Hierarchical, Multi-Sensor Based Classification of Daily

Life Activities: Comparison with State-of-the-Art

Algorithms Using a Benchmark Dataset
Heike Leutheuser*., Dominik Schuldhaus*., Bjoern M. Eskofier
Digital Sports Group, Pattern Recognition Lab, Department of Computer Science, Friedrich-Alexander-University Erlangen-Nuremberg, Germany

Abstract
Insufficient physical activity is the 4th leading risk factor for mortality. Methods for assessing the individual daily life activity
(DLA) are of major interest in order to monitor the current health status and to provide feedback about the individual
quality of life. The conventional assessment of DLAs with self-reports induces problems like reliability, validity, and
sensitivity. The assessment of DLAs with small and light-weight wearable sensors (e.g. inertial measurement units) provides
a reliable and objective method. State-of-the-art human physical activity classification systems differ in e.g. the number and
kind of sensors, the performed activities, and the sampling rate. Hence, it is difficult to compare newly proposed
classification algorithms to existing approaches in literature and no commonly used dataset exists. We generated a publicly
available benchmark dataset for the classification of DLAs. Inertial data were recorded with four sensor nodes, each
consisting of a triaxial accelerometer and a triaxial gyroscope, placed on wrist, hip, chest, and ankle. Further, we developed a
novel, hierarchical, multi-sensor based classification system for the distinction of a large set of DLAs. Our hierarchical
classification system reached an overall mean classification rate of 89.6% and was diligently compared to existing state-of-
the-art algorithms using our benchmark dataset. For future research, the dataset can be used in the evaluation process of
new classification algorithms and could speed up the process of getting the best performing and most appropriate DLA
classification system.

Citation: Leutheuser H, Schuldhaus D, Eskofier BM (2013) Hierarchical, Multi-Sensor Based Classification of Daily Life Activities: Comparison with State-of-the-Art
Algorithms Using a Benchmark Dataset. PLoS ONE 8(10): e75196. doi:10.1371/journal.pone.0075196
Editor: Ioannis P. Androulakis, Rutgers University, United States of America
Received April 24, 2013; Accepted August 11, 2013; Published October 9, 2013
Copyright: ß 2013 Leutheuser et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits
unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: This work was supported by the Bavarian Ministry for Economic Affairs, Infrastructure, Transport, and Technology and the European fund for regional
development (grant number: IUK-1103-0012//IUK383/001; http://www.stmwivt.bayern.de/themen/). Financial and technical support of the study were provided
by adidas AG. We acknowledge support by Deutsche Forschungsgemeinschaft and Friedrich-Alexander-University Erlangen-Nuremberg within the funding
program Open Access Publishing. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing Interests: The authors have declared that no competing interests exist.
* E-mail: [email protected] (HL); [email protected] (DS)
. These authors contributed equally to this work.

Introduction in body weight is continuing in men, adolescents and children.

Wing et al. [4] review the evidence regarding the role of physical
According to the World Health Organization, the 4th leading activity in the treatment of adult overweight and obesity. They
risk factor for mortality is insufficient physical activity [1]. conclude that it is of major interest to develop better ways of
Approximately 3.2 million people of the world population decease measuring exercise. Thereby better types of exercise can be
each year because of insufficient physical activity [1]. Further- defined that will lead to more adherence to exercise and thus long-
more, the risk of all-cause mortality is 20% to 30% higher for term weight loss.
people with inadequate physical activity compared to those who A wide range of studies show that physically active people have
perform moderate physical activities at least 30 minutes a day [1]. higher levels of health-related fitness and lower rates of various
Moderate physical activities are for example walking, ascending chronic diseases compared to physically inactive people [5–11].
stairs or certain household activities. Walking short distances Methods for assessing the individual daily life activity (DLA) are of
(instead of driving) or ascending stairs (instead of taking an major interest in order to monitor the current health status and to
elevator) are modest possibilities to enhance one’s own activity provide feedback about the individual quality of life.
level day by day [2]. Apart from the effect of moderate physical DLAs can be assessed by different methods. An overview of
activity regarding all-cause mortality, it is assumed that the these methods is given by Warren et al. [12]. Self-reports like
participation in 150 minutes of moderate physical activity per questionnaires and activity diaries are a widely used tool to assess
week reduces the risk of ischaemic heart disease by approximately physical activity. They provide physical activity data from a large
30%, the risk of diabetes by 27%, and the risk of breast and colon number of people in short time. However, self-reports induce
cancer by 21% to 25% [1]. problems with reliability, validity and sensitivity [13]. Therefore,
Ogden et al. [3] state that the prevalence of overweight among the current trend is to replace self-reports with automatic DLA
adolescents aged 2 to 19 years and obesity among men increased classification based on small and light-weight wearable sensors like
significantly during 1999 to 2004. They assume that the increase

PLOS ONE | www.plosone.org 1 October 2013 | Volume 8 | Issue 10 | e75196

 Activity Classification Using a Benchmark Dataset

inertial measurement units. These sensors provide a reliable and

objective measurement of physical activity.
Mannini and Sabatini [14] provide an overview of state-of-the-
art human physical activity classification systems. Most of the
approaches used accelerometers but differed in

N number of sensor axes (uniaxial, biaxial and triaxial acceler-

ometer),
N number of sensors and sensor placement,
N sampling rate,
N number of subjects,
N computed features,
N epoch/window size, and
N number and type of activities.
Figure 1. Sensor placement. Four SHIMMER sensor nodes were
Regarding all these differences, it is difficult to compare newly placed on the wrist, chest, hip, and ankle.
proposed methods to existing approaches in the literature. Ideally, doi:10.1371/journal.pone.0075196.g001
newly proposed methods are compared with other approaches in
the literature based on the same benchmark dataset.
the literature [23,24]. To cover the extremities, one sensor was
The purpose of this paper is twofold. First, we provide an placed on the wrist and one on the ankle. Sensors on the wrist
extensive, publicly available dataset of DLAs to be used as a enable a correct classification of activities mainly dominated by the
benchmark for new algorithms in the future (http://www. upper body [17–19,23]. Positioning a sensor on the ankle is
activitynet.org). Second, we propose a novel, hierarchical, multi- heavily used in gait analysis [25] and activity recognition studies
sensor based classification system for DLAs, that is diligently [17,20]. It has already been shown that sensors on the ankle
compared to existing systems. support the recognition of ascending or descending stairs [26].
Our assumption was that sensor fusion of accelerometers and The range for the accelerometers was 66 g. The range of the
recently increasingly available gyroscopes improves the distinction gyroscopes was 6500 deg/s for the sensor nodes wrist, chest, and
of several single activities like ascending or descending stairs. For hip and 62000 deg/s for the sensor node on the ankle, since
data generation, we measured 23 subjects with accelerometer and larger angular velocities are expected in the lower extremities. The
gyroscope sensors placed at four body positions: wrist, hip, chest sampling rate was set to 204.8 Hz and the data was stored on SD
and ankle. Thirteen activities were considered including postures card.
(sitting, lying, standing), household activities (washing dishes, A mobile phone (Samsung Galaxy S2) was used as labeling
vacuuming, sweeping), walking behaviors (normal walking, run- device. An Android-based labeling App (Fig. 2) (running on the
ning, stairs climbing), and sports activities (bicycling, rope mobile phone) was used to label start time and end time of single
jumping). Our proposed classification system consisted of a activities concurrently to data collection.
hierarchical classifier structure that is flexible in its applicability The type of shirt and shoe (Fig. 1) was the same for all
to other activities that were not investigated in the current study. participants. We used four different shirt sizes (S, M, L, XL) in
In a first step, one classifier, in the following sections denoted by order to ensure tight fit and similar measurement conditions. To
BASE, was used to distinguish between several activity groups. In guarantee similar measurement conditions, we measured the chest
a second step, separate classifiers, in the following sections denoted width of each volunteer. Shirt sizes were assigned according to a
by HOUSE, REST, WALK, and BICYCLE, were used to size chart. The volunteers chose the shoe that they felt most
discriminate between the single activities that were included in comfortable in.
each group. In order to compare the proposed approach to
existing algorithms, several state-of-the-art approaches in the
Subjects
literature were implemented and evaluated using the provided
23 healthy subjects (10 female and 13 male, age 2767 years,
benchmark dataset.
body mass index (BMI) 24.0 kg/m263.5 kg/m2, mean6standard
deviation (SD)) were recruited for the study. Of these 23 subjects,
Methods
Hardware Equipment and Sensor Setup
We collected data using four SHIMMER (Shimmer Research,
Dublin, Ireland) sensor nodes [15]. The SHIMMER sensor node
contains a MSP430F1611 microcontroller. The resolution of the
analog-to-digital converter was 12 bit. Each sensor node consisted
of three accelerometer and three gyroscope axes. The four sensor
nodes were placed on the right hip, the chest, the right wrist, and
the left ankle (Fig. 1). These four positions were chosen according
to previously published results, which are mentioned in the
following.
Sensors closely attached to the bodys center of gravity are to be
preferred [16]. Sensors on the chest, the trunk or the hip satisfy
this condition. Sensors on the hip are used in a variety of different Figure 2. Screenshot of App used for data labeling purposes.
studies [16–22]. Sensors on the trunk or chest are also common in doi:10.1371/journal.pone.0075196.g002

PLOS ONE | www.plosone.org 2 October 2013 | Volume 8 | Issue 10 | e75196

 Activity Classification Using a Benchmark Dataset

21 were right handed and two were left handed. The Research intensity ($6 METs) activities. Table 1 lists the durations and the
Ethics Committee of the Friedrich-Alexander-University Erlan- intensities of the executed activities.
gen-Nuremberg confirmed that there is no necessity to obtain the A researcher that labeled the start and end of each activity
approval of the local Ethics Committee. Ethics approval was accompanied the subject during the whole data acquisition. First,
deemed unnecessary because we measured only volunteers that the static activities (sitting, lying, standing) as well as the household
were healthy, in good physical shape and did not suffer from a activities were performed. The subject was told to use the vacuum
disease. All subjects gave written informed consent about their cleaner with the right hand as this was the position for the wrist
participation. All volunteers filled in the Physical Activity sensor. Otherwise, the wrist sensor signal delivered no suitable
Readiness Questionnaire (PAR-Q [27]). The PAR-Q provides a information about the signal pattern of the right hand. Then the
self-administered screening before performing physical activity. subject had to walk on the university campus to another building.
The aim of the PAR-Q is to identify those people who should In this building, walking upstairs (until the third floor) and walking
consult a doctor before performing physical activity. In the study, downstairs (back to the main floor) was recorded. Afterwards, the
only those people who passed the PAR-Q were considered. The subject walked again on the university campus and performed the
content of the PAR-Q can be found on http://www.activitynet. physically more demanding exercises indoor. One exercise
org. The study protocol involved 13 daily life activities that are included walking on a treadmill (h/p/cosmos quasar, h/p/cosmos
normally performed every day. We used unobtrusive sensors that & medical gmbh, Nussdorf-Traunstein, Germany). The treadmill
did not influence the volunteers and did not pose any additional speed was set to 8.3 km/h. Furthermore, the subjects had to
risk to the volunteers. Furthermore, both supervisors of the study bicycle on an ergometer (sanabike 250 F, MESA Medizintechnik
were first aid-trained. We did not conduct research outside our GmbH, Benediktbeuern, Germany) with two different resistance
country of residence. levels (50 W and 100 W). The treadmill speed and the resistance
level were chosen to obtain activities with different MET values
(Table 1). The subjects were told to keep the revolutions per
Data Acquisition and Study Design
minute constant to 70 during the two different resistance levels.
The subjects put the shoes, the T-shirt, the hip-clip, and the
Thus, the differences of the two levels were not due to different
wrist band on (Fig. 1). The SHIMMER sensor nodes were
revolutions per minute. Next, the subject had to perform the
powered on and put on a plate (Fig. 3, top). For offline
activity rope jumping. For this activity, the subject had to perform
synchronization, the plate was dropped down twice and, in five trials with at least five jumps each.
between, the plate was moved up and down. The sensors were As an example, Fig. 4 shows the linear acceleration in vertical
then placed on the dedicated measurement positions. direction of the hip sensor for the activities lying, standing,
Subjects performed a total of 13 activities (Table 1), which were vacuuming, sweeping, walking, and rope jumping.
taken from the ‘‘Compendium of Physical Activity’’ published by After the data acquisition, the SHIMMER sensor nodes were
Ainsworth et al. [28–30]. In these compendiums, physical activity is taken from the dedicated measurement positions and put again on
characterized in four categories depending on Metabolic Equivalent the synchronization plate (Fig. 3). The described synchronization
of Task (MET) values: sedentary (1.0–1.5 METs), light-intensity pattern was again performed in order to check if problems of the
(1.6–2.9 METs), moderate-intensity (3.0–5.9 METs), and vigorous- sensors occurred during the data acquisition. The SHIMMER
sensor nodes were powered off and the kinematic data and the
labeling data were stored on a PC for offline processing.

Preprocessing
Four datasets had to be excluded from further processing. Of
these four datasets, three datasets were excluded because of
problems during gyroscope initialization. The fourth datasets was
excluded as the data of the ankle sensor node was not available. In
total, 19 datasets were used in the following.
The four SHIMMER sensor nodes were synchronized offline.
For this, the first up-down movement of the sensor signal was
manually selected in the linear acceleration of the vertical
direction in all sensor nodes (Fig. 3, bottom, vertical line). This
point constituted the common start point of all sensors.
The labeling was done automatically due to the saved start and
end times of the Android app (Fig. 2). For each labeled activity,
two seconds at the beginning and at the end were cut, in order to
eliminate measuring errors during labeling.

Proposed Classification System

An overview of our proposed classification system is depicted in
Fig. 5. The rectangles indicate single classification systems,
whereas the circles indicate single activities. The general idea
was to set up a hierarchy of classification systems, where each
system solved a different classification problem. The first classifier
Figure 3. Sensor Synchronization. Plate with four SHIMMER sensor in the hierarchy was the BASE classification system. It discrim-
nodes used for synchronization (top) and sinusoidal synchronization inated four activity groups and two single activities (rope jumping
signal (bottom). Dashed red line depicts the synchronization start point. and washing dishes) that did not fit in any of the groups. The
doi:10.1371/journal.pone.0075196.g003 remaining four classifiers constituted the second hierarchy level.

PLOS ONE | www.plosone.org 3 October 2013 | Volume 8 | Issue 10 | e75196

 Activity Classification Using a Benchmark Dataset

Table 1. List of studied activities, abbreviations, durations, and intensities expressed in Metabolic Equivalent of Task (MET).

Activity Abbreviation Duration [min] Intensity [MET]

Sitting SI 1 1.3
Lying LY 1 1.0
Standing ST 1 1.3
Washing dishes WD 2 2.5
Vacuuming VC 1 3.3
Sweeping SW 1 3.3
Walking WK n.a.* 3.5
Ascending stairs AS n.a.** 5.0
Descending stairs DS n.a.** 3.5
Treadmill running RU 2 9.0
Bicycling on ergometer (50 W) BC50 2 3.5
Bicycling on ergometer (100 W) BC100 2 6.8
Rope jumping RJ n.a.*** 8.8

*All subjects had to walk on the university campus from one building to another building.
**All subjects had to climb stairs to the third floor and then back again.
***All subjects had to perform 5 trials with at least 5 jumps each.
doi:10.1371/journal.pone.0075196.t001

Figure 4. Example signals. Linear acceleration in vertical direction of the hip sensor for six activities. A: Lying, B: Standing, C: Vacuuming, D:
Sweeping, E: Walking, F: Rope jumping.
doi:10.1371/journal.pone.0075196.g004

PLOS ONE | www.plosone.org 4 October 2013 | Volume 8 | Issue 10 | e75196

 Activity Classification Using a Benchmark Dataset

We defined a different feature set for the classification system

REST. We extracted the gravitational component of the
acceleration signal by a third-order elliptic low pass filter with
an infinite impulse response and a cut-off frequency at 0.25 Hz
[32]. This means that all three gravitational acceleration
components of all four sensors were used as features. This was
done because only the orientation of the body is important for the
discrimination of the activities sitting, lying and standing. In total,
this resulted in 12 features for each sliding window.

Classification
Since there is no single classifier that is suitable for all
classification tasks [33], the following classification systems were
Figure 5. Illustration of the proposed classification system. used [33,34]: AdaBoost (ADA), classification and regression tree
Rectangles indicate single classification systems BASE, HOUSE, REST, (CART), k-Nearest Neighbor classifier (kNN) and Support Vector
WALK and BICYCLE. Circles indicate single activities VC (vacuuming), SW Machine (SVM) with a radial basis function kernel. In the case of
(sweeping), SI (sitting), LY (lying), ST (standing), WK (walking), RU
(running), AS (ascending stairs), DS (descending stairs), BC 50 (bicycling, AdaBoost, 100 decision stump learners were used. In the case of
50 watt), BC 100 (bicycling, 100 watt), RJ (rope jumping) and WD kNN, k was set to five. The cost parameter of the SVM classifier
(washing dishes). was set to one and the gamma parameter to one divided by the
doi:10.1371/journal.pone.0075196.g005 number of features. For performance assessment, the mean class
dependent classification rate and the overall mean classification
We chose such a hierarchical system, because new activities can be rate were computed with a leave-one-subject-out procedure for all
introduced without retraining all classifiers. The system was five classification systems. In each leave-one-subject-out trial, all
therefore flexible in its application to different activities. epochs of one certain subject were removed from the training set.
In order to evaluate the whole hierarchical classification system,
Preprocessing the classifier with the best performance was chosen for the systems
The data processing was performed in sliding windows with BASE, HOUSE, REST, WALK, and BICYCLE.
50% overlap [17,20,21]. The width of the window was set to 5 s,
comparable to [17,21], which used 6.7 s and 5.12 s, respectively. Comparison to Algorithms in Literature
Our proposed method was compared to six state-of-the-art
Feature Extraction approaches in literature [17,19–21,23,32]. We have chosen these
approaches due to their citation rate and hence, their state-of-the-
We defined a generic feature set for the classification systems
art research impact. Further, all six approaches had intersections
BASE, HOUSE, WALK, and BICYCLE, which were computed
with our study setup regarding the used sensors (accelerometers
for every sliding window. The generic feature set consisted of six
and gyroscopes), the sensor placement, and the performed
features that were computed for every sensor axis and one feature
activities.
that was computed for each of the accelerometer and gyroscope of
An overview of the different approaches is shown in Table 2.
each sensor node. The six features for every sensor axis included
The third column in the table shows according to each publication
four time domain and two frequency domain features.
the kind of sensor and their original placement. We only used
The four time domain features were: identical sensor positions and sensor data. This means that we only
considered at maximum four sensor positions (Fig. 1) of
N minimum amplitude
accelerometer and gyroscope data and disregarded other signals
N maximum amplitude like the heart rate. We modified, for example, the sensor
N mean amplitude placement and axes alignment suggested by Bao and Intille [17].
N variance of amplitude. The authors used five biaxial accelerometers and placed them on
the right hip, the right wrist, the left arm, the right ankle, and the
The minimum and maximum amplitude extracted range left thigh. As we acquired data with triaxial accelerometers, we
information of the amplitude. The mean and variance of the only used two axes for comparison and only the signals of the three
amplitude gave important knowledge about statistics of the signal. sensors placed on the right hip, the right wrist, and the left ankle.
The two frequency domain features were: These were the sensor positions for which we had identical
placements compared to the work by Bao and Intille [17].
N spectral centroid All six state-of-the-art approaches used a lower sampling
N bandwidth. frequency than our proposed sampling frequency of 204.8 Hz.
In order to compare our method with the state-of-the-art
Spectral centroid and bandwidth delivered important informa- approaches, the datasets were down-sampled to the sampling
tion about the frequency distribution of the activities [31]. frequencies used in the corresponding approaches. The down-
The single feature that was computed for each sensor type sampling was performed by a linear interpolation method.
(accelerometer or gyroscope) of one sensor node was the energy. Furthermore, the epoch size was set according to the description
The energy for each sensor type was calculated in three steps. of each publication. We implemented the features and the
First, the sum of the squared values for each axis was calculated. classifiers as described in the different approaches. We compared
Second, the three sums were added together and divided by three. our method only to the suggested final feature set and the classifier
Third, this sum was divided by the number of samples. The energy with whom the best classification results were obtained. In order to
gave important information about the activity level of a person. In use the same evaluation process in each approach, a leave-one-
total, this resulted in 152 features for each sliding window. subject-out cross validation was performed in each of the six

PLOS ONE | www.plosone.org 5 October 2013 | Volume 8 | Issue 10 | e75196

 Table 2. Overview of six state-of-the-art approaches in the literature that were implemented and compared in the present study.

Used Sensor/ Sampling Epoch

Year Authors # Sensor/placement placement rate duration Used features Best classifier Evaluation process

2004 Bao and Intille 5 biaxial accelerometers: accelerometers: right hip, 76.25 Hz 6.7 s (50% mean, energy, frequency decision tree classifier leave-one-subject-out-cross-
[17] right hip, right wrist, left right wrist, left ankle overlap) domain entropy, validation procedure
arm, right ankle, correlation of the
left thigh acceleration signals

PLOS ONE | www.plosone.org

2005 Ravi et al. [21] 1 triaxial accelerometer: right hip 50 Hz 5.12 s (50% mean, standard deviation, plurality voting 10-fold cross-validation
pelvic region overlap) energy, correlation of the
accelerometer signals
2006 Karantonis et al. 1 triaxial accelerometer: right hip 45 Hz 1 s (no median filter, low pass hierarchical, threshold none
[23] right hip overlap) filter, normalized signal based classifier
magnitude area, tilt angle
2006 Pärkkä et al. [23] 2 accelerometers: chest chest and right wrist chest: 200 Hz; 1 s (no peak frequency of up- automatically generated leave-one-subject-out-cross-
and wrist (dominant); wrist: 40 Hz overlap) down chest acceleration, decision tree validation procedure
additional signals like median of up-down chest
altitude, ECG, acceleration, peak power
temperature, of up-down chest
and heart rate acceleration, variance of
back-forth chest
acceleration, sum of
variances of three-
dimensional wrist accelerations

6
2009 Preece et al. [20] 3 triaxial accelerometers: right hip, left ankle 64 Hz 2 s (50% magnitude of first five kNN leave-one-subject-out-cross-
waist, thigh, ankle overlap) components of FFT validation procedure
analysis
2012 Liu et al. [19] 2 triaxial accelerometers: right hip and right wrist 30 Hz 30 s (no hip accelerometer: x-axis: SVM with radial basis leave-one-subject-out-cross-
hip (dominant hip) and overlap) standard deviation, 25th function validation procedure
wrist (dominant hand); percentile, and spectral
ventilation sensor entropy; y-axis: spectral
entropy; z-axis: standard
deviation and 90th
percentile; wrist
accelerometer: x-axis: 25th,
50th, 75th 90th percentiles,
standard deviation,
spectral energy and
entropy; y-axis: all time and
frequency domain features,
z-axis: all time domain
features and spectral energy

doi:10.1371/journal.pone.0075196.t002

October 2013 | Volume 8 | Issue 10 | e75196

Activity Classification Using a Benchmark Dataset
 Activity Classification Using a Benchmark Dataset

state-of-the-art approaches, except in the algorithm of Karantonis Table 4. Mean class dependent classification rates (in
et al. [32]. Karantonis et al. did not apply a training step and used percent) for all 13 activities and overall mean classification
predefined and fixed thresholds. rates of proposed system and state-of-the-art systems [17,19–
Each approach in literature used different activities. We 21,23,32].
evaluated all six approaches on all recorded activities in our
work, except for the algorithm of Karantonis et al. [32]. The
hierarchical, threshold based classifier used in Karantonis et al. [17] [21] [32] [23] [20] [19] Proposed
was optimized for fall detection and therefore not applicable for all
SI 83.1 83.7 81.8 65.5 67.7 38.1 88.9
recorded activities in our work. For the other five approaches, this
means, that we investigated activities that were not considered in LY 94.5 87.6 94.7 98.2 100.0 57.4 100.0
the original works. ST 80.7 60.2 64.0 65.6 48.6 44.4 89.8
WD 88.9 79.0 –* 75.4 56.0 89.6 98.1
Results VC 66.9 17.7 –* 36.0 39.2 42.9 85.4

Table 3 shows the overall mean classification rates after leave- SW 81.2 57.8 –* 60.7 62.6 54.3 89.9
one-subject-out procedure. AdaBoost was the best classifier for the WK 96.2 93.0 98.7 74.3 97.6 88.1 99.0
HOUSE system. kNN was the best classifier for the WALK AS 79.5 18.6 –* 28.5 70.6 29.3 95.5
system. SVM was the best classifier for the BASE, REST and DS 73.1 16.4 –* 44.2 60.2 35.3 95.2
BICYCLE system. Table 4 shows the mean class dependent RU 100.0 98.3 –* 92.7 97.2 94.4 100.0
classification rates and the overall mean classification rates of our
BC50 48.1 50.4 –* 40.2 64.3 47.8 69.1
proposed and the compared algorithms [17,19–21,23,32]. The
confusion matrix of our proposed algorithm can be seen in Table 5. BC100 48.5 11.7 –* 46.2 41.7 48.3 53.5
RJ 99.4 93.4 –* 76.6 86.7 33.3 100.0
Discussion mean 80.0 59.1 84.8 61.8 68.7 54.1 89.6

In this paper, we developed a hierarchical classification system Best results are printed bold.
that was able to distinguish between 13 DLAs. Further, we *The algorithm of (Karantonis et al. [32]) was not applied to all activities, as
activity optimized features were used.
compared our proposed method to six state-of-the-art approaches
doi:10.1371/journal.pone.0075196.t004
in the literature [17,19–21,23,32]. In the following, these two
aspects of this study are discussed in detail.
between activity groups and single activities. The grouping
provided the possibility to use different classifier types for different
Subsystems BASE, REST, HOUSE, WALK, and BICYCLE groups of activities. This enhanced the flexibility of the classifica-
We divided our hierarchical classification system into five tion system.
subsystems. The BASE system is the basis for the differentiation in The overall mean classification rates of the REST system
the four subsystems REST, HOUSE, WALK, and BICYCLE and ranged from 92.7% to 97.4% (Table 3). Thus, all classifiers were
the two activities rope jumping and washing dishes. suitable for the distinction between the static activities. Further-
The best classifier for the BASE system was SVM with an more, the results showed that the reduced feature set was suitable
overall mean classification rate of 97.9% (Table 3). SVM is known for this classification task. The best overall mean classification rate
as a classifier with a good generalization performance [34]. The was obtained with the SVM classifier.
overall mean classification rate of the AdaBoost classifier was The overall mean classification rates of the classifiers for the
rather low compared to CART, kNN and SVM. The reason was HOUSE system ranged from 84.0% to 89.9% (Table 3). Although
the low mean class dependent classification rate of the two single the signal patterns of vacuuming and sweeping are similar (Fig. 4),
activities washing dishes and rope jumping, which heavily the proposed feature set was suitable to distinguish between these
decreased the overall mean classification rate of the AdaBoost two activities. The best classifier of the HOUSE system was
classifier. The number of learners seemed to be too low for this AdaBoost. The reason might be that AdaBoost is an ensemble
classification problem. Further research using these activities system, which reduces the variance and increases the confidence of
might take this into account and increase the number of learners. the classifier decision.
All in all, the high maximum overall mean classification rate of The kNN classifier was the best classifier in the case of the
97.9% showed the applicability of the BASE system to distinguish WALK system and reached an overall mean classification rate of
97.7% (Table 3). Thus, walking patterns at different inclinations
Table 3. Mean classification rates (in percent) of the five and speed levels can be distinguished. It is assumed that the
subsystems (BASE, REST, HOUSE, WALK, BICYCLE). gyroscope of the ankle provides useful information about the
inclination, which is also stated in [26]. In order to further improve
the performance of the hierarchical system, the activities walking
ADA CART kNN SVM and running can be grouped, as well as descending and ascending
stairs. The corresponding two new subsystems can be added in our
BASE 64.8 96.1 97.7 97.9 proposed system after the WALK system (Fig. 5).
HOUSE 89.9 84.0 86.5 85.5 The overall mean classification rates of all classifiers in the case
REST 95.1 92.7 96.1 97.4 of the BICYCLE system were rather low compared to the
WALK 94.7 93.3 97.7 94.3 classification rates of the other four subsystems (Table 3). Since the
revolutions per minute were kept constant, it was hard to
BICYCLE 60.8 53.7 49.4 61.6
distinguish between the two resistance levels. The best classifier
Best results are printed bold. was the SVM, which reached an overall classification rate of
doi:10.1371/journal.pone.0075196.t003 61.6% (Table 3).

PLOS ONE | www.plosone.org 7 October 2013 | Volume 8 | Issue 10 | e75196

 Activity Classification Using a Benchmark Dataset

All in all, the results (Table 3) showed that different classifier reason might be that Pärkkä et al. merged sitting and standing to
types achieved the best overall mean classification rate regarding one class. Thus, the used sensor placements were not able to
each of the five subsystems BASE, HOUSE, REST, WALK and distinguish between these two activities.
BICYCLE. SVM was chosen three of five times as the best The algorithm described by Preece et al. [20] reached the
classifier due to the known good generalization performance. As overall mean classification rate of 68.7% (Table 4). Preece et al.
AdaBoost and kNN achieved better results in two subsystems, focused on the comparison of different feature sets optimized for
applying different classifier types for different groups of activities is dynamic activities. They implemented wavelet features of five
therefore mandatory. This endorses that no single classifier is separate studies [35–39], proposed two own wavelet feature sets
suitable for all classification tasks [33]. and compared each wavelet set to seven time and frequency
domain feature sets. Preece et al. obtained the best result with the
Comparison of Proposed System to State-of-the-art feature set of the magnitudes of the first five components of FFT
Algorithms in Literature analysis (Table 2). This feature set was selected due to dynamic
We compared our hierarchical classification system with six activities. This might be the reason that the static activities sitting
state-of-the-art algorithms in literature [17,19–21,23,32]. and standing were classified with only 67.7% and 48.6%,
The algorithm described by Bao and Intille [17] reached an respectively. Preece et al. did not perform a feature selection as
overall mean classification rate of 80.0% (Table 4). The mean class they wanted to compare different feature sets. The results might
dependent classification rates of sitting, standing, and lying increase, if all features (wavelet, time and frequency domain
(Table 4) were smaller compared to the given classification rates features) were combined in one feature set and an appropriate
in [17]. Especially, the mean class dependent classification rates of feature selection procedure was applied to this feature set before
sitting and standing were considerably higher (94.8% and 95.7% in the classification process.
[17] compared to 80.7% and 83.1%). It is assumed that the The algorithm described by Liu et al. [19] reached the overall
additional sensor on the thigh increased the mean classification rates mean classification rate of 54.1% (Table 4). Liu et al. used the
in [17]. On the other side, the mean class dependent classification epoch size of 30 s. This epoch size is not compatible with the
rates of walking and running were higher using our dataset (89.7% duration of our recorded activities. For the activity rope jumping,
and 87.7% in [17] compared to 96.2% and 100.0%). This might be the subjects had to perform five trials with at least five jumps each.
due to the different sensor position at the lower limb. Hence, the duration of this activity was not always 30 s long.
The algorithm described by Ravi et al. [21] reached an overall Consequently, some rope jumping datasets were not used in the
mean classification rate of 59.1% (Table 4). The reason might be classification process which yielded a low classification result of
that only one sensor on the hip was used. Therefore, the 33.3%. Five of our recorded activities lasted for only one minute
classification of activities including upper and lower extremity (Table 1). Since two seconds were cut at the beginning and at the
motions was challenging. This was indicated by the rather low end of the labeled activities, classification of these five activities was
mean class dependent classification rates of the activities vacuum- based on only one epoch, which might not be enough for robust
ing, sweeping, ascending stairs, descending stairs, and bicycling. classification. The activities were correctly classified with mean
The algorithm described by Karantonis et al. [32] reached the class dependent classification rates from 38.1% to 54.3%. We
best mean class dependent classification rate for walking (Table 4). included three household activities in our study setup. Two of
The reason is that besides the detection of the postural orientation them were only recorded for one minute, and hence, classification
(tilt angle feature), Karantonis et al. used an optimized algorithm rates of only 42.9% and 54.3% (Table 4) were obtained. The third
for walking. All used features were matched to their performed household activity (washing dishes) was classified with 89.6%. This
activities (mainly transitions between standing, lying, and sitting might be because this activity was performed for two minutes
and different fall scenarios). The mean class dependent classifica- (Table 1), which might lead to better performance in [19]. Liu et al.
tion rates of sitting and standing were rather low compared to [19] performed a two-step feature selection. First, they performed a
lying and walking (Table 4). It is assumed that instances of sitting statistical analysis that was followed by the minimal-redundancy-
were misclassified as standing and vice versa, which was also maximal-relevance heuristic [40]. This resulted in a specialized
mentioned in [32]. The focus of Karantonis et al. was to detect feature set (Table 2). The problem with specialized feature sets is
possible falls and hence, this misclassification is not severe. Their that they might not be applicable to all activities. This might be the
focus was a real-time implementation for ambulatory monitoring. reason why only three activities were classified with higher than
Their algorithm could only be applied to a subset of our recorded 80% and activities that were not considered in the original study
activities. Therefore, it is difficult to compare the overall setup [19] like ascending and descending stairs were classified with
performance to the other approaches and our proposed algorithm. low classification rates of 29.3% and 35.3%, respectively.
The algorithm described by Pärkkä et al. [23] reached a rather Our proposed method reached the overall mean classification rate
low overall mean classification rate of 61.8% compared to the of 89.6% (Table 4). It is assumed that different number of sensors
other algorithms (Table 4). The reason might be that only two due to different study setups and different identical sensor positions
sensors (one sensor on the chest and one sensor on the wrist) were influenced the results. We used four sensor positions and compared
used. Therefore, as mentioned before, the classification of activities these four sensor positions to one sensor position [21,32], two sensor
including lower extremity motions was challenging. This is positions [20,23], and three sensor positions [17,19]. Thus, by using
indicated by the low mean class dependent classification rates of more sensors more complex activities can be classified.
ascending stairs, descending stairs, and bicycling. Although We suggest to use sensors near the body’s center of mass (hip
vacuuming includes motions that should be recognized by the and chest) in order to cover a wide range of basic activities such as
chest and wrist sensor, the mean class dependent classification rate sitting, standing, lying, and walking. Nevertheless, the mean class
was low. It is assumed that instances of vacuuming were dependent classification rates of sitting and standing were smaller
misclassified as sweeping, whose signal patterns are similar to compared to lying (Table 4). This trend coincides with the results
signal patterns of vacuuming (Fig. 4). Regarding the discrimination of the other approaches in Table 4. It is assumed that an
of the static activities, the mean class dependent classification rate additional sensor on the thigh improves the performance of the
of lying was rather high compared to sitting and standing. The classification system.

PLOS ONE | www.plosone.org 8 October 2013 | Volume 8 | Issue 10 | e75196

 Activity Classification Using a Benchmark Dataset

Moreover, we suggested to use sensors on lower and upper approach by Bao and Intille seemed to be applicable for a large
extremities (wrist and ankle) to distinguish between more complex set of DLAs, especially under semi-naturalistic conditions.
activities like ascending stairs, descending stairs, vacuuming, and
sweeping. Nevertheless, vacuuming was often misclassified as Compared to Bao and Intille, we only collected data under
sweeping and vice versa (Table 5). In this case, an additional controlled conditions. Thus, data under more realistic conditions
sensor on the other wrist might incorporate additional information would be desirable and it is planned to integrate these more
into the classification system. Descending and ascending stairs realistic conditions in our work.
were often misclassified as walking (Table 5). In order to reduce However, our proposed method showed that the classification of
the misclassification, specialized gait features might improve the DLAs can benefit from sensor data fusion of accelerometer and
results for the WALK system. Nevertheless, given the flexibility of gyroscope (Table 4). It is assumed that especially the gyroscope
the proposed classification system, the incorporation of these ideas improves the classification of activities which include rotational
is straightforward. movements like washing dishes, ascending stairs or rope jumping.
Most approaches found in literature only used accelerometers.
The high mean class dependent classification rates of washing
Nevertheless, Lee and Mase [24] and Najafi et al. [41] used the
dishes (98.1%) and rope jumping (100.0%) (Table 4) showed again
combination of accelerometer and gyroscope for the classification
the applicability of the BASE system to distinguish between single
of body postures and walking behaviors like ascending stairs or
activities and activity groups.
walking. However, the classification systems were optimized for a
The high overall mean classification rate of the BASE system
certain subset of activities and cannot be applied to the recorded
(Table 3) showed that it is possible to classify the merged activity activities in our work. Koskimaki et al. [42] used the combination
group of the two resistance levels of bicycling. The rather low of accelerometer and gyroscope for the classification of activities of
mean class dependent classification rates of the two resistance workers on industrial assembly lines. Altun and Barshan [43] used
levels of bicycling 69.1% and 53.5% (Table 4) showed the the combination of accelerometer and gyroscope for the classifi-
challenge to distinguish between the single resistance levels. The cation of nineteen DLAs. Koskimaki et al. [42] and Altun and
reason is that the low resistance level was misclassified as the high Barshan [43] computed features which were not optimized for a
resistance level and vice versa, which is confirmed by observations certain subset of activities and therefore, it is planned to compare
in the confusion matrix. Bicycling with the lower resistance level our approach with [42] and [43], too.
(50 W) was correctly classified in 620 cases and misclassified as For the performance assessment, a leave-one-subject-out
bicycling with the higher resistance level (100 W) in 250 cases procedure was applied. This procedure results in a small bias
(Table 5). Bicycling with the higher resistance level was correctly and a large variance of the true error rate estimator [44]. Due to
classified in 480 cases and misclassified as bicycling with the lower the sample size, the leave-one-subject-out procedure was preferred
resistance level in 410 cases. to for example a 10-fold cross-validation, which shows a good
The algorithm described by Bao and Intille [17] was the best result performance for a larger sample size.
that we obtained with an algorithm used for comparison that we Since, multiple subjects perform multiple activities in different
applied to all performed activities. The reasons might be the following: ways, there might be a high intersubject variability. In order to
setup a generalized system that shows good performance for an
N The study design of Bao and Intille and our study design had unknown subject, the classification systems were trained based on
three sensor positions in common. In the other approaches, at multiple subjects.
maximum two sensor positions were identical. All in all, our hierarchical, multi-sensor based classification
N The approach by Bao and Intille was applied on twenty system had problems in the distinction of the different activities of
activities under both controlled conditions in the laboratory and the HOUSE and BICYCLE systems. Both systems have to be
semi-naturalistic conditions outside the laboratory. Thus, the improved. The activities sitting and lying were classified with mean

Table 5. Confusion matrix of our proposed algorithm. Each entry represents the number of classified epochs.

SI LY ST WD VC SW WK AS DS RU BC50 BC100 RJ

SI 383 0 22 14 1 2 0 0 2 0 7 0 0
LY 0 435 0 0 0 0 0 0 0 0 0 0 0
ST 0 0 386 27 9 0 0 0 0 0 8 0 0
WD 0 0 10 896 5 1 0 0 0 0 1 0 0
VC 0 0 0 0 369 55 0 0 0 0 6 2 0
SW 0 0 1 6 43 643 4 7 5 0 4 2 0
WK 0 0 0 0 4 4 1995 5 7 0 0 0 0
AS 0 0 0 0 0 3 10 279 0 0 0 0 0
DS 0 0 0 0 0 1 11 0 237 0 0 0 0
RU 0 0 0 0 0 0 0 0 0 886 0 0 0
BC50 0 0 0 0 4 22 0 1 0 0 620 250 0
BC100 0 0 0 1 0 6 0 0 0 0 410 480 0
RJ 0 0 0 0 0 0 0 0 0 0 0 0 236

The confusion matrices of each leave-one-subject-out trial were summed up.

doi:10.1371/journal.pone.0075196.t005

PLOS ONE | www.plosone.org 9 October 2013 | Volume 8 | Issue 10 | e75196

 Activity Classification Using a Benchmark Dataset

class dependent classification rates below 90%. An additional first step is to provide feedback about the individual quality of life.
sensor on the thigh could increase these classification rates. We In this field, the classification of DLAs is of major interest.
calculated 152 features for each sliding window for each In this paper, a novel, hierarchical, multi-sensor based
classification system (except for the REST system). A high number classification system was developed, which reached an over all
of features leads to high computational complexity in real-time mean classification rate of 85.8%. We considered the classification
applications or in embedded systems. Hence, an automatic of 13 DLAs. Furthermore, our proposed system was compared to
reduction of this feature set like sequential forward selection [34] state-of-the-art algorithms in literature using the same dataset. The
should be applied. Nevertheless, the overall mean classification comparison showed that the proposed data fusion of accelerometer
rate of 89.6% showed the applicability of our proposed system to and gyroscope provided a useful tool to distinguish between
classify the acquired 13 activities. complex activities like ascending stairs or descending stairs.
The hierarchical structure of our proposed system has four A multitude of activity classification systems has been proposed
advantages: in literature, and to date it is not clear which solution is
outperforming the others and is applicable to a variety of real
N Different classifiers can be used for the classification of world scenarios. It is mandatory for the community to provide
different activity groups. benchmark datasets and reference implementations. This will help
N Additional activities can easily be integrated without retraining to speed up the process of getting the best performing and most
the complete system. appropriate DLA classification system into much needed real
N In many applications, in which the further classification of the world applications. We are inviting fellow scientists to share their
data and implementations on our newly erected internet platform
activities in the activity groups is not necessary, the HOUSE,
REST, WALK, and BICYCLE system can easily be neglected. (http://www.activitynet.org).
N A different window size can be chosen for the BASE, HOUSE,
Acknowledgments
REST, WALK, and BICYCLE system, which might increase
the classification rate. We thank all subjects for participating in the study.

Author Contributions
Conclusion
Conceived and designed the experiments: HL DS BME. Performed the
Physically inactive people have to be motivated to be more experiments: HL DS. Analyzed the data: HL DS. Contributed reagents/
active so that their risk of various chronic diseases will decrease. A materials/analysis tools: HL DS. Wrote the paper: HL DS BME.

References
1. World Health Organization (http://www.who.int/gho/ncd/risk_factors/ 16. Trost SG, McIver KL, Pate RR (2005) Conducting accelerometer-based activity
physical_activity_text/en/).Prevalence of insufficient physical activity. Accessed assessments in field-based research. Med Sci Sports Exerc 37 (supplement 11):
on August 19th 2013. 531–543.
2. Pate RR, Pratt M, Blair SN, Haskell WL, Macera CA, et al. (1995) Physical 17. Bao L, Intille SS (2004) Activity recognition from user-annotated acceleration
activity and public health: A recommendation from the centers for disease data. In: Ferscha A, Mattern F, editors, Pervasive Computing, Springer Berlin
control and prevention and the american college of sports medicine. J Am Med Heidelberg, volume 3001 of Lecture Notes in Computer Science. pp. 1–17.
Assoc 273: 402–407. 18. Ermes M, Pärkkä J, Mäntyjärvi J, Korhonen I (2008) Detection of daily activities
3. Ogden C, Carroll M, Curtin L, McDowell M, Tabak C, et al. (2006) Prevalence and sports with wearable sensors in controlled and uncontrolled conditions.
of overweight and obesity in the united states, 1999–2004. J Am Med Assoc 295: Trans Info Tech Biomed 12: 20–26.
1549–1555. 19. Liu S, Gao RX, John D, Staudenmayer JW, Freedson PS (2012) Multisensor
4. Wing R (1999) Physical activity in the treatment of the adulthood overweight data fusion for physical activity assessment. IEEE Trans Biomed Eng 59: 687–
and obesity: current evidence and research issues. Med Sci Sports Exerc 31: 696.
5547–5552. 20. Preece SJ, Goulermas JY, Kenney LPJ, Howard D (2009) A comparison of
5. Physical Activity Guidelines Advisory Committee (2008). Physical activity feature extraction methods for the classification of dynamic activities from
guidelines advisory committee report, 2008. accelerometer data. IEEE Trans Biomed Eng 56: 871–879.
6. Blair SN, Cheng Y, Holder JS (2001) Is physical activity or physical fitness more 21. Ravi N, Dandekar N, Mysore P, Littman ML (2005) Activity recognition from
important in defining health benefits? Med Sci Sports Exerc 33 (supplement 6): accelerometer data. In: Proc. 17th Conf. Innovative Applications Artificial
379–399. Intelligence (IAAI). pp. 1541–1546.
22. Ward DS, Evenson KR, Vaughn A, Rodgers AB, Troiano RP (2005)
7. Lee IM, Paffenbarger RSJ (2000) Associations of light, moderate, and vigorous
Accelerometer use in physical activity: best practices and research recommen-
intensity physical activity with longevity. The harvard alumni health study.
dations. Med Sci Sports Exerc 37 (supplement 11): 582–588.
Am J Epidemiol 151: 293–299.
23. Pärkkä J, Ermes M, Korpipää P, Mäntyjärvi J, Peltola J, et al. (2006) Activity
8. Oguma Y, Sesso H, Paffenbarger R, Lee I (2002) Physical activity and all cause
classification using realistic data from wearable sensors. IEEE Trans Inf Technol
mortality in women: a review of the evidence. Br J Sports Med 36: 162–172.
Biomed 10: 119–128.
9. Berlin JA, Colditz GA (1990) A meta-analysis of physical activity in the
24. Najafi B, Aminian K, Paraschiv-Ionescu A, Loew F, Bla CJ, et al. (2003)
prevention of coronary heart disease. Am J Epidemiol 132: 612–628. Ambulatory system for human motion analysis using a kinematic sensor:
10. Powell KE, Thompson PD, Caspersen CJ, Kendrick JS (1987) Physical activity Monitoring of daily physical activity in the elderly. IEEE Trans Biomed Eng 50:
and the incidence of coronary heart disease. Annu Rev Public Health 8: 253– 711–723.
287. 25. Mariani B, Hoskovec C, Rochat S, Bla C, Penders J, et al. (2010) 3D gait
11. Warburton DER, Nicol CW, Bredin SSD (2006) Health benefits of physical assessment in young and elderly subjects using foot-worn inertial sensors.
activity: the evidence. CMAJ 174: 801–809. J Biomech 43: 2999–3006.
12. Warren JM, Ekelund U, Besson H, Mezzani A, Geladas N, et al. (2010) 26. Schuldhaus D, Kugler P, Leible M, Jensen U, Schlarb H, et al. (2012)
Assessment of physical activity - a review of methodologies with reference to Classification of surfaces and inclinations during outdoor running using shoe-
epidemiological research: a report of the exercise physiology section of the mounted inertial sensors. In: Proc. 21st International Conf. Pattern Recognition
european association of cardiovascular prevention and rehabilitation. (ICPR). pp. 2258–2261.
Eur J Cardiov Prev R 17: 127–139. 27. Adams R (1999) Revised physical activity readiness questionnaire. Can Fam
13. Shephard RJ (2003) Limits to the measurement of habitual physical activity by Physician 45: 992–995.
questionnaires. Br J Sports Med 37: 197–206. 28. Ainsworth BE, Haskell WL, Leon AS, Jacobs DR, Montoye HJ, et al. (1993)
14. Mannini A, Sabatini M (2010) Machine learning methods for classifying human Compendium of physical activities: Classification of energy costs of human
physical activity from on-body accelerometers. Sensors 10: 1154–1175. physical activities. Med Sci Sports Exerc 25: 71–80.
15. Burns A, Greene BR, McGrath MJ, O’Shea TJ, Kuris B, et al. (2010) Shimmer - 29. Ainsworth BE, Haskell WL, Whitt MC, Irwin ML, Swartz AM, et al. (2000)
a wireless sensor platform for noninvasive biomedical research. IEEE Sensors J Compendium of physical activities: an update of activity codes and met
10: 1527–1534. intensities. Med Sci Sports Exerc 32 (supplement 9): 498–516.

PLOS ONE | www.plosone.org 10 October 2013 | Volume 8 | Issue 10 | e75196

 Activity Classification Using a Benchmark Dataset

30. Ainsworth BE, Haskell WL, Herrmann SD, Meckes N, Bassett DJJ, et al. (2011) 38. Sekine M, Tamura T, Akay M, Fujimoto T, Togawa T, et al. (2002)
2011 compendium of physical activities: a second update of codes and met Discrimination of walking patterns using wavelet-based fractal analysis. IEEE
values. Med Sci Sports Exerc 43: 1575–1581. Trans Neural Syst Rehabil Eng 10: 188–196.
31. Agostini G, Longari M, Pollastri E (2003) Musical instrument timbres 39. Sekine M, Tamura T, Togawa T, Fukui Y (2000) Classification of waist-
classification with spectral features. EURASIP J Adv Sig Pr 2003: 5–14. acceleration signals in a continuous walking record. Med Eng Phys 22: 285–291.
32. Karantonis DM, Narayanan MR, Mathie M, Lovell NH, Celler BG (2006) 40. Peng H, Long F, Ding C (2005) Feature selection based on mutual information
Implementation of a realtime human movement classiffer using a triaxial criteria of max-dependency, max-relevance, and min-redundancy. IEEE Trans
accelerometer for ambulatory monitoring. IEEE Trans Inf Technol Biomed 10: Pattern Anal Mach Intell 27: 1226–1238.
156–167. 41. Lee SW, Mase K (2002) Activity and location recognition using wearable
33. Duda R, Hart P, Stork D (2000) Pattern Classification. New York, NY: Wiley- sensors. IEEE Pervasive Comput 1: 24–32.
Interscience, 2nd edition. 42. Koskimaki H, Huikari V, Siirtola P, Laurinen P, Roning J (2009) Activity
34. Theodoridis S, Koutroumbas K (2008) Pattern Recognition. San Diego, CA: recognition using a wrist-worn inertial measurement unit: A case study for
Academic Press, 4th edition. industrial assembly lines. In: Proc. 17th Mediterranean Conf. Control and
35. Tamura T, Sekine M, Ogawa M, Togawa T, Fukui Y (1997) Classification of
Automation (MED). pp. 401–405.
acceleration waveforms during walking by wavelet transform. Method Inform
43. Altun K, Barshan B (2010) Human activity recognition using inertial/magnetic
Med 36: 356–359.
sensor units. In: Salah A, Gevers T, Sebe N, Vinciarelli A, editors, Human
36. Nyan MN, Tay FEH, Seah KHW, Sitoh YY (2006) Classification of gait
Behavior Understanding, Springer Berlin Heidelberg, volume 6219 of Lecture
patterns in the time-frequency domain. J Biomech 39: 2647–2656.
37. Wang N, Ambikairajah E, Lovell NH, Celler BG (2007) Accelerometry based Notes in Computer Science. pp. 38–51.
classification of walking patterns using time-frequency analysis. In: Proc. 29th 44. Witten IH, Frank E, Hall MA (2011) Data Mining - Practical Machine Learning
Annual International Conf. IEEE Engineering Medicine and Biology Society Tools and Techniques. Elsevier.
(EMBS). pp. 4899–4902. doi:10.1109/IEMBS.2007.4353438.

PLOS ONE | www.plosone.org 11 October 2013 | Volume 8 | Issue 10 | e75196