kambhamettu
Department of Computer
and Information Sciences
Newark, Delaware
chandrak@udel.edu
Kenneth E. Barner
Department of Electrical
and Computer Engineering
Newark, Delaware
barner@udel.edu
Abstract—With recent progress in pervasive healthcare,
physical activity recognition with wearable body sensors has
become an important and challenging area in both research and
industrial communities. Here, we address a novel technique for
a sensor platform that performs physical activity recognition by
leveraging a class specific regularizer term into the dictionary
pair learning objective function. The proposed algorithm jointly
learns a synthesis dictionary and an analysis dictionary in
order to simultaneously perform signal representation and
classification once the time-domain features have been extracted.
Specifically, the class specific regularizer term ensures that the
sparse codes belonging to the same class will be concentrated
thereby proving beneficial for the classification stage. In order
to develop a more practical approach, we employ a combination
of an alternating direction method of multipliers and a l1 − ls
minimization method to approximately minimize the objective
function. We validate the effectiveness of our proposed model
by employing it on two activity recognition problem and an
intensity estimation problem, both of which include a large
number of physical activities. Experimental results demonstrate
that classifiers built in this dictionary learning based framework
outperforms state of art algorithms by using simple features,
thereby achieving competitive results when compared with
classical systems built upon features with prior knowledge.
Keywords— Body sensor networks, Wearable Sensors, Hu-
man Activity Recognition, Physical Activity Monitoring, In-
tensity Estimation of Physical Activity, Inertial sensor, Heart
Rate, Sensor localization, Sparse Representation, Dictionary
Pair Learning
I. INTRODUCTION AND RELATED WORK
The classification of basic physical activities (walk, run,
cycle) and basic postures (lie, stand, sit) is well researched
[1], [2], [3], [4], and good recognition performance can
be obtained using a single 3D-accelerometer and simple
classifiers. Moreover, recent studies have focused on
estimating the intensity of an activity (i.e., light, moderate or
vigorous) ( e.g., in [5], [6]) by means of metabolic equivalent
(MET), a parameter that attributes to the energy expenditure
of a physical activity.
Over the years, many studies have analyzed both simple
and complex human activities reported by wearable sensors.
A large number of these have focused on determining which
features in this activity data are the most significant, and
how these data can be most effectively employed to classify
the activities [3], [4], [7], [8]. Other research has focused on
which computational model is the most relevant to represent
human activity data [9], [10], [11]. Despite such research
efforts, the scalability of handling large intra-class variations
and the robustness of many human-activity recognition
techniques to the model parameters remains limited.
To make physical activity monitoring practicable in
everyday life scenarios, an activity-recognition framework
must be able to manage a wide range of everyday, household,
or sport activities and must handle a variety of potential
users. Although current research has focused on increasing
the number of activities that are recognized, each increase in
the number of activities causes the classification performance
to fall off. Secondly, recording and using only a small set of
a few activities in basic activity recognition, without having
simultaneous background activities, limits the applicability of
the developed algorithms. Recently, infilling of missing pixels,
and image and speech classification problems have been
successfully addressed by dictionary learning particularly
because of its robustness towards missing data and noise. Real-
time scenarios might include activity switching, thus requires
testing on a wider range of activities than were used for
training [12]. Thus, successful activity-recognition requires
dictionary learning frameworks [12] that are sufficiently
robust for classification even with limited training data.
In this paper, we present a unified dictionary pair learning
framework by incorporating a centralized class specific
regularizer term to solve the wearable sensor based
classification problem. In the dictionary pair learning
framework, the training samples of each class contribute
identically to the dictionary, thus generating a dictionary
having training samples corresponding to each class. This
can result in instability and high residual error, which is
detrimental to the recognition performance. Using traditional
class specific dictionary learning approaches [35] does
allow learning a dictionary for each class, but might
result in interdependence in sparse codes and erroneous
978-1-5090-4780-2/17/$31.00 ©2017 IEEE

discrimination. Thus, the main contribution is to explicitly
incorporate centralized class specific sparse codes to the
dictionary pair learning objective function to obtain superior
classification performance by making sparse codes in the same
class concentrated. A new alternate minimization algorithm
incorporated with a l1 − ls minimization method is developed
to facilitate convergence of the non-convex objective function.
To the best of our knowledge, dictionary learning frameworks,
and specifically centralized class dependent dictionary pair
learning frameworks, have not to date been used in wearable
sensor-based applications. Consequently, our novel dictionary
learning-based framework algorithm will instigate future
research on this method’s potential applicability for accurate
sensor-based data classifications and for other physiological-
signal classifications.
II. PROPOSED METHODOLOGY
The proposed framework involves two steps: data process-
ing and recognition. Data processing incorporates preprocess-
ing, segmentation and feature extraction stages. In the prepro-
cessing stage, raw sensory data is synchronized, timestamped,
and labeled; 3D-acceleration and heart rate data are acquired.
During segmentation, this collected data is segmented with a
sliding window, using a window size of 512 samples. During
the feature extraction stage, signal features extracted from
the segmented 3D-acceleration data are calculated for each
of the three axis separately and for the 3 axes together.
Inclusion of Heart Rate (HR) monitor with the commonly
used inertial sensors proved specifically useful for physical
activity intensity estimation. Mean and gradient are calculated
on both the raw and normalized heart rate signals from the HR
data. Overall, a total of 137 basic features are computed : 133
features from Inertial Measurement Unit (IMU) acceleration
data and 4 features from HR data.
A. Dictionary Pair Learning Framework
Dictionary Pair Learning classification algorithm jointly
learns a synthesis dictionary and an analysis dictionary to
attain the objective of signal representation and discrimination
[13]. To explain discriminative dictionary learning, a set of p-
dimensional training samples from K classes can be defined
as X = [X1, ........, Xk, ...., XK], where Xk ∈ Rp×n
is the
training sample set of class k, and n is the number of samples
of each class. Discriminative dictionary learning (DL) methods
intents to learn an efficient data representation model from X
for classification by employing the class label information of
training data, and can be formulated as:
min
D,A
X − DA 2
F + λ||A||p + Ψ(D, A, Y) (1)
Here λ 0 is a scalar constant; Y denotes the class label
matrix of samples in X; D is the synthesis dictionary to be
learned; and A is the coding coefficient matrix of X over
D. The data fidelity term X − DA 2
F in the training model
(equation (1)), ensures the representation ability of D; ||A||p
is the lp-norm regularizer on A; and Ψ(D, A, Y) represents
discrimination promotion function that ensures discrimination
power of D and A [13]. Using an analysis dictionary defined
by P ∈ RmK×p
, the code A can be analytically represented
as A = PX, thereby making the representation of X efficient.
Based on this concept, the DPL model learns an analysis
dictionary P together with the synthesis dictionary D, leading
to the following DPL model [13]:
P∗
, D∗
= arg min
P,D
X − DPX 2
F + Ψ(D, P, X, Y) (2)
Here Ψ (D, P, X, Y) represents the discrimination function;
D and P form a dictionary pair where the analysis dictionary
P is used to analytically code X, and the synthesis dictionary
D is used to reconstruct X. The learned structured synthesis
dictionary D = [D1, ..Dk, ., DK] and the structured analysis
dictionary P = [P1, ....., Pk......PK ] form a sub-dictionary pair
corresponding to class k. Thus incorporating the structured
analysis dictionary P, we ensure that the sub-dictionary Pk
projects the samples from class i,i = k to a nearly null
space thereby making the coefficient matrix PX nearly block
diagonal. By adopting variable matrix A to relax the non-
convex problem, the following DPL model can be defined as:
P∗
, A∗
, D∗
= arg min
P,A,D
K
k=1
Xk − DkAk
2
F
+τ PkXk − Ak
2
F + λ||Pk, Xi||2
F
(3)
Here Xk represents the complementary data matrix of Xk in
the training set X, λ > 0 denotes a scalar constant, and di is
the ith
atom of synthesis dictionary D.
B. Dictionary Pair Learning model based on centralized class
specific regularizer term
The objective is to incorporate a centralized class specific
regularizer term [14] to obtain a unified dictionary pair
learning framework. In DPL model, training samples of each
class contribute equivalently to the dictionary, thus generating
a dictionary consisting of training samples in corresponding
class, resulting in instability and higher residual error. The
main contribution here is to explicitly incorporate centralized
class specific sparse codes [21] to the dictionary pair learning
objective function, thereby, making sparse codes in the same
class concentrated.
To attain this objective,we denote the mean of each row of
sparse code A as E(A). The regularizer term can be formulated
as :
R(Ak) = η
N
n=1
(Ak)·n − E(Ak) 2
2 (4)
where η is the tradeoff parameter between the reconstruction
error and the degree of deviation from the sparse code to their
centers and A·n represent the nth column of a sparse matrix
A. Incorporating the class specific regularizer term to the DPL
objective function, equation (3) can be formulated as:

P∗
, A∗
, D∗
= arg min
P,A,D
K
k=1
Xk − DkAk
2
F + τ PkXk − Ak
2
F
+λ||Pk, Xi||2
F + η
N
n=1
(Ak)·n − E(Ak) 2
2
(5)
In order to solve this equation, we integrate alternating
direction method of multipliers with l1 − ls method to
facilitate convergence. The alternating direction method of
multipliers (ADMM) solve convex optimization problems
by fixing some variables and solving for the other variable,
thereby decomposing the problem into smaller sub-problems
making each of them easier to handle. The minimization can
be alternated between the following steps :
1:Fix A, update P and D
P∗
= arg min
P
K
k=1
τ PkXk − Ak
2
F + λ||Pk, Xi||2
F (6)
D∗
= arg min
D
K
k=1
Xk − DkAk
2
F (7)
The closed form solution for P can be obtained as:
P∗
= τAkXT
k (τXkXT
k + λ ¯Xk
¯XT
k + Y I)−1
(8)
The closed form solution for D can be obtained by intro-
ducing a variable S as in [13]
Dr+1
= arg min
D
K
k=1
Xk − DkAk
2
F + ρ Dk − Sr
k + Tr
k
2
F
(9)
The solution for D and P are the same as in the DPL
framework [13] using ADMM algorithm.
2. Fix D and P, update A
A∗
= arg min
A
K
k=1
tr[(Xk − DkAk)⊤
(Xk − DkAk)]
+tr[τ(PkXk − Ak)⊤
(PkXk − Ak)]
+η
N
n=1
(Ak)·n − E(Ak) 2
2
(10)
Here, tr represents the trace of a matrix. We use l1 − ls
optimization method to solve for A. Ignoring the constant
terms, the function can be simplified using l1 −ls optimization
method.
A∗
= arg min
A
K
k=1
tr[−2(Xk
⊤
Dk + τXk
⊤
Pk
⊤
)Ak
+Ak
⊤
(Dk
⊤
Dk + τI)Ak] + η
N
n=1
(Ak)·n − E(Ak) 2
2
(11)
Defining t = (Xk
⊤
Dk + τXk
⊤
Pk
⊤
) and
S = (Dk
⊤
Dk + τI) and using property tr(A⊤
BA) =
N
n=1 A⊤
·nB(A·n), equation 11 can be rewritten as
A∗
= −2
N
n=1
tn·(Ak)·n +
N
n=1
(Ak)⊤
·nS(Ak)·n+
η
N
n=1
[(
N − 1
N
)2
(Ak)⊤
·n(Ak)·n − 2
N − 1
N2
(Ak)⊤
·n
N
m=1,m=n
(Ak)·m]
(12)
The objective function now reduces to equation (13)
f(Ak)ln = (Ak)2
ln(Sll + η[
N − 1
N
]2
) + 2(Ak)ln(
R
q=1q=l
Slq(Ak)qn)
−2η(Ak)ln[
N − 1
N2
N
m=1,m=n
(Ak)mn] − 2(Ak)lntT
ln
(13)
Here (Ak)ln denotes all elements in Ak except the element in
the ‘l’ th row and the ‘n’ th column. Defining
dln = tT
ln + η[
N − 1
N2
N
m=1,m=n
(Ak)lm] (14)
f(Ak)ln becomes a piece-wise parabolic function as in
[21]. Adopting the convexity and monotonic property of the
parabolic function and the problem transformation defined in
[21] , we have the closed-form solution as function reaches
the minimum at the unique point.
(Ak)ln =
dln − [S( Âk
ln
)]ln
Sll + η[N−1
N ]2
(15)
In the testing phase, the analysis sub-dictionary Pk is
trained to produce small coefficients for samples from classes
other than k, and thus can only generate significant coding
coefficients for samples from class k. Meanwhile, the synthesis
sub-dictionary Dk is trained to reconstruct the samples of
class k from their projective coefficients PkXk, i.e residual
will be small. Conversely, since PkXi will be small and
Dk is not trained to reconstruct Xi, the residual ai will be
much larger. Thus, if the query sample y is from class k, its
projective coding vector Pk will more likely be large, while its
projective coding vectors Pi will be small. Therefore, class-
specific reconstruction residual are used to identify the class
label of testing sample.
III. EXPERIMENTAL RESULTS AND DISCUSSION
A. PAMAP2 Database
The proposed algorithm is evaluated over the activity
recognition and intensity estimation classification problems
defined on the recently released PAMAP2 Physical Activity
Monitoring Data Set. Briefly, this dataset captures 18 physical
activities performed by 9 subjects wearing 3 IMUs (Inertial
measurement unit) and a HR (heart rate) monitor. Here, the
activity classification task consists of 15 different activity
classes represented as lie, sit, stand, walk, run, cycle, Nordic
walk, drive car, ascend, descend stairs, vacuum, iron, fold

Annotated Class Class 1 Class 2 Class 3 Class 4 Class 5 Class 6 Class 7 Class 8 Class 9 Class 10 Class 11 Class 12 Class 13 Class 14 Class 15
Predicted Class 1 1660 2 3 19 12 0 0 0 1 0 0 0 30 0 1
Predicted Class 2 1 1257 134 73 18 4 0 4 0 3 1 0 130 25 5
Predicted Class 3 0 83 1262 126 33 1 4 0 6 2 2 0 156 20 8
Predicted Class 4 1 29 128 1379 103 0 0 0 0 0 3 0 506 40 0
Predicted Class 5 0 0 10 170 693 5 0 5 0 2 8 0 634 11 19
Predicted Class 6 0 0 0 0 0 409 49 199 101 0 0 0 15 0 6
Predicted Class 7 0 0 0 0 0 73 358 49 102 7 0 1 41 0 0
Predicted Class 8 0 0 0 0 0 77 35 1825 224 28 0 0 4 0 0
Predicted Class 9 0 0 0 0 0 4 21 111 1568 0 0 0 2 0 3
Predicted Class 10 0 4 1 3 10 8 7 25 78 1312 0 0 19 0 5
Predicted Class 11 0 0 0 0 14 0 0 0 6 0 791 0 18 0 6
Predicted Class 12 0 0 0 0 30 0 33 0 35 0 0 2668 2 0 0
Predicted Class 13 0 9 43 208 350 12 6 1 9 10 2 0 1056 34 8
Predicted Class 14 0 1 21 322 145 0 0 0 0 0 2 0 353 57 0
Predicted Class 15 0 0 0 0 18 43 2 16 46 0 8 3 17 0 242
TABLE I: Confusion matrix using proposed framework on PAMAP2 -AR dataset.
laundry, clean house, play soccer and jump rope. This
classification task is referred to as the PAMAP2 Activity
Recognition (PAMAP2-AR) task. The intensity-estimation
classification task aims to distinguish activities of light,
moderate, and vigorous effort based on the MET of the
various physical activities, as provided by [15] and is referred
to as the PAMAP2 Intensity Estimation (PAMAP2 - IE)
task. Therefore, intensity classes are defined as activities of
light effort (< 3.0 METs) (lie, sit, stand, drive a car, iron,
fold laundry, clean house, watch TV, work at a computer),
moderate effort (3.0-6.0 METs) (walk, cycle, descend stairs,
vacuum and Nordic walk), or vigorous effort (> 6.0 METs)
(run, ascend stairs, jump rope, play soccer).
Using these two defined classification tasks, the proposed
method is compared with C4.5 Decision Tree [16] and
Adaboost algorithms [16]. For the evaluation procedure,
we randomly selected 75% of the data for training and
25% for testing and averaged recognition results over 10
repetitions. Aside from providing classification accuracy, we
also report on the average computation time for classifying
one query activity of competing algorithms in the experiments.
1) PAMAP2 -Results on Intensity Estimation Task: First
presented are classification results over a intensity estimation
task defined on PAMAP2 datset. In [16], the C4.5 decision tree
algorithm and Adaboost classifier were tested on the PAMAP2
dataset and demonstrated an accuracy of 70.07% and 73.93%.
The overall confusion matrix using the proposed framework
for the three intensity estimation tasks (i.e., light (Class 1),
moderate (Class 2) and vigorous (Class 3) tasks) is given
in Table II. An independent performance assessment of the
proposed framework results in an accuracy of 85.89% on the
AR-IE task, demonstrating that our framework outperforms
the C4.5 decision tree and AdaBoost classifiers (Table III).
As shown in Table III, the proposed framework achieves the
highest recognition rate of 85.89%. As shown in Table IV, the
proposed algorithm is approximately more than 5 times faster
than Adaboost. Clearly, the proposed approach outperforms
other methods.
Annotated activity Class 1 Class 2 Class 3
Predicted Class 1 11655 1338 331
TABLE II: Overall confusion matrix using dictionary pair learning framework on
PAMAP2-IE dataset. The table shows how different annotated activities are classified
into different classes.
Method Proposed Adaboost C4.5 Decision Tree
Accuracy 85.89% 73.93% 70.07%
TABLE III: Recognition results on PAMAP2-IE Dataset.
Computation time 0.24s 10.82s 2.39s
TABLE IV: Computation time for classifying one query activity on PAMAP2-IE Dataset.
2) PAMAP2 - Results on Activity Recognition Task:
To investigate the proposed framework’s performance on
PAMAP2− AR task, we computed its performance on 15
classes based activity-recognition tasks. We find that our
framework outperforms the C4.5 decision tree with an accu-
racy of 73.17% versus 71.59%, and it gives competitive results
when compared to an AdaBoost classifier on the PAMAP2 AR
task (Table V).
Accuracy 73.17% 71.78% 71.59%
TABLE V: Recognition results on PAMAP2-AR Dataset.
Computation time 1.52s 11.25s 9.39s
TABLE VI: Computation time for classifying one query activity on PAMAP2-AR
Dataset.
In addition to competitive accuracy, our framework provides
the additional advantages of lower computation time for clas-
sifying a query activity(Table VI). Further examination of our
results indicate that averaged over 10 test runs, the confusion
matrix of the best-performing classifier on the PAMAP2-
AR task yields an overall accuracy of 73.17%, showing that
some activities are recognized with high accuracy, such as
lying, walking, or even distinguishing between ascending
and descending. Overall, misclassifications, where activities
belonging to one class are mistakenly classified as belonging
to its neighboring classes, are lower. In [16] and [17], the eval-
uation technique was leave-one-activity-out (LOAO) where an

Reference/Algorithm Class 1 Class 2 Class 3 Class 4 Class 5 Class 6
Class 1 1877 29 0 0 0 0
Class 2 572 1190 15 0 0 0
Class 3 4 17 1918 0 0 5
Class 4 1 2 1 1588 51 79
Class 5 0 8 0 10 1320 68
Class 6 0 12 0 18 19 1495
TABLE VII: Overall confusion matrix using proposed framework on SBHAR dataset.
The table shows how different annotated activities are classified into different classes.
Method Proposed Multi-Class Naives Decision
SVM [18] Bayes [19] Tree [19]
Accuracy 91.31% 89.0% 82.5% 86.8%
TABLE VIII: Recognition results on SBHAR Dataset.
activity monitoring system is used on a previously unknown
activity. Our framework also takes in data randomly.
Thus evaluation is based on a completely unknown activity
from an unknown user, while training is performed using a
different activity with a different user in our random 75%-
25% validation approach. These types of subject-independent
and activity-independent validation techniques are preferred
for physical activity monitoring since they provide results with
more practical meaning. Using our framework, we can not only
achieve good classifier performance but also eliminate the need
of pre-training a particular activity for classification. Thus, our
proposed method makes it possible to design a robust physical
activity monitoring system that has the desired generalization
characteristics.
B. SBHAR
This is a dataset built from the recordings of 30 subjects
performing activities of daily living (ADL) while carrying a
waist-mounted smartphone with embedded inertial sensors.
This dataset includes six activities (walking, walking upstairs,
walking downstairs, sitting, standing and laying) performed
wearing a smartphone (Samsung Galaxy S II) on the waist.
The 3-axial linear acceleration and 3-axial angular velocity
at a constant rate of 50Hz was captured using the phone’s
embedded accelerometer and gyroscope. The sensor signals
are pre-processed by applying noise filters and then sampled
in fixed-width sliding windows of 2.56 sec and 50% overlap
(128 readings/window). From each window, a vector of
features was obtained by calculating variables from the time
and frequency domain (e.g. mean, standard deviation, signal
magnitude area, entropy, signal-pair correlation, etc.)
We evaluate our proposed approach and compare with
Naive’s Bayes Classifier [18], Decision Tree [18], Multiclass
SVM [18] and other state-of-the-art approaches. Our
approaches outperform the state-of-the-art approaches as
illustrated in Table VIII and the confusion matrices for the
CSCDPL approach is illustrated in table VII.
Our approaches consistently outperform all the competing
approaches and the basic reason for the good recognition
performance, even with only a few training examples, is
that the new regularizer constraint encourages the input
signals from the same class to have similar sparse codes and
those from different classes to have dissimilar sparse codes
thereby maintaining a high classification accuracy even using
a smaller training set.
IV. CONCLUSION
We present a novel dictionary learning framework to evalu-
ate the robustness of activity recognition and intensity estima-
tion of aerobic activities using data from wearable sensors. The
main contribution is integrating class specific centralized regu-
larizer term into objective function for dictionary pair learning.
The proposed objective function is efficiently optimized by
combination of alternating direction method of multipliers
and l1 − ls minimization method. Experimental results show
that the classifiers built in this framework achieves impressive
classification performance over two recognition tasks and
outperforms state-of-the-art methods along with being trained
activity and subject independent. Both of these are important
considerations for developing systems that must be robust,
scalable and must perform well in real world settings. Based
on our promising results, we will further extend this work by
incorporating an unsupervised feature learning framework that
learns multiple representation layers, thus capturing features at
a variety of abstraction levels while simultaneously preserving
the spatial smoothness between similar activities. We will
also consider implementing this framework in exergaming
environments, thus testing it on large-scale group datasets
containing more types of human activities. Our system will
be useful in a variety of healthcare applications, including
calculating energy expenditures, determining context aware-
ness, and making accurate classifications in activity based
environments.

REFERENCES
[1] Xi Long, Bin Yin, and Ronald M Aarts, “Single-accelerometer-based
daily physical activity classification,” in Engineering in Medicine and
Biology Society, 2009. EMBC 2009. Annual International Conference of
the IEEE. IEEE, 2009, pp. 6107–6110.
[2] Kwangsoo Kim Inho Lee Sun Ha Jee Mi-hee Lee, Jungchae Kim and
Sun Kook Yoo, “Physical activity recognition using a single tri-axis
accelerometer,” in Proceedings of the world congress on engineering
and computer science, 2009, vol. 1.
[3] Juha Pärkkä, Miikka Ermes, Panu Korpipää, Jani Mäntyjärvi, Johannes
Peltola, and Ilkka Korhonen, “Activity classification using realistic data
from wearable sensors,” Information Technology in Biomedicine, IEEE
Transactions on, vol. 10, no. 1, pp. 119–128, 2006.
[4] Miikka Ermes, Juha Pärkkä, and Luc Cluitmans, “Advancing from
offline to online activity recognition with wearable sensors,” in En-
gineering in Medicine and Biology Society, 2008. EMBS 2008. 30th
Annual International Conference of the IEEE. IEEE, 2008, pp. 4451–
4454.
[5] Emmanuel Munguia Tapia, Stephen S Intille, William Haskell, Kent
Larson, Julie Wright, Abby King, and Robert Friedman, “Real-time
recognition of physical activities and their intensities using wireless
accelerometers and a heart rate monitor,” in Wearable Computers, 2007
11th IEEE International Symposium on. IEEE, 2007, pp. 37–40.
[6] Juha Parkka, Miikka Ermes, Kari Antila, Mark van Gils, Ari Manttari,
and Heikki Nieminen, “Estimating intensity of physical activity: a
comparison of wearable accelerometer and gyro sensors and 3 sensor
locations,” in Engineering in Medicine and Biology Society, 2007. EMBS
2007. 29th Annual International Conference of the IEEE. IEEE, 2007,
pp. 1511–1514.
[7] Nishkam Ravi, Nikhil Dandekar, Preetham Mysore, and Michael L
Littman, “Activity recognition from accelerometer data,” in AAAI, 2005,
vol. 5, pp. 1541–1546.
[8] Attila Reiss and Didier Stricker, “Introducing a modular activity
monitoring system,” in Engineering in Medicine and Biology Society,
EMBC, 2011 Annual International Conference of the IEEE. IEEE, 2011,
pp. 5621–5624.
[9] Pavel Dohnálek, Petr Gajdoˇs, and Tomáˇs Peterek, “Tensor modifica-
tion of orthogonal matching pursuit based classifier in human activity
recognition,” in Nostradamus 2013: Prediction, Modeling and Analysis
of Complex Systems, pp. 497–505. Springer, 2013.
[10] Hristijan Gjoreski, Simon Kozina, Matjaz Gams, Mitja Lustrek, Juan An-
tonio Álvarez-Garc´ıa, Jin-Hyuk Hong, Julian Ramos, Anind K Dey,
Maurizio Bocca, and Neal Patwari, “Competitive live evaluations of
activity-recognition systems,” Pervasive Computing, IEEE, vol. 14, no.
1, pp. 70–77, 2015.
[11] Ana M Martınez, Geoffrey I Webb, Shenglei Chen, and Nayyar A Zaidi,
“Scalable learning of bayesian network classifiers,” 2015.
[12] Sherin M Mathews, Chandra Kambhamettu, and Kenneth E Barner,
“Maximum correntropy based dictionary learning framework for phys-
ical activity recognition using wearable sensors,” in International
Symposium on Visual Computing. Springer, 2016, pp. 123–132.
[13] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng, “Projec-
tive dictionary pair learning for pattern classification,” in Advances in
Neural Information Processing Systems, 2014, pp. 793–801.
[14] Bao-Di Liu, Liangke Gui, Yuting Wang, Yu-Xiong Wang, Bin Shen, Xue
Li, and Yan-Jiang Wang, “Class specific centralized dictionary learning
for face recognition,” Multimedia Tools and Applications, pp. 1–19.
[15] Barbara E Ainsworth, William L Haskell, Melicia C Whitt, Melinda L
Irwin, Ann M Swartz, Scott J Strath, WILLIAM L O Brien, David R
Bassett, Kathryn H Schmitz, Patricia O Emplaincourt, et al., “Com-
pendium of physical activities: an update of activity codes and met
intensities,” Medicine and science in sports and exercise, vol. 32, no.
9; SUPP/1, pp. S498–S504, 2000.
[16] Attila Reiss, Gustaf Hendeby, and Didier Stricker, “A novel confidence-
based multiclass boosting algorithm for mobile physical activity moni-
toring,” Personal and Ubiquitous Computing, vol. 19, no. 1, pp. 105–
121, 2015.
[17] Attila Reiss, Personalized mobile physical activity monitoring for
everyday life, Ph.D. thesis, Technical University of Kaiserslautern, 2014.
[18] Davide Anguita, Alessandro Ghio, Luca Oneto, Xavier Parra, and
Jorge L Reyes-Ortiz, “Human activity recognition on smartphones using
a multiclass hardware-friendly support vector machine,” in International
Workshop on Ambient Assisted Living. Springer, 2012, pp. 216–223.
[19] Charissa Ann Ronaoo and Sung-Bae Cho, “An optimal feature selection
method based on random forests for activity recognition with smartphone
sensors,” pp. 448–450, 2014.

Centralized Class Specific Dictionary Learning for Wearable Sensors based physical activity recognition

More Related Content

What's hot (19)

Similar to Centralized Class Specific Dictionary Learning for Wearable Sensors based physical activity recognition (20)

Recently uploaded (20)

Centralized Class Specific Dictionary Learning for Wearable Sensors based physical activity recognition