See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224345069

A Comparison of Feature Extraction Methods for the Classification of

Dynamic Activities From Accelerometer Data

Article  in  IEEE transactions on bio-medical engineering · April 2009

DOI: 10.1109/TBME.2008.2006190 · Source: IEEE Xplore

CITATIONS READS
486 6,073

4 authors, including:

Stephen J. Preece Laurence Paul Joseph Kenney

University of Salford University of Salford
89 PUBLICATIONS   2,519 CITATIONS    148 PUBLICATIONS   2,981 CITATIONS   

SEE PROFILE SEE PROFILE

David Howard
University of Salford
103 PUBLICATIONS   2,164 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

upper limb FES View project

Running View project

All content following this page was uploaded by Laurence Paul Joseph Kenney on 05 June 2014.

The user has requested enhancement of the downloaded file.

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 56, NO. 3, MARCH 2009 871

A Comparison of Feature Extraction Methods

for the Classification of Dynamic Activities
From Accelerometer Data
Stephen J. Preece*, John Yannis Goulermas, Member, IEEE, Laurence P. J. Kenney, and David Howard

Abstract—Driven by the demands on healthcare resulting from to play an important role in large-scale epidemiological studies
the shift toward more sedentary lifestyles, considerable effort has in this area [6], [7]. Furthermore, such systems can also be used
been devoted to the monitoring and classification of human activity. to assess the effectiveness of different interventions aimed at
In previous studies, various classification schemes and feature ex-
traction methods have been used to identify different activities from increasing levels of physical activity and for motivating individ-
a range of different datasets. In this paper, we present a comparison uals to become more physically active.
of 14 methods to extract classification features from accelerome- The success of a given rehabilitation program is often judged
ter signals. These are based on the wavelet transform and other by not only the levels of activity but also the type of activ-
well-known time- and frequency-domain signal characteristics. To ity that an individual can return to after treatment. In addition,
allow an objective comparison between the different features, we
used two datasets of activities collected from 20 subjects. The first as fall risk increases with age, so a better understanding of the
set comprised three commonly used activities, namely, level walk- factors contributing to fall risk becomes more important. Ambu-
ing, stair ascent, and stair descent, and the second a total of eight latory monitoring of various activities, including the time spent
activities. Furthermore, we compared the classification accuracy in sit–stand transitions, has shown promise as a predictor of
for each feature set across different combinations of three differ- fall risk [8]. Furthermore, both type and intensity of individ-
ent accelerometer placements. The classification analysis has been
performed with robust subject-based cross-validation methods us- ual’s activity are of interest to urban designers, and designers,
ing a nearest-neighbor classifier. The findings show that, although manufacturers, and purchasers of certain medical devices (e.g.,
the wavelet transform approach can be used to characterize non- advanced responsive pacemakers and orthopedic implants).
stationary signals, it does not perform as accurately as frequency- In addition to health-related applications, portable systems,
based features when classifying dynamic activities performed by which can accurately identify the activity of the user, have the
healthy subjects. Overall, the best feature sets achieved over 95%
intersubject classification accuracy. potential to play a fundamental role in a ubiquitous computing
scenario [9], [10]. In this field, computing devices use infor-
Index Terms—Activity classification, ambulatory monitoring, mation from a variety of sensors to determine the context of a
machine learning, wavelet transform.
situation. Different devices can then use the context information
to deliver an appropriate service. For example, a mobile phone
I. INTRODUCTION may detect when a person is driving a vehicle and automatically
VER THE past decade, there has been considerable re- divert a call.
O search effort directed toward the monitoring and clas-
sification of physical activity patterns from body-fixed sensor
With recent advances in miniaturized sensing technology, it
is now possible to collect and store acceleration data from indi-
data [1], [2]. This has been motivated by a number of impor- vidual body segments over extended periods of time. Although
tant health-related applications. For example, with the trend this technology offers the ideal platform for monitoring daily
toward more sedentary lifestyles, there is growing interest in activity patterns, effective algorithms are also required to inter-
the link between levels of physical activity and common health pret the accelerometer data in the context of different activities.
problems, such as diabetes, cardiovascular disease, and osteo- Previous studies have shown machine learning or artificial in-
porosis [3]. As self-reported measures have been shown to be telligence approaches to be effective for identifying a range of
unreliable [4], [5], systems for activity profiling are beginning different activities from body-fixed sensor data [11]–[14]. These
techniques typically operate via a two-stage process [15]. First,
Manuscript received April 1, 2008; revised May 31, 2008 and August 1, features are derived from windows of accelerometer data. A clas-
2008. First published October 31, 2008; current version published April 15, sifier is then used to identify the activity corresponding to each
2009. This work was supported by the EU Framework VI under Contract IST- separate window of data. A range of different approaches has
2002-1-001837 (Healthy AIMS). Asterisk indicates corresponding author.
∗ S. J. Preece is with the Centre for Rehabilitation and Human Perfor- been used to obtain features from accelerometer data, with some
mance Research, University of Salford, Salford M6 6PU, U.K. (e-mail: researchers deriving features directly from the time-varying ac-
[email protected]). celeration signal [12], [16]–[18] and others from a frequency
J. Y. Goulermas is with the Department of Electrical Engineering and
Electronics, University of Liverpool, Liverpool L69 3BX, U.K. (e-mail: analysis [11], [13], [19], [20]. More recently, wavelet analy-
[email protected]). sis has been used to derive the so-called time-frequency fea-
L. P. J. Kenney and D. Howard are with the Centre for Rehabilitation and tures [14], [21]–[24].
Human Performance Research, University of Salford, Salford, M6 6PU, U.K.
(e-mail: [email protected]; [email protected]). With wavelet analysis, the original signal is decomposed into
Digital Object Identifier 10.1109/TBME.2008.2006190 a series of coefficients, which carry both spectral and temporal

0018-9294/$25.00 © 2009 IEEE

Authorized licensed use limited to: UNIVERSITY OF SALFORD. Downloaded on May 21, 2009 at 04:12 from IEEE Xplore. Restrictions apply.
872 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 56, NO. 3, MARCH 2009

information about the original signal. From these coefficients,

it is possible to identify localized temporal instances at which
there is a change in frequency characteristics of the original
signal [25]. This concept has been applied successfully to ac-
celerometer signals in order to identify points in the signal at
which the subject changes from one activity to another [22], [24].
As well as being used to locate discrete temporal events, wavelet
analysis can also be used to derive time-frequency features that
characterize the original signal. However, it is not clear whether
such time-frequency features lead to more effective activity
classification than the more commonly used time-domain or
frequency-domain features.
The overall aim of this study was to extensively compare
the performance of a number of previously reported and novel
wavelet features with a range of time-domain and frequency-
domain features for the classification of different activities.
Many previous wavelet-based studies have investigated level Fig. 1. Activity monitors attached to the ankle and thigh.
walking, stair ascent, and stair descent [21]–[23], but have not
compared their performance against simpler approaches. There- attachment has been illustrated in Fig. 1 for the ankle and thigh
fore, our first research aim was to compare features for this three- placements.
activity classification problem. As a second aim, we sought to Ten male and ten female subjects participated in the study. As
compare the same features for a larger set of activities that large individual variation has been reported for accelerometer
represents a more challenging problem. Additionally, since the signals corresponding to the same activity [27], subjects with
performance of a given set of features can be dependent on the a range of ages and body mass indices were recruited into the
location of the sensor, we compared accuracy for the different study. The mean [standard deviation (SD)] age of the subjects
features across a number of different lower limb placements. was 31 (7) years, mean (SD) height was 1.71 (0.07) m, and the
It was felt that this study would underpin the development of mean (SD) weight was 68 (10) kg. The subjects covered a wide
an off-the-shelf activity monitor that could be used to classify range of body mass indices from 19 to 30 with mean (SD) 24
activity patterns across different subjects. (3). Each subject gave informed consent to participate in the trial
after approval had been obtained from the Ethical Committee at
II. METHOD the University of Salford.
A number of studies have shown that static postures can
A. Data Collection be differentiated from dynamic activity by applying a single
Accelerometer data were collected using Pegasus activity threshold to some measure of acceleration variability [28], [29].
monitors developed by ETB, U.K. Each of these units contained Provided sensors are attached to more than one body segment,
a triaxial accelerometer, with a dynamic range of ±5 g, which it is possible to accurately identify different static postures us-
was sampled with 10-bit resolution. With these devices, it is pos- ing a threshold-based approach [30], [31]. However, the situ-
sible to sample accelerometer data at a user-defined frequency ation is more complicated with only a single sensor. In this
and to store these data for up to 24 h. A sampling frequency of scenario, more complex signal processing along with an appro-
64 Hz was selected for this study as this is sufficiently higher priate biomechanical model is required to differentiate between
than the 20 Hz sampling required to assess daily activity [26]. different postures, postural transitions, and continuous dynamic
A number of previous activity classification studies have used activity [32]. For this study, we chose to investigate the clas-
wavelet analysis to derive features from accelerometer signals sification of continuous dynamic activities. This choice was
collected at relatively high sampling frequencies (>250 Hz). motivated by previous studies that have used a range of differ-
However, for this study, 64 Hz was chosen as this is a realistic ent features to characterize acceleration signals [11], [14], [16],
sampling frequency that could be implemented by an off-the- [17], [21]–[24], [33]–[35].
shelf activity monitor. No antialiasing filtering was applied to Subjects completed a total of eight different activities (level
the acceleration data. walking, walking upstairs and downstairs, jogging, running,
For each subject, data were collected with three activity moni- hopping on the left and right leg, and jumping). Each of these
tors. These were attached to waist (at the sacrum), the thigh (just activities was part of a continuous circuit that started in a build-
above the knee), and the ankle (just above the lateral maleollus). ing and then followed a route around the university campus.
To secure each unit in place, specialized bandage (Fabrifoam) This circuit was described to the subjects before the start of
was first positioned around each of the body segments and the data collection. During the trial, the experimenter recorded the
activity monitors, which were backed with Velcro, adhered to sequence of activities with a portable video camera giving mini-
the underwrapped bandage. Once in position, additional ban- mal prompting to the subject. With this design, the subjects were
dage was then wrapped over each sensor to ensure no move- free to move at their preferred pace and to transition between
ment could occur from the overlying clothing. This method of different activities when they felt most comfortable.

Authorized licensed use limited to: UNIVERSITY OF SALFORD. Downloaded on May 21, 2009 at 04:12 from IEEE Xplore. Restrictions apply.
PREECE et al.: COMPARISON OF FEATURE EXTRACTION METHODS FOR THE CLASSIFICATION OF DYNAMIC ACTIVITIES 873

et al. [24] (2 s) and Wang et al. [14] (2.56 s) that had used sim-
ilar length windows. It was not possible to use shorter windows
as signals with less than 128 samples could not be fully de-
composed into wavelet coefficients appropriate for comparison
with other studies (see Section 2.2). Longer windows limited the
amount of data that could be extracted from short-duration ac-
tivities, such as stair walking. Pilot work also showed minimal
differences between classification accuracies calculated from
frequency-domain features derived from 2 s or 3 s windows.
If a window corresponded to a transition between two ac-
tivities, it was excluded from subsequent analysis. Given the
continuous nature of the circuit completed by the subjects, there
was a disproportionate number of windows of data that cor-
responded to level walking. Therefore, in order to balance the
Fig. 2. GUI used to annotate the accelerometer data from the video record. distribution of the different activities, only a randomly chosen
subset of these windows was used in the final analysis.
To ensure that there were sufficient data to address the first
research question, the circuit involved stair walking both inside
and outside the building as well as level walking in a number of B. Wavelet Features
different environments. In addition to these three everyday activ- A number of previous activity classification studies have de-
ities, both jogging and running were also included in the circuit. rived time-frequency features obtained using the filter bank in-
For the first of these two activities, subjects were instructed to terpretation of the discrete wavelet transform [22], [24]. With
perform a gentle jog over a 50 m distance, and for the second to this approach, the original time-domain signal (maximum fre-
perform a fast run over the same distance. Both these activities quency f ) is initially decomposed into a coarse approximation
have been used in previous classification studies [11], [12], [36] and detail information by low-pass filtering (bandpass [0, f /2])
and their recognition could prove invaluable in any activity mon- and high-pass filtering (bandpass [f /2, f ]), respectively [39].
itoring system for sports rehabilitation. We wanted to collect With wavelet decomposition, the half-band filters are designed
data across a range of different modes of locomotion and there- to enable perfect reconstruction of the original signal and to
fore included three additional activities: hopping (on each leg) avoid aliasing effects. In subsequent levels of decomposition,
and jumping. Both hopping [37] and jumping [36] have been the approximation signal from the previous level is split into a
used in previous activity monitoring studies and are also used in second approximation and a detail coefficient. This process is
sports rehabilitation. In order to include each of these activities repeated to the desired decomposition level. For further details,
as part of the circuit, each subject was required to hop (on each see [40]. Five separate studies were identified, which had previ-
leg separately) over a 15 m distance and to jump, moving both ously used wavelet features for classification of accelerometer
legs together, over the same distance. data [14], [21]–[24]. These studies were then used as a basis for
Just prior to data collection, the three activity monitoring defining seven sets of wavelet features (Table I).
units were synchronized with each other and with the clock of The first set of wavelet features was proposed by Tamura
a laptop computer. This procedure was repeated at the end of et al. [23]. With this approach, the accelerometer signal is de-
each experiment to ensure that the units had not drifted relative composed using the wavelet transform and the features defined
to each other. Resynchronization was not needed as the units as signal power measurements, calculated as the sum of the
never drifted by more than three to four samples (0.05 s). Cus- squared detail coefficients at levels 4 and 5. Tamura et al. [23]
tom software was developed in Matlab (The MathWorks, USA) sampled acceleration data at 250 Hz. Given our lower sampling
so that the video data could be synchronized with the laptop frequency of 64 Hz, we calculated the two features from detail
and thus the accelerometer data. Following data collection, this coefficients corresponding to the same frequency bands as those
software was used to annotate the accelerometer data with the used by Tamura et al. [23]. This process of identifying corre-
transition points between each of the different activities (see sponding wavelet coefficients for our lower sampling frequency
Fig. 2). This method allowed for rapid and accurate labeling of was performed for other wavelet feature sets where needed.
the data, which was particularly important for identifying stair The second set of features was taken from Nyan et al. [24].
ascent and descent as these activities only lasted approximately These features are calculated in a similar way to Tamura et al.
10 s. A small pilot study demonstrated minimal (<1 s) intertester [23]; however, rather than treating the scales separately, the
variability in the identification of the activity transition points. summations at levels 4 and 5 are added together. Features sug-
Once activity transition points had been identified, features gested by Sekine et al. [22] form the basis of the third set of
were calculated from 2-s (128-sample) consecutive windows features. Again, there are two features, the first being the total
that overlapped by 1 s. The use of a 50% overlap between suc- of the summations of the detail signal at levels 6 and 7. This
cessive sliding windows has been shown to be effective in pre- quantity is divided by the number of steps (N ), which is ob-
vious studies of activity classification [11], [38]. The choice tained by counting the number of times the signal, reconstructed
of a 2 s window was motivated by previous studies by Nyan from levels 6 and 7, changes sign. For the second feature, the

Authorized licensed use limited to: UNIVERSITY OF SALFORD. Downloaded on May 21, 2009 at 04:12 from IEEE Xplore. Restrictions apply.
874 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 56, NO. 3, MARCH 2009

TABLE I
SUMMARY OF THE DIFFERENT WAVELET FEATURES

total of the summations of the detail signal from levels 4 to 7 quency used by Sekine et al. [21] (1024 Hz), they were able
is normalized against the sum of the squares from the original to calculate the fractal dimension from the variance of the de-
signal. Although [22] used a Coiflet wavelet mother for wavelet tail coefficients across seven different levels. Due to our lower
decomposition, our preliminary investigation showed improved sampling frequency of 64 Hz, fractal dimension was estimated
classification with a Daubechies wavelet mother. This was there- from variance progression across three levels. Although this
fore used for subsequent analysis. Both [24] and [22] collected may lead to poorer discriminate ability for this feature set, the
data at 256 Hz. use of additional detail coefficients was not possible with our
Most previous activity classification studies have used lower sampling frequency.
wavelet analysis to derive only a small number of features. In In addition to the five sets of wavelet features described ear-
contrast, Wang et al. [14] used wavelet packet analysis to derive lier, we experimented with some alternative wavelet features.
33 features from a triaxial accelerometer signal. With wavelet Two additional feature sets were then included in this study
packet analysis, the detail coefficients are split into a further (Table I). For both of these feature sets, each component of the
approximation and detail coefficients. This allows additional in- 64-Hz triaxial acceleration signal was decomposed to five levels
formation to be extracted from the original signal. The features using a Daubechies 2 wavelet mother. A sixth wavelet feature
suggested by Wang et al. [14] involved summing the squares of set was then defined as the sum of the squared detail coefficients
the detail and wavelet packet approximation coefficients across at levels 1–5. These five features were calculated for each com-
different levels. In addition, they calculated standard deviations ponent of acceleration, thus giving a total of 15 features. The
and rms values of detail and wavelet packet approximation co- seventh feature set was obtained in a similar way, but the sums
efficients at a number of different levels. In their study, Wang of the absolute values were used to provide a different type of
et al. [14] sampled accelerometer data at 50 Hz; therefore, our combining norm. All wavelet features in Table I were derived
data were resampled to this frequency. for every window of accelerometer data using Matlab ver.7.4
The fifth set of features is based on the concept of fractal (The Mathworks, USA).
dimension, which was used by Sekine et al. [21] to character-
ize accelerometer signals. The fractal dimension quantifies the
variance progression of the detail coefficient over the different C. Time- and Frequency-Domain Features
wavelet scales and as such gives a measure of the complexity For additional comparison, we also employed three sets
within the original signal [41]. Given the high sampling fre- of time-domain features and four sets of frequency-domain

Authorized licensed use limited to: UNIVERSITY OF SALFORD. Downloaded on May 21, 2009 at 04:12 from IEEE Xplore. Restrictions apply.
PREECE et al.: COMPARISON OF FEATURE EXTRACTION METHODS FOR THE CLASSIFICATION OF DYNAMIC ACTIVITIES 875

TABLE II TABLE III

SUMMARY OF THE TIME- AND FREQUENCY-DOMAIN FEATURES CLASSIFICATION ACCURACIES (%) FOR THE THREE-ACTIVITY CLASSIFICATION
PROBLEM (LEVEL WALKING, STAIR ASCENT, AND STAIR DESCENT) WITH THE
WAVELET FEATURES (TABLE I)

features (Table II). Within each of these seven sets, the features
were derived individually for each of the three components of
the triaxial accelerometer signal. Mean and standard deviation shown to be effective in previous activity recognition stud-
(SD) have been used in previous studies [34] to characterize ies [11], [12] and selects the activity that is closest to the
windows of accelerometer data. As an extension of this set, we feature under question using the Euclidean distance metric in
defined the multiple statistics features set, which additionally the multidimensional feature space. We employed kNN as our
included median and 25th and 75th percentile [33]. Low-pass recognition engine, due to its implementational simplicity and
filtering is commonly used to separate the dc and ac components flexibility, and the fact that it can allow analysis of the classi-
of an accelerometer signal [42]. Previous studies have defined fication decisions. With leave-one-subject-out cross validation,
features as the mean dc and the mean of the rectified ac sig- the classifier is trained with data from all subjects except one
nal [16], [17]. These two statistics were therefore used to define and then tested with data from the excluded subject. This pro-
the third set of time-domain features. cess is repeated until each subject has been used once as the
In order to derive frequency-domain features, a fast Fourier testing dataset. With this approach, the overall accuracy is cal-
transform (FFT) was performed on each 2 s window. The prin- culated as the average test classification result of each train-test
cipal frequency was defined as the first of the frequency-domain repetition. Cross validation is a popular statistical resampling
feature sets (fourth in Table II). This has been used previously as procedure [44] and we use it here to evaluate the accuracy of the
an addition to time-domain measures in order to improve classi- kNN classifier for a given set of features. The mean accuracy
fication accuracy [35]. The second frequency-domain feature set of all train-test repetitions can be influenced by a small number
was chosen to be spectral energy, which is defined to be the sum of subjects who may bias the overall result. Therefore, in order
of the squared FFT coefficients [11], [43]. A recent study carried to compare the performance of two sets of features, the Mann–
out by Bao and Intille [11] obtained high levels of classification Whitney U test was used to test for differences in the two distri-
accuracy using a mixed set of time and frequency-domain fea- butions of testing accuracies. This test was chosen as it was not
tures. Therefore, this was included as the sixth set of features. In possible to guarantee that these distributions were normally dis-
addition to spectral energy [11], the sixth set included dc, corre- tributed. A significance level of p < 0.01 was used throughout.
lations between axes, and frequency-domain entropy. This latter To address our first research aim, only windows of data that
feature gives a measure of the normalized information entropy corresponded to level walking, stair ascent, and stair descent
of the FFT components and allows for differentiation between were included in the analysis. For this three-activity classifica-
activities that have simple acceleration patterns and those with tion problem, accuracy was determined for the waist-mounted
more complex patterns [11]. The final frequency-domain feature accelerometer for each of the seven sets of wavelet features
set was defined as the magnitude of the first five components and for each of the seven sets of time-frequency features. This
of the FFT power spectrum. As with the other feature sets, this process was then repeated for the thigh and then the ankle-
set of parameters was derived separately for each of the three mounted sensor. To establish whether it would be possible to
components of acceleration. Although it is more common to improve classification accuracy using data from more than one
use the power spectrum of FFT coefficients, preliminary studies sensor, the analysis was performed for all seven possible com-
showed that the magnitudes gave improved accuracy and were binations of the three sensors (as shown in the first column of
therefore used for the final analysis. Table III). Once classification accuracies had been determined
for the three-activity problem, the process was repeated with
D. Activity Classification windows of accelerometer data from all eight activities.
In order to compare the discriminate ability of each of the
III. RESULTS
different features sets, a k-nearest neighbor (kNN) classifier
was implemented and its accuracy determined using leave-one- Table III gives the classification accuracies for the wavelet
subject-out cross validation. This type of classifier has been feature sets and different accelerometer placements for the

Authorized licensed use limited to: UNIVERSITY OF SALFORD. Downloaded on May 21, 2009 at 04:12 from IEEE Xplore. Restrictions apply.
876 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 56, NO. 3, MARCH 2009

TABLE IV TABLE V
CLASSIFICATION ACCURACIES (%) FOR THE THREE-ACTIVITY CLASSIFICATION CLASSIFICATION ACCURACIES (%) FOR THE EIGHT-ACTIVITY CLASSIFICATION
PROBLEM (LEVEL WALKING, STAIR ASCENT, AND STAIR DESCENT) WITH THE PROBLEM WITH THE WAVELET FEATURES (SEE TABLE I)
TIME AND FREQUENCY FEATURES (TABLE II)

TABLE VI
CLASSIFICATION ACCURACIES (%) FOR THE EIGHT-ACTIVITY CLASSIFICATION
PROBLEM WITH THE TIME AND FREQUENCY FEATURES (SEE TABLE II)

three-activity classification problem. Table IV illustrates the

same information but for the time-frequency features. Overall,
for the three-activity problem, the highest classification accuracy
for a single sensor (97% ± 3%) was obtained using FFT com-
ponents derived from the ankle-mounted unit. This distribution
of accuracies was significantly higher than those obtained from
all other feature sets derived from a single sensor (p < 0.01).
In general, for the wavelet feature sets, the highest performance
was obtained using the sum of the absolute values for each sen-
sor configuration (Table III). However, the performance of this
feature set was, in some cases, not significantly better than the
wavelet feature set proposed by Wang et al. [14].
In order to establish whether, in general, the time-frequency
features outperformed the wavelet features, a number of statis-
tical tests were performed. First, the performance of the best set
termine whether differences in accuracy existed between the two
of time-frequency features was compared with the best set of
types of features, comparisons were made between the best and
wavelet features for each sensor configuration. With the excep-
second-best performing time/frequency and wavelet features.
tion of the waist-mounted sensor, the time-frequency feature sets
These comparisons showed that, with the exception of features
significantly outperformed the wavelet feature sets (p < 0.01) in
derived from a waist-mounted sensor, the time/frequency fea-
every case (Tables III and IV). Further testing was then carried
tures significantly outperformed the wavelet features (p < 0.01)
out by comparing the second-best performing time-frequency
(Tables V and VI).
feature set with the second-best performing wavelet feature set
The classification accuracies reported in Tables V and VI
for every sensor configuration. These tests also showed the time-
represent an average across all of the eight different activities.
frequency feature sets to significantly outperform the wavelet
Although these data would suggest that FFT features give bet-
feature sets in every case (p < 0.01).
ter classification accuracy than wavelet features, it is not clear
The results of the eight-activity classification problem dis-
whether this result is true across all activities. To investigate this
played similar trends to the three-activity problem for both
further, sensitivity and specificity were calculated separately for
the wavelet features (Table V) and the time-frequency features
each of the eight activities for the ankle-mounted sensor with the
(Table VI). Again, the highest wavelet classification accura-
best-performing wavelet feature set (sum of the absolute values)
cies, for each of the sensor configurations, were obtained us-
and the best performing time/frequency feature set (magnitude
ing the sum of the absolute values (Table V). However, the
of the FFT components). This comparison (Table VII) shows
performance of this feature set was, in some cases, not signif-
that, for each of the different activities, the FFT feature set
icantly better than the sum of the squares feature set. For the
outperforms the wavelet feature set.
time/frequency features, maximal classification accuracy for a
single sensor (92% ± 7%) was again obtained when the indi-
vidual FFT components were derived from the ankle-mounted IV. DISCUSSION
unit (Table VI). However, this distribution of accuracies was not This study was designed to compare the discriminative ability
significantly different to those obtained using FFT coefficients of wavelet features with time/frequency features for two activ-
derived from the thigh-mounted sensor (p = 0.16). Again, to de- ity classification problems: a simple three-activity problem and

Authorized licensed use limited to: UNIVERSITY OF SALFORD. Downloaded on May 21, 2009 at 04:12 from IEEE Xplore. Restrictions apply.
PREECE et al.: COMPARISON OF FEATURE EXTRACTION METHODS FOR THE CLASSIFICATION OF DYNAMIC ACTIVITIES 877

TABLE VII respond to a single activity, the frequency content of the signal
SENSITIVITY AND SPECIFICITY FOR EACH ACTIVITY FOR THE BEST
PERFORMING TIME/FREQUENCY AND WAVELET FEATURE SETS
varies little with time. Wavelet analysis allows for the analysis of
nonstationary signals. However, it is not clear whether parame-
ters derived from wavelet coefficients represent a more effective
means of characterizing short windows of data than standard
frequency-domain techniques. In this study, data were collected
from 20 healthy subjects. Analysis of these data showed that
features derived from an FFT analysis outperformed those de-
rived from wavelet coefficients. This may reflect the suitability
of standard frequency-domain techniques for characterizing the
short-duration stationary signals, which were characteristic of
our subject group.
This study found surprisingly good levels of classification
accuracy when using simple time-domain features. A number of
other studies have reported high levels of classification accuracy
an eight-activity problem. In addition, classification accuracies using time-domain features. For example, Pirttikangas [34] used
were compared for three individual lower limb placements, the means and SDs from a number of body-worn accelerometers to
waist, thigh, and ankle, as well as some of their combinations. In accurately classify (>90%) a wide range of activities. Similarly,
general, similar levels of accuracy were found when data from a Fahrenberg et al. [16] used mean dc and mean rectified ac in
waist-mounted sensor were used to obtain either time/frequency a hierarchical classification to differentiate between a range of
or wavelet features. However, for both the ankle- and thigh- static postures and movements. For the current study, this set
mounted sensors, time/frequency features significantly outper- normally outperformed the other time-domain features and often
formed the wavelet features. For both classification problems, gave comparable accuracy to the FFT component feature set.
the optimal accelerometer placement for a single sensor was The highest classification accuracy for a single sensor was ob-
shown to be on the ankle. tained for the FFT component feature set and the ankle-mounted
Five previous studies were identified, which had used wavelet sensor. This feature set consistently outperformed both the en-
features to discriminate between level walking, stair ascent, and ergy feature and the larger set proposed by Bao and Intille [11].
stair descent. Of these five studies, only Nyan et al. [24] and As they studied a larger range of activities than those of this
Wang et al. [14] reported intersubject classification accuracies. study, direct comparison of classification accuracies is not pos-
The remaining three studies simply demonstrated significant sible. However, their reported maximum classification accuracy
differences between wavelet parameters corresponding to each of 84% using data from five sensors is similar to the maximum
of the three activities [21]–[23]. Nyan et al. [24] used a simple accuracy (90%) achieved in our study for the eight-activity prob-
threshold-based classification scheme that required the manual lem. Huynh and Schiele [37] also compared the discriminative
selection of arbitrary thresholds for both of their features. With ability of individual FFT components with simple time-domain
this approach, they obtained accuracies ranging from 97% to features, spectral energy, and spectral entropy for a range of
99%. The use of thresholds determined by the experimenter activities including walking, jogging, and hopping. In agree-
reduces the system’s ability for fully automatic classification. ment with this study, they found the FFT component to have
In our paper, we aimed to build an automated system that can higher discriminative ability than the other features. However,
be trained by a set of supervised subjects and activity scenar- they were unable to identify a single component that performed
ios. This system can then be applied to new subjects, instru- best for each activity. Although, in this study, the first five com-
mented with the same sensors, without any further supervision. ponents were used as input to the classifier, it is possible to use
In their study, Nyan et al. [24] collected data using two shoulder- a larger or smaller number of components. Fig. 3 illustrates how
mounted accelerometers so that their results are not directly the classification accuracy changes as the number of compo-
comparable to those in this study. nents varies. It can be seen that using the first six components
Wang et al. [14] studied level walking, stair ascent/descent, produces maximal accuracy for both the three-activity and eight-
and walking up/down a slope using data collected from a waist- activity problems. Although, for the three-activity problem, an
mounted accelerometer. Using a multilayer perceptron neural almost perfect result is achieved, with the eight-activity problem
network classifier, they obtained classification accuracies rang- a maximum accuracy of only 94% is possible. Inspection of the
ing from 89% to 92% for these five activities. However, in their corresponding confusion matrix (Table VIII) showed that jump-
study, an individual normalization scheme was used in which the ing was often confused with a number of other activities. When
features were divided by those obtained from a 5-s flat walking this activity was excluded, the accuracy increased to 97%.
session. When unadjusted features were used for classification, There are a number of limitations of this study. First, subjects
similar levels of accuracy to those found in this study were performed each of the separate activities while being videoed by
obtained. the experimenter. Under these conditions, it is possible that in-
In order to minimize computational power requirements, ac- dividuals may subconsciously modify their habitual movement
tivity classification algorithms typically work with relatively patterns. However, some method is required for annotating the
short windows of sensor data. As these windows typically cor- sensor data. The video method, used in this study, was selected

Authorized licensed use limited to: UNIVERSITY OF SALFORD. Downloaded on May 21, 2009 at 04:12 from IEEE Xplore. Restrictions apply.
878 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 56, NO. 3, MARCH 2009

most appropriate feature sets for other subjects groups, such as

the elderly or neurologically impaired.

REFERENCES
[1] K. Aminian and B. Najafi, “Capturing human motion using body-fixed
sensors: Outdoor measurement and clinical applications,” Comput. Anim.
Virtual Worlds, vol. 15, pp. 79–94, 2004.
[2] W. Zjlstra and K. Aminian, “Mobility assessment in older people: New
possibilities and challenges,” Eur. J. Aging, vol. 4, pp. 3–12, 2007.
[3] World Health Organization Global Strategy on Diet, Physical Activity and
Health. Geneva, Switzerland: WHO, 2004.
[4] R. A. Washburn and H. J. Montoye, “The assessment of physical activity
by questionnaire,” Amer. J. Epidemiol., vol. 123, pp. 563–576, 1986.
[5] B. E. Ainsworth, D. R. Jacobs, Jr., and A. S. Leon, “Validity and relia-
bility of self-reported physical activity status: The lipid research clinics
Fig. 3. Plot to show the accuracy of activity recognition as the number of questionnaire,” Med. Sci. Sports Exerc., vol. 25, pp. 92–98, 1993.
FFT coefficients is increased. The dashed line shows the relationship for the [6] C. J. Riddoch, C. Mattocks, K. Deere, J. Saunders, J. Kirkby, K. Tilling,
three-activity problem (level walking, stair ascent, and stair descent) and the S. D. Leary, S. N. Blair, and A. R. Ness, “Objective measurement of levels
solid line shows the relationship for the eight-activity problem. and patterns of physical activity,” Arch. Dis. Child., vol. 92, pp. 963–969,
2007.
TABLE VIII [7] R. R. Pate, K. A. Pfeiffer, S. G. Trost, P. Ziegler, and M. Dowda, “Phys-
CONFUSION MATRIX SHOWING CLASSIFICATION RESULTS FOR THE ical activity among children attending preschools,” Pediatrics, vol. 114,
EIGHT-ACTIVITY PROBLEM USING FEATURES DEFINED AS THE MAGNITUDES pp. 1258–1263, 2004.
OF THE FIRST TEN FFT COMPONENTS OBTAINED FROM THE [8] B. Najafi, K. Aminian, F. Loew, Y. Blanc, and P. A. Robert, “Measurement
ANKLE-MOUNTED SENSOR of stand–sit and sit–stand transitions using a miniature gyroscope and its
application in fall risk evaluation in the elderly,” IEEE Trans. Biomed.
Eng., vol. 49, no. 8, pp. 843–851, Aug. 2002.
[9] N. Streitz and P. Nixon, “The disappearing computer,” Commun. ACM,
vol. 48, pp. 32–35, 2005.
[10] J. Coutaz, J. L. Crowley, S. Dobson, and D. Garlan, “Context is key,”
Commun. ACM, vol. 48, pp. 49–53, 2005.
[11] L. Bao and S. S. Intille, “Activity recognition from user-annotated acceler-
ation data,” Pers Comput., Lecture Notes in Computer Science, vol. 3001,
pp. 1–17, 2004.
[12] U. Maurer, A. Rowe, A. Smailagic, and D. Siewiorek, “Location and
activity recognition using eWatch: A wearable sensor platform,” Ambient
Intell. Everday Life, Lecture Notes in Computer Science, vol. 3864, pp. 86–
102, 2006.
[13] J. Parkka, M. Ermes, P. Korpipaa, J. Mantyjarvi, J. Peltola, and I. Korho-
nen, “Activity classification using realistic data from wearable sensors,”
IEEE Trans. Inf. Technol. Biomed., vol. 10, no. 1, pp. 119–128, Jan. 2006.
[14] N. Wang, E. Ambikairajah, N. H. Lovell, and B. G. Celler, “Accelerometry
based classification of walking patterns using time-frequency analysis,”
in Proc. 29th Annu. Conf. IEEE Eng. Med. Biol. Soc., Lyon, France, 2007,
as it was believed to be more accurate than self-observation pp. 4899–4902.
[15] S. J. Preece, J. Y. Goulermas, L. P. J. Kenney, K. Meijer, R. H. Crompton,
by the subject. Another limitation is that only relatively young, and D. Howard, “Activity identification using body-mounted sensors—
healthy subjects were included in the study. Clearly, it is not pos- a review of classification techniques,” Physiol. Meas., 2008, to be
sible to generalize our findings, that frequency domain features published.
[16] J. Fahrenberg, F. Foerster, M. Smeja, and W. Muller, “Assessment of pos-
perform better than wavelet features, to other subject groups, ture and motion by multichannel piezoresistive accelerometer recordings,”
such as the elderly or patients with neurological impairment. Psychophysiology, vol. 34, pp. 607–612, 1997.
For such individuals, jerkiness of movement may lead to iso- [17] J. Fahrenberg, W. Muller, F. Foerster, and M. Smeja, “A multi-channel
investigation of physical activity,” Psychophysiology, vol. 10, pp. 209–
lated frequency transients that maybe better characterized using 217, 1996.
wavelet features. Further research is thus needed to determine [18] F. R. Allen, E. Ambikairajah, N. H. Lovell, and B. G. Celler, “Classifi-
the most appropriate features for activity classification for dif- cation of a known sequence of motions and postures from accelerometry
data using adapted Gaussian mixture models,” Physiol. Meas., vol. 27,
ferent patient groups. pp. 935–951, 2006.
For this study, a single classifier (kNN) was used to evaluate [19] T. Prill and J. Fahrenberg, “Simultaneous assessment of posture and limb
the discriminatory ability of the different feature sets. Although movements (e.g., periodic leg movements) with calibrated multiple ac-
celerometry,” Physiol. Meas., vol. 27, pp. N47–N53, 2006.
it is possible to use other methods to identify optimal features, [20] S. J. Preece, J. Y. Goulermas, L. P. J. Kenney, and D. Howard, “A com-
this method was chosen for its simplicity, flexibility, and pop- parison of different feature generation methods in activity classification,”
ularity. In general, different classifiers can have different sub- in Proc. Int. Conf. Ambul. Monit. Phys. Act. Movement, Rotterdam, The
Netherlands, 2008, p. 87.
sets of optimal features and a larger evaluation study would be [21] M. Sekine, T. Tamura, M. Akay, T. Fujimoto, T. Togawa, and Y. Fukui,
needed to perform comparisons between different classifiers. “Discrimination of walking patterns using wavelet-based fractal analysis,”
In conclusion, the findings of this study suggest that fu- IEEE Trans. Neural Syst. Rehabil. Eng., vol. 10, no. 3, pp. 188–196, Sep.
2002.
ture activity monitoring systems, aimed at healthy individuals, [22] M. Sekine, T. Tamura, T. Togawa, and Y. Fukui, “Classification of waist-
should consider using an FFT feature set derived from an ankle- acceleration signals in a continuous walking record,” Med. Eng. Phys.,
mounted sensor. Further research is required to determine the vol. 22, pp. 285–291, 2000.

Authorized licensed use limited to: UNIVERSITY OF SALFORD. Downloaded on May 21, 2009 at 04:12 from IEEE Xplore. Restrictions apply.
PREECE et al.: COMPARISON OF FEATURE EXTRACTION METHODS FOR THE CLASSIFICATION OF DYNAMIC ACTIVITIES 879

[23] T. Tamura, M. Sekine, M. Ogawa, T. Togawa, and Y. Fukui, “Classifi- Stephen J. Preece received the B.Sc. degree in
cation of acceleration waveforms during walking by wavelet transform,” physics in 1996 and the Ph.D. degree in applied
Methods Inf. Med., vol. 36, pp. 356–359, 1997. mathematics in 2001, both from the University of
[24] M. N. Nyan, F. E. Tay, K. H. Seah, and Y. Y. Sitoh, “Classification of gait Birmingham, Birmingham, U.K.
patterns in the time-frequency domain,” J. Biomech., vol. 39, pp. 2647– He has been engaged in research on a range
2656, 2006. of biomedical engineering projects. He is currently
[25] S. Mallat and W. L. Hwang, “Singularity detection and processing with a Postdoctoral Research Fellow in the field of
wavelets,” IEEE Trans. Inf. Theory, vol. 38, no. 2, pp. 617–643, Mar. human performance research at the University of
1992. Salford, Salford, U.K. His current research interests
[26] C. V. C. Bouten, K. T. M. Koekkoek, M. Verduin, R. Kodde, and J. D. include algorithm development in activity monitoring
Janssen, “A triaxial accelerometer and portable data processing unit for and electromyography methodology for characteriz-
the assessment of daily physical activity,” IEEE Trans. Biomed. Eng., ing lower limb muscle function during gait.
vol. 44, no. 3, pp. 136–147, Mar. 1997.
[27] E. A. Heinz, K. S. Kunze, S. Sulistyo, H. Junker, P. Lukowicz, and
G. Troster, “Experimental evaluation of variations in primary features
used for accelerometric context recognition,” Ambient Intell., Lecture
Notes in Computer Science, vol. 2875, pp. 252–263, 2003.
[28] M. J. Mathie, A. C. Coster, N. H. Lovell, and B. G. Celler, “Detection of
John Yannis Goulermas (S’98–M’98) received the
daily physical activities using a triaxial accelerometer,” Med. Biol. Eng.
B.Sc. (Hons, First Class) degree in computation in
Comput., vol. 41, pp. 296–301, 2003.
1994 from the University of Manchester Institute
[29] H. B. Bussmann, P. J. Reuvekamp, P. H. Veltink, W. L. Martens, and
of Science and Technology (UMIST), Manchester,
H. J. Stam, “Validity and reliability of measurements obtained with an
U.K., and the M.Sc. degree by research in 1996 and
“activity monitor” in people with and without a transtibial amputation,”
Phys. Ther., vol. 78, pp. 989–998, 1998. the Ph.D. degree in 2000, both from the Department
of Electrical Engineering, where he worked in the
[30] K. Aminian, P. Robert, E. E. Buchser, B. Rutschmann, D. Hayoz, and
field of on machine vision.
M. Depairon, “Physical activity monitoring based on accelerometry: Vali-
He has been a Scientific Developer in industry
dation and comparison with video observation,” Med. Biol. Eng. Comput.,
vol. 37, pp. 304–308, 1999. and a Postdoctoral Researcher in virtual reality and
biomechanics. He joined the Department of Electri-
[31] J. B. J. Bussmann, Y. M. van de Laar, M. P. Neeleman, and H. J. Stam,
cal Engineering and Electronics of the University of Liverpool, Liverpool, U.K.,
“Ambulatory accelerometry to quantify motor behaviour in patients after
in 2005. His current research interests include pattern recognition, optimization,
failed back surgery: A validation study,” Pain, vol. 74, pp. 153–161, 1998.
[32] B. Najafi, K. Aminian, A. Paraschiv-Ionescu, F. Loew, C. J. Bula, and and image analysis, with applications in biomedical engineering and industrial
monitoring.
P. Robert, “Ambulatory system for human motion analysis using a kine-
matic sensor: Monitoring of daily physical activity in the elderly,” IEEE
Trans. Biomed. Eng., vol. 50, no. 6, pp. 711–723, Jun. 2003.
[33] M. Ermes, J. Parkka, J. Mantyjarvi, and I. Korhonen, “Detection of daily
activities and sports with wearable sensors in controlled and uncontrolled
conditions,” IEEE Trans. Inf. Technol. Biomed., vol. 12, no. 1, pp. 20–26,
Jan. 2008. Laurence P. J. Kenney received the B.Sc. degree in
[34] P. Pirttikangas, K. Fujinami, and T. Nakajima, “Feature selection and mechanical engineering in 1986 and the Ph.D. degree
activity recognition from wearable sensors,” Ubiquitous Comput. Syst., in engineering design in 1993, both from the Univer-
Lecture Notes in Computer Science, vol. 4239, pp. 516–527, 2006. sity of Salford, Salford, U.K.
[35] F. Foerster and J. Fahrenberg, “Motion pattern and posture: Correctly From 1998–2000, he was a Postdoctoral Re-
assessed by calibrated accelerometers,” Behav. Res. Methods Instrum. searcher with Roessingh Research and Development
Comput., vol. 32, pp. 450–457, 2000. in The Netherlands. He rejoined the University of
[36] K. Zhang, P. Werner, M. Sun, F. X. Pi-Sunyer, and C. N. Boozer, “Mea- Salford in 2000, where he was engaged in the estab-
surement of human daily physical activity,” Obes. Res., vol. 11, pp. 33–40, lishment of the Centre for Rehabilitation and Human
2003. Performance Research and where he currently is a
[37] T. Huynh and B. Schiele, “Analyzing features for activity recognition,” Senior Research Fellow. His current research inter-
in Proc. Conf. Smart Objects Ambient Intell.: Innov. Context-Aware Serv.: ests include the development and evaluation of functional electrical stimulation
Usages Technol., Grenoble, France, 2005, pp. 159–164. and prosthetic devices.
[38] K. van Laerhoven and O. Cakmakci, “What shall we teach our pants?,” in
Proc. 4th Int. Symp. Wearable Comput., 2000, pp. 77–83.
[39] S. G. Mallat, “A thoery for multi-resolution signal decomposition: The
wavelet representation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 11,
no. 7, pp. 674–693, Jul. 1989.
[40] T. Chau, “A review of analytical techniques for gait data. Part 2: Neural
network and wavelet methods,” Gait Posture, vol. 13, pp. 102–120, 2001. David Howard received the B.E. degree in mechan-
[41] B. B. Mandelbrot, The Fractal Geometry of Nature. New York: Free- ical and production engineering from Brunel Univer-
man, 1983. sity, Uxbridge, U.K., in 1980, and the Ph.D. degree in
[42] D. M. Karantonis, M. R. Narayanan, M. Mathie, N. H. Lovell, and B. G. mechanical engineering from the University of Bath,
Celler, “Implementation of a real-time human movement classifier using Bath, U.K., in 1987.
a triaxial accelerometer for ambulatory monitoring,” IEEE Trans. Inf. He is currently with the University of Salford,
Technol. Biomed., vol. 10, no. 1, pp. 156–167, Jan. 2006. Salford, U.K., in the Centre for Rehabilitation and
[43] A. Sugimoto, Y. Hara, T. W. Findley, and K. Yoncmoto, “A useful method Human Performance Research. His current research
for measuring daily physical activity by a three-direction monitor,” Scand. interests include applying gait simulation to design-
J. Rehabil. Med., vol. 29, pp. 37–42, 1997. assistive devices, controlling functional electrical
[44] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification, 2nd ed. stimulation, activity monitoring using body-mounted
New York: Wiley, 2001. sensors, and real-world monitoring of patients using assistive devices.

Authorized
View publication stats licensed use limited to: UNIVERSITY OF SALFORD. Downloaded on May 21, 2009 at 04:12 from IEEE Xplore. Restrictions apply.