Professional Documents
Culture Documents
AbstractWith recent progress in pervasive healthcare, features in this activity data are the most significant, and
physical activity recognition with wearable body sensors has how these data can be most effectively employed to classify
become an important and challenging area in both research and the activities [3], [4], [7], [8]. Other research has focused on
industrial communities. Here, we address a novel technique for
a sensor platform that performs physical activity recognition by which computational model is the most relevant to represent
leveraging a class specific regularizer term into the dictionary human activity data [9], [10], [11]. Despite such research
pair learning objective function. The proposed algorithm jointly efforts, the scalability of handling large intra-class variations
learns a synthesis dictionary and an analysis dictionary in and the robustness of many human-activity recognition
order to simultaneously perform signal representation and techniques to the model parameters remains limited.
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 To make physical activity monitoring practicable in
thereby proving beneficial for the classification stage. In order everyday life scenarios, an activity-recognition framework
to develop a more practical approach, we employ a combination must be able to manage a wide range of everyday, household,
of an alternating direction method of multipliers and a l1 ls or sport activities and must handle a variety of potential
minimization method to approximately minimize the objective
function. We validate the effectiveness of our proposed model users. Although current research has focused on increasing
by employing it on two activity recognition problem and an the number of activities that are recognized, each increase in
intensity estimation problem, both of which include a large the number of activities causes the classification performance
number of physical activities. Experimental results demonstrate to fall off. Secondly, recording and using only a small set of
that classifiers built in this dictionary learning based framework a few activities in basic activity recognition, without having
outperforms state of art algorithms by using simple features,
thereby achieving competitive results when compared with simultaneous background activities, limits the applicability of
classical systems built upon features with prior knowledge. the developed algorithms. Recently, infilling of missing pixels,
and image and speech classification problems have been
Keywords Body sensor networks, Wearable Sensors, Hu- successfully addressed by dictionary learning particularly
man Activity Recognition, Physical Activity Monitoring, In- because of its robustness towards missing data and noise. Real-
tensity Estimation of Physical Activity, Inertial sensor, Heart time scenarios might include activity switching, thus requires
Rate, Sensor localization, Sparse Representation, Dictionary testing on a wider range of activities than were used for
Pair Learning training [12]. Thus, successful activity-recognition requires
dictionary learning frameworks [12] that are sufficiently
I. I NTRODUCTION AND R ELATED W ORK robust for classification even with limited training data.
The classification of basic physical activities (walk, run,
cycle) and basic postures (lie, stand, sit) is well researched In this paper, we present a unified dictionary pair learning
[1], [2], [3], [4], and good recognition performance can framework by incorporating a centralized class specific
be obtained using a single 3D-accelerometer and simple regularizer term to solve the wearable sensor based
classifiers. Moreover, recent studies have focused on classification problem. In the dictionary pair learning
estimating the intensity of an activity (i.e., light, moderate or framework, the training samples of each class contribute
vigorous) ( e.g., in [5], [6]) by means of metabolic equivalent identically to the dictionary, thus generating a dictionary
(MET), a parameter that attributes to the energy expenditure having training samples corresponding to each class. This
of a physical activity. can result in instability and high residual error, which is
detrimental to the recognition performance. Using traditional
Over the years, many studies have analyzed both simple class specific dictionary learning approaches [35] does
and complex human activities reported by wearable sensors. allow learning a dictionary for each class, but might
A large number of these have focused on determining which result in interdependence in sparse codes and erroneous
sliding window, using a window size of 512 samples. During P , A , D = arg min kXk Dk Ak k2F
P,A,D
k=1
(3)
the feature extraction stage, signal features extracted from
the segmented 3D-acceleration data are calculated for each + kPk Xk Ak k2F + ||Pk , Xi ||2F
of the three axis separately and for the 3 axes together. Here Xk represents the complementary data matrix of Xk in
Inclusion of Heart Rate (HR) monitor with the commonly the training set X, > 0 denotes a scalar constant, and di is
used inertial sensors proved specifically useful for physical the ith atom of synthesis dictionary D.
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 B. Dictionary Pair Learning model based on centralized class
features from Inertial Measurement Unit (IMU) acceleration specific regularizer term
data and 4 features from HR data.
The objective is to incorporate a centralized class specific
A. Dictionary Pair Learning Framework regularizer term [14] to obtain a unified dictionary pair
learning framework. In DPL model, training samples of each
Dictionary Pair Learning classification algorithm jointly
class contribute equivalently to the dictionary, thus generating
learns a synthesis dictionary and an analysis dictionary to
a dictionary consisting of training samples in corresponding
attain the objective of signal representation and discrimination
class, resulting in instability and higher residual error. The
[13]. To explain discriminative dictionary learning, a set of p-
main contribution here is to explicitly incorporate centralized
dimensional training samples from K classes can be defined
class specific sparse codes [21] to the dictionary pair learning
as X = [X1 , ........, Xk , ...., XK ], where Xk Rpn is the
objective function, thereby, making sparse codes in the same
training sample set of class k, and n is the number of samples
class concentrated.
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 To attain this objective,we denote the mean of each row of
training data, and can be formulated as: sparse code A as E(A). The regularizer term can be formulated
as : XN
min kX DAk2F + ||A||p + (D, A, Y) (1)
D,A R(Ak ) = k(Ak ) n E(Ak )k22 (4)
n=1
Here > 0 is a scalar constant; Y denotes the class label where is the tradeoff parameter between the reconstruction
matrix of samples in X; D is the synthesis dictionary to be error and the degree of deviation from the sparse code to their
learned; and A is the coding coefficient matrix of X over centers and A n represent the nth column of a sparse matrix
D. The data fidelity term kX DAk2F in the training model A. Incorporating the class specific regularizer term to the DPL
(equation (1)), ensures the representation ability of D; ||A||p objective function, equation (3) can be formulated as:
X
K
P , A , D = arg min kXk Dk Ak k2F + kPk Xk Ak k2F
P,A,D
k=1 X
N X
N
A = 2 tn (Ak )n + (Ak )
n S(Ak )n +
X
N
+||Pk , Xi ||2F + k(Ak ) n E(Ak )k22 n=1 n=1
n=1 X N 1
N
N 1 X
N
2
(5) [( ) (Ak )n (A k )n 2 (A k )n (Ak )m ]
N N2
n=1 m=1,m6=n
In order to solve this equation, we integrate alternating (12)
direction method of multipliers with l1 ls method to
facilitate convergence. The alternating direction method of The objective function now reduces to equation (13)
multipliers (ADMM) solve convex optimization problems
by fixing some variables and solving for the other variable, N 1 2 X
R
thereby decomposing the problem into smaller sub-problems f (Ak )ln = (Ak )2ln (Sll + [ ] ) + 2(Ak )ln ( Slq (Ak )qn )
N
q=1q6=l
making each of them easier to handle. The minimization can
N 1 X
N
be alternated between the following steps :
2(Ak )ln [ (Ak )mn ] 2(Ak )ln tTln
N2
m=1,m6=n
1:Fix A, update P and D (13)
XK
P = arg min kPk Xk Ak k2F + ||Pk , Xi ||2F (6) Here (Ak )ln denotes all elements in Ak except the element in
P
k=1 XK the l th row and the n th column. Defining
D = arg min kXk Dk Ak k2F N 1 X
(7) N
D
k=1 dln = tTln + [ (Ak )lm ] (14)
N2
m=1,m6=n
The closed form solution for P can be obtained as: f (Ak )ln becomes a piece-wise parabolic function as in
P = A X T ( X X T + X XT + Y I)1
k k k k k k (8) [21]. Adopting the convexity and monotonic property of the
parabolic function and the problem transformation defined in
The closed form solution for D can be obtained by intro-
[21] , we have the closed-form solution as function reaches
ducing a variable S as in [13]
the minimum at the unique point.
XK
Dr+1 = arg min kXk Dk Ak k2F + kDk Srk + Trk k2F ln
dln [S(Ak )]ln
D (Ak )ln = (15)
k=1
(9) Sll + [ NN1 ]2
The solution for D and P are the same as in the DPL In the testing phase, the analysis sub-dictionary Pk is
framework [13] using ADMM algorithm. trained to produce small coefficients for samples from classes
other than k, and thus can only generate significant coding
2. Fix D and P, update A coefficients for samples from class k. Meanwhile, the synthesis
X
K
sub-dictionary Dk is trained to reconstruct the samples of
A = arg min tr[(Xk Dk Ak ) (Xk Dk Ak )]
A class k from their projective coefficients Pk Xk , i.e residual
k=1
will be small. Conversely, since Pk Xi will be small and
+tr[ (Pk Xk Ak ) (Pk Xk Ak )] (10) Dk is not trained to reconstruct Xi , the residual ai will be
X
N
much larger. Thus, if the query sample y is from class k, its
+ k(Ak )n E(Ak )k22 projective coding vector Pk will more likely be large, while its
n=1
projective coding vectors Pi will be small. Therefore, class-
Here, tr represents the trace of a matrix. We use l1 ls specific reconstruction residual are used to identify the class
optimization method to solve for A. Ignoring the constant label of testing sample.
terms, the function can be simplified using l1 ls optimization
method. III. E XPERIMENTAL RESULTS AND DISCUSSION
XK
A. PAMAP2 Database
A = arg min tr[2(Xk Dk + Xk Pk )Ak
A The proposed algorithm is evaluated over the activity
k=1
X
N recognition and intensity estimation classification problems
+Ak (Dk Dk + I)Ak ] + k(Ak ) n E(Ak )k22 defined on the recently released PAMAP2 Physical Activity
n=1 Monitoring Data Set. Briefly, this dataset captures 18 physical
(11) activities performed by 9 subjects wearing 3 IMUs (Inertial
measurement unit) and a HR (heart rate) monitor. Here, the
Defining t = (Xk Dk + Xk Pk ) and
activity classification task consists of 15 different activity
S = (Dk Dk + I) and using property tr(A BA) =
P classes represented as lie, sit, stand, walk, run, cycle, Nordic
N
n=1 An B(An ), equation 11 can be rewritten as 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