The MobiAct Dataset: Recognition of Activities of Daily Living using
Smartphones
George Vavoulas
1
, Charikleia Chatzaki
1,2
, Thodoris Malliotakis
1
, Matthew Pediaditis
1,2
and Manolis Tsiknakis
1,2
1
Technological Educational Institute of Crete, Department of Informatics Engineering,
Biomedical Informatics and eHealth Laboratory, Estavromenos, 71004, Heraklion, Crete, Greece
2
Foundation for Research and Technology – Hellas, Institute of Computer Science,
Computational BioMedicine Laboratory, Vassilika Vouton, 71110, Heraklion, Crete, Greece
Keywords: Human Activity Recognition, Activities of Daily Living, Smartphone, Accelerometer, Dataset.
Abstract: The use of smartphones for human activity recognition has become popular due to the wide adoption of
smartphones and their rich sensing features. This article introduces a benchmark dataset, the MobiAct
dataset, for smartphone-based human activity recognition. It comprises data recorded from the
accelerometer, gyroscope and orientation sensors of a smartphone for fifty subjects performing nine
different types of Activities of Daily Living (ADLs) and fifty-four subjects simulating four different types
of falls. This dataset is used to elaborate an optimized feature selection and classification scheme for the
recognition of ADLs, using the accelerometer recordings. Special emphasis was placed on the selection of
the most effective features from feature sets already validated in previously published studies. An important
qualitative part of this investigation is the implementation of a comparative study for evaluating the
proposed optimal feature set using both the MobiAct dataset and another popular dataset in the domain. The
results obtained show a higher classification accuracy than previous reported studies, which exceeds 99%
for the involved ADLs.
1 INTRODUCTION
Human activity recognition is the process of
identifying and recognizing the activities and goals
of one or more humans from an observed series of
actions. In recent years, human activity recognition
has evoked notable scientific interest due to its
frequent use in surveillance, home health
monitoring, human-computer interaction, ubiquitous
health care, as well as in proactive computing.
Human activities can be further decomposed as a set
of basic and complex activities, namely activities of
daily living (ADLs) and instrumental activities of
daily living (IADLs). Typical approaches use vision
sensors, inertial sensors and a combination of both.
Exploiting the increasing tendency of smartphone
users, latest reports introduce systems which use
smartphone sensors to recognize human activities
(Kwapisz et al., 2011; Siirtola and Röning 2012;
Khan et al., 2010; Lee and Cho, 2011).
The aim of this work is to introduce a benchmark
dataset and to present an optimized system in terms
of feature selection and classification for recognition
of ADLs based on smartphone's triaxial
accelerometer data. The “MobiAct” dataset contains
records of the accelerometer, gyroscope and
orientation sensors of a smartphone from fifty
subjects performing nine different types of ADLs
and fifty-four subjects performing four different
types of falls. In order to achieve an optimized
recognition system, special emphasis was placed on
the selection of the most effective features from
feature sets already validated in published studies.
Furthermore, a comparison study was performed to
evaluate the proposed optimal feature set with the
MobiAct dataset, as well as with an additional
dataset. The results show higher classification
accuracy than previous reported studies.
2 RELATED WORK
As already said, human activity recognition has
evoked notable scientific interest in recent years. A
recent study (Bayat et al., 2014) proposes a
smartphone-based recognition system, in which the
Vavoulas, G., Chatzaki, C., Malliotakis, T., Pediaditis, M. and Tsiknakis, M.
The MobiAct Dataset: Recognition of Activities of Daily Living using Smar tphones.
In Proceedings of the International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AWE 2016), pages 143-151
ISBN: 978-989-758-180-9
Copyright
c
2016 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
143
application of a low-pass filter and a combination of
Multilayer Perceptron, LogitBoost and Support
Vector Machine (SVM) classifiers reached an
overall accuracy of 91.15% when the smartphone
was held in the hand of the user. Samples were
recorded from four volunteers while performing six
activities: slow and fast running, walking, aerobic
dance, ascending stairs (“stairs up”) and descending
stairs (“stairs down”). The sampling rate was set at
100 Hz while a window of 1.28 seconds with 50%
overlap was used for feature extraction.
Anjum and Ilyas (2013) introduced a similar
approach with ten users performing seven different
activities which included walking, running, stairs up,
stairs down, cycling, driving and remaining inactive,
by carrying the smartphone in various positions. A
sampling rate of 15 Hz and matching time windows
of 5 seconds were used. Based on the ranking of the
information gain, nine features were selected from
the auto correlation function. For the classification
process Naϊve Bayes, C4.5 Decision Tree, K-Nearest
Neighbor and SVM classifiers were tested. The C4.5
Decision Tree performed better than the other
classifiers with an accuracy of 95.2%.
Zheng et al. (2014) proposed a two-phase method
to achieve recognition of four different types of
activities (sitting, standing, walking and running)
using tri-axial acceleration data from a Samsung
galaxy SIII smartphone. Five subjects performed the
activities with the phone placed loosely in a pocket.
Records of two minutes were used for the training
phase while for the testing phase data from
continuous records of several days were used. A
sampling rate of 100 Hz was used. In order to achieve
noise reduction, the authors deployed Independent
Components Analysis, specifically the fastICA
algorithm, in combination with the wavelet transform
for feature extraction. For the classification, a Support
Vector Machine was employed using the WEKA
toolkit. A maximum accuracy of 98.78% was
reported for a leave-one-out validation.
Based on tri-axial accelerometer data of a
smartphone, Buber and Guvensan (2014) developed
a recognition system for the following activities:
walking, jogging, jumping, stairs up, stairs down,
sitting, standing and biking. Five volunteers
performed those activities with the smartphone
placed in the front pocket of their trousers. The
sampling rate was set at 20 Hz and a 10 second
moving window was used for feature extraction. The
evaluation was performed with two feature selection
algorithms (OneRAttributeEval and ReliefF
AttributeEval) and six classification algorithms (J48,
K-Star, Bayes Net, Naïve Bayes, Random Forest,
and k-NN) using 10-fold cross-validation. The
authors resulted in a combination of 15 features with
k-NN to perform best at a recognition rate of 94%.
Fan et al. (2013) studied three different decision
tree models based on a) the activity performed by the
user and the position of the smartphone (vector), b)
only the position and c) only the activity. Fifteen users
performed five kinds of activities: stationary, walking,
running, stairs up and stairs down with the smartphone
placed into a carrying bag, a trouser pocket or in the
hand. Ten-second samples of accelerometer data were
recorded for each different kind of activity and position
of smartphone. The authors concluded that the model
based only on the activity outperformed the other two
with an accuracy of 88.32%.
In another study (Siirtola and Roning, 2013),
accelerometer data from a smartphone were
recorded with a sampling frequency of 40Hz while
seven volunteers were performing five different
activities: walking, running, cycling, driving a car,
and sitting/standing. In each recording, four
smartphones were placed in various positions,
namely, trousers’ front pocket, jacket’s pocket, at
backpack, at brachium and one was held at the ear
only when it was physically allowed. For feature
extraction a sliding window of 7.5 seconds with 25%
overlap in an online (on device) application and one
with 50% overlap in an offline application, were
used. Classification was achieved using five
classifiers based on quadratic discriminant analysis
arranged in a three stage decision tree topology.
Average recognition rate of almost 98.9% was
reported in the offline and 90% in the online system.
Exploiting the accelerometer sensor of a
smartphone (Dernbach et al. 2012) developed a
system for recognizing simple (biking, stairs up,
driving, lying, running, sitting, standing and
walking) and complex (cooking, cleaning etc.)
activities performed by ten participants. The
sampling frequency was set at 80 Hz maximum
although variations in the sampling rate were
reported. Multiple windows sizes of 1, 2, 4, 8 and 16
seconds with 50% overlap were used. The placement
of the smartphone, in terms of position and
orientation, was left at each user’s will. Although
complex activities were classified with an accuracy
of 50%, simple activities were classified with 93%
accuracy with a Multilayer Perceptron and a window
size of 2 seconds.
Saputri et al. (2014) proposed a system for
activity recognition in which twenty-seven subjects
performed six types of activities, namely, walking,
jogging, running, stairs up, stairs down and hopping.
The smartphone was placed in the front trouser
ICT4AWE 2016 - 2nd International Conference on Information and Communication Technologies for Ageing Well and e-Health
144
Table 1: Overview of the methodology and results followed by the related studies.
Study
No of
subjects
Activities
1
Sampling
Frequency
Window
size/overlap
No of
Features
Smartphone
position
Algorithms
2
Performance
(Bayat et
al., 2014)
4
RUN, SWL
FWL,
ADN, STU,
STN
100Hz 1.28s/50% 18
hand of the
user
J48, K-
Star, BN,
NB, RF,
kNN
MLP & LB
&SVM:
91,15%
Accuracy
(Anjum
and Ilyas,
2013)
10
RUN, STN,
STU, BIK,
STC, DRI,
INA
15Hz 5s 9*
various
positions
NB, C4.5,
KNN,
SVM
C4.5:
95.2%.
(Zheng et
al., 2014)
5
SIT, STD,
WAL, RUN
100Hz -
ICA +
Wavelet
freely in
pocket
SVM 98.78%
(Buber
and
Guvensan
, 2014)
5
WAL,
JOG, STN,
STU, SIT,
JUM, BIK
20Hz 10s 15
front
pocket
J48, K-
Star, BN,
NB, RF,
kNN
k-NN: 94%
(Fan et
al., 2013)
15
STC,
WAL,
RUN,
STU,STN
- 10s 10*
bag,
trouser
pocket &
hands
ID3 DC 80.29%
(Siirtola
and
Roning,
2013)
7
WAL,
RUN, BIK,
DRI,
SIT/STD
40Hz
7.5s/25%
online app
7.5s/50%
offline app
76
5
smartphone
s : various
position
DC &
QDA
90% online
98.9%
offline
(Dernbac
h et al.,
2012)
10
BIK, STU,
DRI, LAY,
RUN, SIT,
STD WAL.
80Hz
1,2,4,8,16/50
%
6
user’s
choice
(position &
orientation)
MLP, NB,
BN, DT,
B-FT, K-
star
MLP: 93%
2s window
(Saputri
et al.,
2014)
27
WAL, RUN,
STN,STU,
HOP
50Hz 2s 21
front
pocket
ANN 93%
1
WAL: Walking, JOG: Jogging, STN: Stairs down, STU: Stairs up, SIT: Sitting, STD: Standing, RUN: Running, BIK: Biking, LAY:
Laying down, STC: Static, ADN: Aerobic dancing, HOP: Hopping, DRI: Driving, INA: Inactivity.
2
J48: Weka implementation of C4.5 Decision Tree, LR: Logistic Regression, MLP: Multilayer Perceptron, kNN: k-Nearest Neighbors,
SMO: Sequential Minimal Optimization, NB: Naïve Bayes, SVM: Support Vector Machines, RF: Random Forest, DT: Decision Table, B-
FT: Best-First Tree
*
Feature set includes that number of features but is not limited to.
pocket using a sampling rate of 50 Hz. In the feature
extraction process, the window size was set at 2
seconds, while feature selection was performed
using a self-devised three-staged genetic algorithm.
The use of an Artificial Neural Network produced
93% accuracy in the activity recognition.
The above non-exhaustive review on ADLs
recognition systems using smartphone embedded
inertial sensors reveals that several research studies
have already been published, reporting acceptable
results while employing various different data
processing and analysis approaches. However, there
is an inherent weakness of conducting objective
comparisons between different implementations,
because of the heterogeneity of the acquired raw
data, as shown in Table 1. The issue of
differentiation in smartphone positions, sampling
frequency and the kinds of activities addressed,
along with the relatively small number of subject
recordings is addressed in the following work with
the use of the developed MobiAct dataset.
3 THE MOBIACT DATASET
3.1 Dataset Description
MobiAct is a publicly available dataset (available for
download from www.bmi.teicrete.gr) which includes
data from a smartphone when participants are
performing different types of activities and a range
of falls. It is based on the previously released
MobiFall dataset (Vavoulas et al. 2014), which was
initially created with fall detection in mind. The fact
that MobiFall included various activities of daily
living made it also suitable for research in human
activity recognition. In its current version, and with
The MobiAct Dataset: Recognition of Activities of Daily Living using Smartphones
145
a more generic name, MobiAct is introduced for the
first time in the context of this study.
It encompasses four different types of falls and
nine different ADLs from a total of 57 subjects with
more than 2500 trials, all captured with a smartphone.
The activities of daily living were selected based on
the following criteria: a) Activities which are fall-like
were firstly included. These include sequences where
the subject usually stays motionless at the end, in
different positions, such as sitting on a chair or
stepping in and out of a car; b) Activities which are
sudden or rapid and are similar to falls, like jumping
and jogging; c) The most common everyday activities
like walking, standing, ascending and descending
stairs (“stairs up” and “stairs down”). These activities
were included from the start of the effort, since our
ultimate objective has been to extend our work
towards recognition of not only falls, but also
complex everyday activities and, eventually,
behaviours. Moreover, the fact that such activities are
included is an advantage concerning human activity
recognition (HAR) in general. As a result, MobiAct is
suitable investigating both fall detection and HAR.
Table 2 and Table 3 summarize all captured activities
(and activity codes), their present trial counts,
durations and a short description for each activity.
3.2 Dataset Acquisition Details
All activities related to the design of the acquisition
protocol and the acquisition of the MobiAct dataset
itself were performed at the Technological
Educational Institute of Crete. Data were recorded
from the accelerometer, gyroscope and orientation
sensors of a Samsung Galaxy S3 smartphone with
the LSM330DLC inertial module (3D accelerometer
and gyroscope). The orientation sensor is software-
based and derives its data from the accelerometer
and the geomagnetic field sensor. The gyroscope
was calibrated prior to the recordings using the
device’s integrated tool. For the data acquisition, an
Android application has been developed for the
recording of raw data for the acceleration, the
angular velocity and orientation (Vavoulas et al.
2013). In order to achieve the highest sampling rate
possible the parameter “SENSOR_DELAY
FASTEST” was enabled. Finally, each sample was
stored along with its timestamp in nanoseconds.
The techniques applied in the majority of
published studies focusing on smartphone-based
activity recognition, require the smartphone to be
rigidly placed on the human body and with a specific
orientation. For this purpose a strap is frequently
used. In contrast to this and in an attempt to simulate
every-day usage of mobile phones, our device was
located in a trousers’ pocket freely chosen by the
subject in any random orientation. For the falls, the
subjects used the pocket on the opposite side of the
direction of the fall to protect the device from
damage. For the simulation of falls a relatively hard
mattress of 5 cm in thickness was employed to
dampen the fall (Vavoulas et al. 2014).
3.3 Dataset Participants
For the generation of the MobiAct dataset 57
subjects (42 men and 15 women) were recorded
while performing the predefined activities. The
subjects’ age spanned between 20 and 47 years
(average: 26), the height ranged from 160 cm to 189
cm (average: 175), and the weight varied from 50 kg
to 120 kg (average: 76). 50 subjects completed
Table 2: Falls recorded in the MobiAct dataset.
Code Activity Trials Duration Description
FOL Forward-lying 3 10s Fall Forward from standing, use of hands to dampen fall
FKL Front-knees-lying 3 10s Fall forward from standing, first impact on knees
SDL Sideward-lying 3 10s Fall sideward from standing, bending legs
BSC Back-sitting-chair 3 10s Fall backward while trying to sit on a chair
Table 3: Activities of Daily Living recorded in the MobiAct dataset.
Code Activity Trials Duration Description
STD Standing 1 5m Standing with subtle movements
WAL Walking 1 5m Normal walking
JOG Jogging 3 30s Jogging
JUM Jumping 3 30s Continuous jumping
STU Stairs up 6 10s Stairs up (10 stairs)
STN Stairs down 6 10s Stairs down (10 stairs)
SCH Sit chair 6 6s Sitting on a chair
CSI Car step in 6 6s Step in a car
CSO Car step out 6 6s Step out of a car
ICT4AWE 2016 - 2nd International Conference on Information and Communication Technologies for Ageing Well and e-Health
146
successfully all ADLs and 54 subjects completed all
falls. In total 10 trials had to be removed from the
dataset due to errors in acquisition.
4 METHODS
4.1 Datasets for Comparison and
Evaluation
Our intention for generating MobiAct was to enable
testing and benchmarking between various methods
for human activity recognition with smartphones. As
a result a comparison to other existing and publically
available datasets is of significant value. The most
suitable such public dataset is the WISDM dataset
(Kwapisz et al., 2011). Both WISDM and MobiAct
datasets include a large set of the same ADLs,
namely walking, jogging, stairs up, stairs down,
sitting and standing, in a common file format.
Moreover, the position of the mobile device is
equally treated in both datasets since it is up to each
subject to freely select the orientation it will be
placed into the pocket.
Other freely available datasets, such as the
DALIAC dataset (Leutheuser et al., 2013) and the
UCI dataset (Anguita et al., 2012) could not be used
for comparison since they differ significantly in
terms of the recorded ADLs and the data acquisition
conditions, which should be overlapping as much as
possible among all the datasets under consideration.
For example, the DALIAC dataset uses multiple
accelerometer nodes statically placed on the human
body. It does not use smartphone-based inertial
sensors and therefore it is not suitable for the study
at hand. The UCI data were recorded with a specific
position for the smartphone (waist mounted). In
addition, the UCI dataset does not include the
jogging activity, which is part of both MobiAct and
WISDM datasets, but instead includes the lying
down activity, which is not part of MobiAct and
WISDM. Apart from these differences, significant
differences in the data format prevented the
utilization of the UCI dataset.
4.2 Pre-processing
In order to extract features from the two selected
datasets a common file format and sampling rate for
both must be achieved. Following MobiAct’s file
format, the WISDIM raw data file was split into
smaller files based on the subject’s ID and activity.
Linear interpolation and subsampling was applied on
the MobiAct data in order to achieve a 20Hz sampling
frequency which is what is used for the production of
the WISDM dataset. 20Hz as a sampling frequency is
also reported by Shoaib et al. (2015) as being suitable
for the recognition of ADLs from inertial sensors. In
MobiAct, the duration of some types of activities was
smaller than 10 sec, which is the time window for
feature extraction that the WISDM study uses
(Kwapisz et al., 2011). To achieve a minimum of 10
sec trial duration especially in trials of stairs up, stairs
down and sitting on chair the last sample of each file
in question was padded.
4.3 Reproduction of the WISDM Study
An important qualitative part of this investigation is
the validation of the feature extraction techniques
through the reproduction of a published
computational pipeline and the comparison of the
results. For this purpose the reported study (Kwapisz
et al., 2011) was selected, which uses the WISDM
dataset. Our hypothesis is that, if the results of the
reproduction of the WISDM study are
approximately the same as the published results,
then the feature set defined could be used for a
comparison to other feature sets, such as the one
reported by Vavoulas et al. (2014).
The results from the reproduction of the WISDM
study are presented in Table 4. In general the
reproduced and the reported results have the same
behaviour in both studies. Some minor deviations
may be due to slight differences in the windowing
and feature extraction methodology, since, as
previously mentioned, we had to split the WISDM
data into smaller files.
4.4 Feature Extraction & Feature Sets
In attempting to estimate with the parameters for an
optimal computational and analysis pipeline, it is
obvious that the selection of a respective optimal
feature set is of paramount importance. To construct
this feature set, a combination of features from the
study using the precursor of MobiAct (Vavoulas et
al. 2014) and the WISDM study (Kwapisz et al.,
2011) were used.
4.4.1 Feature Set A (FSA)
This feature set consists of 68 features based on the
reported work in (Vavoulas et al. 2014). For most of
the features a value was extracted for each of the
three axes (x, y, z). In detail, the following features
were computed within each time window:
The MobiAct Dataset: Recognition of Activities of Daily Living using Smartphones
147
21 features in total from: Mean, median,
standard deviation, skew, kurtosis, minimum
and maximum of each axis (x, y, z) of the
acceleration.
1 feature from: The slope SL defined as:
=
(
−
)
+(
−
)
+(
−
)
(1)
4 features from: Mean, standard deviation,
skew and kurtosis of the tilt angle TA
i
between
the gravitational vector and the y-axis (ssince
the orientation of the smartphone was not
predefined it is expected that the negative y-
axis will not be always pointing towards the
vertical direction). The tilt angle is defined as:

=sin

+
+
(2)
where x, y and z is the acceleration in the
respective axis.
11 features from: Mean, standard deviation,
minimum, maximum, difference between
maximum and minimum, entropy of the energy
in 10 equal sized blocks, short time energy,
spectral centroid, spectral roll off, zero crossing
rate and spectral flux from the magnitude of the
acceleration vector.
31 additional features were calculated from the
absolute signals of the accelerometer, including
mean, median, standard deviation, skew,
kurtosis, minimum, maximum and slope.
4.4.2 Feature Set B (FSB)
A total of 43 features were generated in accordance
to the WISDM study reported by Kwapisz, Weiss
and Moore (2011) as variants of six basic features.
For each of the three axes, the average acceleration,
standard deviation, average absolute difference, time
between peaks and binned distribution (x10 bins)
were calculated in addition to the average resultant
acceleration as a single feature.
4.4.3 Optimal Feature Set (OFS)
Following elaborate experimentation (totally 70
different experimental setting) in which a) various
combinations of window size (10, 5, 2 sec) and
overlap (0%, 50%, 80%) were tested, b) features
were removed or added into the feature vector based
on observations of the achieved accuracy, and c)
different classifiers were employed, such as IBk,
J48, Logistic regression, Multilayer Perceptron and
LMT (from the WEKA’s algorithm set), an optimal
feature set, in our view, has been produced. All
experiments were conducted using 10-fold cross-
validation. Specifically, the two feature sets (FSA
and FSB), obtained using a time window of 5 sec
and 80% overlap, were at first combined to form one
new feature set. Subsequently weak features,
identified through a trial-and-error approach, were
taken out in an iterative process until the best overall
accuracy for both datasets (MobiAct and WISDM)
was obtained. A total number of 64 features were
thus retained to form the optimal feature set. The
features excluded from FSA were kurtosis for the x,
y and z axes and spectral centroid. The features
excluded from FSB were: time between peaks,
binned distribution and average absolute difference.
The optimal feature set was also calculated by using
a 10s window and no overlap as defined in the
WISDM study for a final comparison to their results,
as shown in Table 4.
Table 4: Classification results (% accuracy) in comparison to the WISDM published results (10s window size, no overlap).
Activity
Published Results Reproduced Results (FSB)
Results using the optimal
feature set (OFS)
J48 Logistic
Multilayer
Perceptron
J48 Logistic
Multilayer
Perceptron
J48 Logistic
Multilayer
Perceptron
Walking 89.9 93.6 91.7 90.8 93.8 95.3 99.4 98.3 99.8
Jogging 96.5 98.0 98.3 98.5 98.6 99.0 99.1 99.4 99.6
Upstairs 59.3 27.5 61.5 65.5 53.2 79.3 85.2 79.5 92.5
Downstairs 55.5 12.3 44.3 55.6 49.7 69.4 87.4 77.4 91.5
Sitting 95.7 92.2 95.0 97.0 94.1 94.6 97.0 97.5 98.0
Standing 93.3 87.0 91.9 97.0 94.6 90.4 99.4 97.0 99.4
Overall 85.1 78.1 91.7 88.3 87.5 92.4 96.7 94.9 98.2
ICT4AWE 2016 - 2nd International Conference on Information and Communication Technologies for Ageing Well and e-Health
148
Table 5: Classification results using the optimal feature set (5s window size, 80% overlap).
Dataset/Classifier: MobiAct/IBk MobiAct/J48 WISDM/IBk WISDM/J48
Activity TPRate FPRate TPRate FPRate TPRate FPRate TPRate FPRate
Walking 1.000 0.000 1.000 0.000 1.000 0.000 0.998 0.002
Jogging 1.000 0.000 1.000 0.000 0.999 0.000 0.998 0.001
Upstairs 0.993 0.001 0.930 0.004 0.992 0.001 0.939 0.006
Downstairs 0.982 0.000 0.921 0.003 0.991 0.001 0.937 0.007
Sitting 1.000 0.000 0.999 0.000 0.999 0.000 0.996 0.000
Standing 1.000 0.000 1.000 0.000 0.999 0.000 0.996 0.000
Accuracy: 99.88 % 99.30 % 99.79 % 98.63 %
4.5 Classifiers
The classifiers selected for the final testing of the
optimal feature set were the IBk (with 1 nearest
neighbor), the J48 decision tree, Logistic regression
and Multilayer perceptron, included in WEKA (Hall
et al. 2009) with default parameters. The first two
produced the best overall results, whilst the
remaining two were used for a comparison to the
WISDM study since they were also reported there.
5 RESULTS
The experimental results obtained using the optimal
feature set are shown in Table 5. It is worth noticing
that with both classifiers the overall accuracy is
close to 99% for both datasets. The best accuracy for
the MobiAct dataset is obtained with the IBk
classifier. IBk generally appears to have a relative
better performance with 94% accuracy, a fact that
has already been reported elsewhere (Buber and
Guvensan, 2014). Also, IBk performs better than J48
for the WISDM dataset as well. The weakness in
accurately recognizing activities which produce
similar signals, such as stairs up and stairs down, is
noticeable with J48. Nevertheless, IBk recognizes
these activities effectively. An additional noticeable
point is that IBk performs slightly better in
classifying the walking activity, which has been
observed to be often misclassified as a stairs up or
stairs down activity.
Considering the comparison of the results when
using FSB (reproduced results) and OFS with the
WISDM dataset, for all the classifiers used, OFS
outperforms FSB (Table 4). A possible explanation
to this may be the higher number of features used in
OFS. This finding is in line with related published
evidence. As reported by Siitrola and Roning (2013),
accuracy of 98.9% achieved with the use of a large
feature set (75 features).
6 CONCLUSIONS
The study’s objective was to estimate an optimal
computational and analysis pipeline which
accurately recognizes ADLs exploiting an extensive
dataset of motion data collected from a smartphone.
As a result of this investigation a set of 64 features
that proved to perform best with two datasets was
extracted. These features were the outcome of many
tests, through a trial and error process that removed
weak features such as kurtosis and spectral centroid.
It is noticeable that absolute values of kurtosis in all
three axes improve the performance of classification
and hence were included in the final optimal feature
set. The spectral centroid is the key feature, which
affects the results of activity recognition negatively.
The stairs up and stairs down activities exhibit the
worst accuracy among all those performed in the
tests. This observation is also seen in other reports
and may be related with the random device
orientation or the dynamic and temporal resolution
of the accelerometer sensor.
The best overall accuracy of 99.88% is achieved
when using the IBk classification algorithm on the
MobiAct dataset in combination with the optimal
feature set mentioned above. This is the best
reported classification result to date, when
comparing with the most recent studies presented in
Table 1. This result is the outcome of a 10-fold
cross-validation which is a very common evaluation
approach in the related studies, although we expect
to decrease when using a leave-one-out cross-
validation, which is a more realistic scenario. It is
the intention of the authors to advance into such
validation scenarios in the near future. For the above
results a sampling rate of 20Hz, a window size of 5
seconds and an overlap of 80% have been used.
These values are proposed as the optimal for this
experimental setup. The usage of two independent
datasets ensures robustness of the results, always
within the limits of each dataset.
The MobiAct Dataset: Recognition of Activities of Daily Living using Smartphones
149
Finally, the experimental results obtained
indicate that the MobiAct can be considered as a
benchmark dataset since it includes a relatively large
number of records and a wide range of activities in
an easy to manage data format. Furthermore, since
the placement of the smartphone is freely chosen by
the subject in any random orientation we believe that
it represents real life conditions as close as possible.
The next step towards developing a real-life
application requires that a) orientation data is used in
a more efficient manner and b) assessment and
optimization of power consumption (battery usage)
requirements for the feature extraction and
classification algorithms, is thoroughly studied.
ACKNOWLEDGEMENTS
This work is partly funded by the WeMP – Wellness
Management Platform funded by FORTHnet S.A.
The authors gratefully thank all volunteers for
their contribution in the generation of the MobiAct
dataset.
REFERENCES
Anguita, D., Ghio, A., Oneto, L., Parra, X. and Reyes-
Ortiz, J. L., 2012, 'Human Activity Recognition on
Smartphones Using a Multiclass Hardware-Friendly
Support Vector Machine', in Ambient Assisted Living
and Home Care, Springer Berlin Heidelberg, Vitoria-
Gasteiz.
Anjum, A. and Ilyas, M. U., 2013, 'Activity recognition
using smartphone sensors', Consumer
Communications and Networking Conference
(CCNC), 11-14 January 2013, pp. 914-919, DOI:
10.1109/CCNC.2013.6488584.
Bayat, A., Pomplun, M. and Tran, D. A., 2014, 'A Study
on Human Activity Recognition Using Accelerometer
Data from Smartphones', 11th International
Conference on Mobile Systems and Pervasive
Computing (MobiSPC'14), 2014, pp. 450-457,
DOI:10.1016/j.procs.2014.07.009.
Buber, E. and Guvensan, A. M., 2014, 'Discriminative
time-domain features for activity recognition on a
mobile phone', IEEE 9th International Conference on
Intelligent Sensors, Sensor Networks and Information
Processing (ISSNIP), 21-24 April 2014, pp. 1-6,
DOI:10.1109/ISSNIP.2014.6827651.
Dernbach, S., Das, B., Krishnan, N. C., Thomas, B. L. and
Cook, D. J., 2012, 'Simple and Complex Activity
Recognition through Smart Phones', 8th International
Conference on Intelligent Environments (IE), 26-29
June 2012, pp. 214-221.
Fan, L., Wang, Z. and Wang, H., 2013, 'Human Activity
Recognition Model Based on Decision Tree',
Proceedings of the 2013 International Conference on
Advanced Cloud and Big Data (CBD '13), 2013, pp.
64-68, DOI: 10.1109/CBD.2013.19.
Hall, M., Frank, E., Holmes, G., Pfahringer, B.,
Reutemann, P. and Witten, I. H., 2009, 'The WEKA
data mining software: an update', ACM SIGKDD
Explorations Newsletter, vol 11, no. 1, pp. 10-18.
Khan, A. M., Lee, Y. K., Lee, S. Y. and Kim, T. S., 2010,
'Human Activity Recognition via an Accelerometer-
Enabled-Smartphone Using Kernel Discriminant
Analysis', 5th International Conference in Future
Information Technology (FutureTech), 2010, DOI:
10.1109/FUTURETECH.2010.5482729.
Kwapisz, J. R., Weiss, G. M. and Moore, S. A., 2011,
'Activity recognition using cell phone accelerometers',
ACM SIGKDD Explorations Newsletter, 31 March
2011, pp. 74-82, DOI: 10.1145/1964897.1964918.
Lee, Y. S. and Cho, S. B., 2011, 'Activity Recognition
Using Hierarchical Hidden Markov Models on a
Smartphone with 3D Accelerometer', Hybrid Artificial
Intelligent Systems, 2011, pp. 460-467, DOI:
10.1007/978-3-642-21219-2_58.
Leutheuser, H., Schuldhaus, D. and Eskofier, B. M., 2013,
'Hierarchical, Multi-Sensor Based Classification of
Daily Life Activities: Comparison with State-of-the-
Art Algorithms Using a Benchmark Dataset', PLoS
ONE, vol 8, no. 10, DOI:
10.1371/journal.pone.0075196.
Saputri, T. R. D., Khan, A. M. and Lee, S. W., 2014,
'User-Independent Activity Recognition via Three-
Stage GA-Based Feature Selection', International
Journal of Distributed Sensor Networks, 2014, p. 15,
DOI: 10.1155/2014/706287.
Shoaib, M., Bosch, S., Incel, O. D. and Scholten, H., 2015,
'A Survey of Online Activity Recognition Using
Mobile Phones', Sensors, vol 15, pp. 2059-2085.
Siirtola, P. and Röning, J., 2012, 'Recognizing Human
Activities User-independently on Smartphones Based
on Accelerometer Data', International Journal of
Interactive Multimedia and Artificial Intelligence,
2012, pp. 38-45, DOI: 10.1155/2014/706287.
Siirtola, P. and Roning, J., 2013, 'Ready-to-use activity
recognition for smartphones', IEEE Symposium on
Computational Intelligence and Data Mining (CIDM),
16-19 April 2013, pp. 59-64, DOI:
10.1109/CIDM.2013.6597218.
Vavoulas, G., Pediaditis, M., Chatzaki, C., Spanakis, E. G.
and Tsiknakis, M., 2014, 'The MobiFall Dataset: Fall
Detection and Classification with a Smartphone',
International Journal of Monitoring and Surveillance
Technologies Research (IJMSTR), 2014, p. 13, DOI:
10.4018/ijmstr.2014010103.
Vavoulas, G., Pediaditis, M., Spanakis, E. and Tsiknakis,
M., 2013, 'The MobiFall Dataset: An Initial
Evaluation of Fall Detection Algorithms Using
Smartphones', IEEE 13th International Conference on
Bioinformatics and Bioengineering (BIBE).
ICT4AWE 2016 - 2nd International Conference on Information and Communication Technologies for Ageing Well and e-Health
150
Zheng, L., Cai, Y., Lin, Z., Tang, W., Zheng, H., Shi, H.,
Liao, B., and Wang, J., 2014, 'A Novel Activity
Recognition Approach Based on Mobile Phone',
Multimedia and Ubiquitous Engineering, 2014, pp.
59-65, DOI:10.1007/978-3-642-54900-7_9.
The MobiAct Dataset: Recognition of Activities of Daily Living using Smartphones
151