A Comparison of Feature Extraction Methods for the Classification of Dynamic Activities From Accelerometer Data

TL;DR: The findings show that, although the wavelet transform approach can be used to characterize nonstationary signals, it does not perform as accurately as frequency-based features when classifying dynamic activities performed by healthy subjects.

Abstract: Driven by the demands on healthcare resulting from the shift toward more sedentary lifestyles, considerable effort has been devoted to the monitoring and classification of human activity. In previous studies, various classification schemes and feature extraction methods have been used to identify different activities from a range of different datasets. In this paper, we present a comparison of 14 methods to extract classification features from accelerometer signals. These are based on the wavelet transform and other well-known time- and frequency-domain signal characteristics. To allow an objective comparison between the different features, we used two datasets of activities collected from 20 subjects. The first set comprised three commonly used activities, namely, level walking, stair ascent, and stair descent, and the second a total of eight activities. Furthermore, we compared the classification accuracy for each feature set across different combinations of three different accelerometer placements. The classification analysis has been performed with robust subject-based cross-validation methods using a nearest-neighbor classifier. The findings show that, although the wavelet transform approach can be used to characterize nonstationary signals, it does not perform as accurately as frequency-based features when classifying dynamic activities performed by healthy subjects. Overall, the best feature sets achieved over 95% intersubject classification accuracy.

Summary

1. Introduction

• Over the last decade there has been considerable research effort directed towards the monitoring and classification of physical activity patterns from body-fixed sensor data [1, 2].
• A mobile phone may detect when a person is driving a vehicle and automatically divert a call.
• Effective algorithms are also required to interpret the accelerometer data in the context of ~3~ different activities.
• 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.
• It was felt that this work would underpin the development of an off-the-shelf activity monitor which could be used to classify activity patterns across different subjects.

2.1 Data collection

• Accelerometer data was collected using Pegasus activity monitors developed by ETB, UK.
• A sampling frequency of 64Hz was selected for this study as this is sufficiently larger than the 20Hz sampling required to assess daily activity [26].
• To secure each unit in place specialised bandage (FabriFoam®) was first positioned around each of the body segments and the activity monitors, which were backed with Velcro®, adhered to the underwrapped bandage.
• A number of studies have shown that static postures can be differentiated from dynamic activity by applying a single threshold to some measure of acceleration variability [28, 29].
• Subjects completed a total of eight different activities (level walking, walking upstairs and downstairs, jogging, running, hopping on the left and right leg and jumping).

2.2 Wavelet features

• A number of previous activity classification studies have derived time-frequency features obtained using the filter bank interpretation of the discrete wavelet transform (DWT) [22, 24].
• The first set of wavelet features was proposed by Tamura et al. [23].
• Again there are two features, the first being the total of the summations of the detail signal at levels 6 and 7.
• In contrast Wang et al. [14] used wavelet packet analysis to derive 33 features from a tri-axial accelerometer signal.
• Given the high sampling frequency used by Sekine et al. [21] (1024 Hz), they were able to calculate the fractal dimension from the variance of the detail coefficients across seven different levels.

2.3 Time and frequency-domain features

• For additional comparison, the authors also employed three sets of time-domain features and four sets of frequency-domain features (Table II).
• Within each of these seven sets, the features were derived individually for each of the three components of the tri-axial accelerometer signal.
• These two statistics were therefore used to define the third set of time-domain features.
• This has been used previously as an addition to time-domain measures in order to improve classification accuracy [35].
• The second frequency-domain feature set was chosen to be spectral energy, which is defined to be the sum of the squared ~11~ FFT coefficients [11, 43].

2.4 Activity classification

• In order to compare the discriminate ability of each of the different features sets, a k-Nearest Neighbour (kNN) classifier was implemented and its accuracy determined using leave-onesubject-out cross validation.
• This process is repeated until each subject has been used once as the testing dataset.
• Cross validation is a popular statistical resampling procedure [44] and the authors use it here to evaluate the accuracy of the kNN classifier for a given set of features.
• This test was chosen as it was not possible to guarantee that these distributions were normally distributed.
• For this three-activity classification problem, accuracy was determined for the waist-mounted accelerometer for each of the seven sets of wavelet features and for each of the seven sets of time/frequency features.

3. Results

• Table III gives the classification accuracies for the wavelet feature sets and different accelerometer placements for the three-activity classification problem.
• Table IV illustrates the same information but for the time/frequency features.
• This distribution of accuracies was significantly higher than those obtained from all other feature sets derived from a single sensor (p<0.01).
• In order to establish whether, in general, the time/frequency features outperformed the wavelet features, a number of statistical tests were performed.
• For the time/frequency features, maximal classification accuracy for a single sensor (92±7%) was again obtained when the individual FFT components were derived from the ankle-mounted unit (Table VI).

4. Discussion

• This study was designed to compare the discriminative ability of wavelet features with time/frequency features for two activity classification problems: a simple three-activity problem and an eight-activity problem.
• Analysis of this data showed that features derived from an FFT analysis outperformed those derived from wavelet coefficients.
• Their reported maximum classification accuracy of 84% using data from five sensors is similar to the maximum accuracy (90%) achieved in their study for the eight-activity problem.
• For such individuals, jerkiness of movement may lead to isolated frequency transients which maybe better characterised using wavelet features.

5. Conclusion

• This study was performed on healthy individuals.
• More specifically, the highest levels of classification accuracy were obtained from individual FFT components.
• The study also compared classification accuracies across three different sensor placements and showed a sensor mounted at the ankle to outperform the thigh and waist sensors for most feature sets.
• Further work is required to determine the most appropriate features sets for other subjects groups, such as the elderly or neurologically impaired. ~18~.

Figures

Table III: Classification accuracies (%) obtained using leave-one-out cross validation for the three-activity classification problem (level walking, stair ascent and stair descent) with the wavelet features (table I). Accuracies have been reported for each of the different accelerometer combinations.

Table II: Summary of the time and frequency-domain features

Table VII: Sensitivity and specificity for each activity for the best performing time/frequency and wavelet feature sets.

Table VI: Classification accuracies (%) obtained using leave-one-out cross validation for the eight-activity classification problem with the time and frequency features (table II). Accuracies have been reported for each of the different accelerometer combinations.

Table VII: Sensitivity and specificity for each activity for the best performing time/frequency and wavelet feature sets.

Table I: Summary of the different wavelet features. The term cDj refers to the detail coefficient at the jth level of decomposition. All other nomenclature is explained within the table.

Table V: Classification accuracies (%) obtained using leave-one-out cross validation for the eight-activity classification problem with the wavelet features (table 1). Accuracies have been reported for each of the different accelerometer combinations.

Table IV: Classification accuracies (%) obtained using leave-one-out cross validation for the three-activity classification problem (level walking, stair ascent and stair descent) with the time and frequency features (table II). Accuracies have been reported for each of the different accelerometer combinations.

Full text

A comparison of feature extraction
methods for the classification of dynamic
activities from accelerometer data
Preece, SJ, Goulermas, JY, Kenney, LPJ and Howard, D
Title A comparison of feature extraction methods for the classification of
dynamic activities from accelerometer data
Authors Preece, SJ, Goulermas, JY, Kenney, LPJ and Howard, D
Publication title IEEE Transactions on Biomedical Engineering
Publisher Institute of Electrical and Electronics Engineers
Type Article
USIR URL This version is available at: http://usir.salford.ac.uk/id/eprint/12578/
Published Date 2009
USIR is a digital collection of the research output of the University of Salford. Where copyright
permits, full text material held in the repository is made freely available online and can be read,
downloaded and copied for non-commercial private study or research purposes. Please check the
manuscript for any further copyright restrictions.
For more information, including our policy and submission procedure, please
contact the Repository Team at: library-research@salford.ac.uk.

~1~
A Comparison of Feature Extraction Methods for the Classification
of Dynamic Activities from Accelerometer Data
S. J. Preece, J. Y. Goulermas, L. P. J. Kenney, D. Howard
Abstract
Driven by the demands on healthcare resulting from the shift towards more sedentary
lifestyles considerable effort has been devoted to the monitoring and classification of human
activity. In previous work, various classification schemes and feature extraction methods
have been used to identify different activities from a range of different datasets. In this paper,
we present a comparison of fourteen methods to extract classification features from
accelerometer signals. These are based on the wavelet transform and other well-known time-
and frequency-domain signal characteristics. To allow an objective comparison between the
different features, we used two datasets of activities collected from twenty subjects. The first
set comprised three commonly used activities, level walking, stair ascent and descent and
the second a total of eight activities. Furthermore, we compared the classification accuracy
for each feature set across different combinations of three different accelerometer
placements. The classification analysis has been performed with robust subject-based cross-
validation methods using a Nearest-Neighbour classifier. The findings show that, although
the wavelet transform approach can be used to characterise non-stationary signals, it does
not perform as accurately as frequency-based features when classifying dynamic activities
performed by healthy subjects. Overall, the best feature sets achieved over 95% inter-subject
classification accuracy.
1. Introduction
Over the last decade there has been considerable research effort directed towards the
monitoring and classification of physical activity patterns from body-fixed sensor data [1, 2].
This has been motivated by a number of important health-related applications. For example,

~2~
with the trend towards more sedentary lifestyles there is a growing interest in the link
between levels of physical activity and common health problems, such as diabetes,
cardiovascular disease and osteoporosis [3]. As self reported measures have been shown to
be unreliable [4, 5], systems for activity profiling are beginning to play an important role in
large-scale epidemiological studies in this area [6, 7]. Furthermore, such systems can also be
used to assess the effectiveness of different interventions aimed at increasing levels of
physical activity and for motivating individuals to become more physically active.
The success of a given rehabilitation programme is often judged by not only the levels of
activity, but also the type of activity that an individual can return to after treatment. In
addition, as fall risk increases with age, so a better understanding of the factors contributing
to fall risk becomes more important. Ambulatory monitoring of various activities, including the
time spent in sit-stand transitions have shown promise as predictors of fall-risk [8]. Further,
both type and intensity of individuals’ activity are of interest to urban designers, and
designers, manufacturers and purchasers of certain medical devices (e.g. advanced
responsive pacemakers and orthopaedic implants).
In addition to health-related applications, portable systems which can accurately identify the
activity of the user have the potential to play a fundamental role in a ubiquitous computing
scenario [9, 10]. In this field, computing devices use information from a variety of sensors to
determine the context of a situation. Different devices can then use the context information to
deliver an appropriate service. For example, a mobile phone may detect when a person is
driving a vehicle and automatically divert a call.
With recent advances in miniaturised sensing technology, it is now possible to collect and
store acceleration data from individual body segments over extended periods of time.
Although this technology offers the ideal platform for monitoring daily activity patterns,
effective algorithms are also required to interpret the accelerometer data in the context of

~3~
different activities. Previous studies have shown machine learning or artificial intelligence
approaches to be effective for identifying a range of different activities from body-fixed sensor
data [11-14]. These techniques typically operate via a two-stage process [15]. Firstly,
features are derived from windows of accelerometer data. A classifier is then used to identify
the activity corresponding to each separate window of data. A range of different approaches
has been used to obtain features from accelerometer data, with some researchers deriving
features directly from the time-varying acceleration signal [12, 16-18] and others from a
frequency analysis [11, 13, 19, 20]. More recently wavelet analysis has been used to derive
so-called time-frequency features [14, 21-24].
With wavelet analysis the original signal is decomposed into a series of coefficients which
carry both spectral and temporal information about the original signal. From these
coefficients, it is possible to identify localised temporal instances at which there is a change
in frequency characteristics of the original signal [25]. This concept has been applied
successfully to accelerometer 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 which
characterise 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 walking, stair ascent and stair descent [21-23], but have not
compared their performance against simpler approaches. Therefore our first research aim
was to compare features for this three-activity classification problem. As a second aim, we
sought to compare the same features for a larger set of activities, which represents a more

~4~
challenging problem. Additionally, since the performance of a given set of features can be
dependent on the location of the monitor, we compared accuracy for the different features
across a number of different lower limb placements. It was felt that this work would underpin
the development of an off-the-shelf activity monitor which could be used to classify activity
patterns across different subjects.
2. Methods
2.1 Data collection
Accelerometer data was collected using Pegasus activity monitors developed by ETB, UK.
Each of these units contained a tri-axial accelerometer, with dynamic range of ±5g, which
was sampled a with 10-bit resolution. With these devices it is possible to sample
accelerometer data at a user-defined frequency and to store this data for up to 24 hours. A
sampling frequency of 64Hz was selected for this study as this is sufficiently larger than the
20Hz sampling required to assess daily activity [26]. A number of previous activity
classification studies have used wavelet analysis to derive features from accelerometer
signals collected at relatively high sampling frequencies (>250Hz). However, for this study
64Hz was chosen as this is a realistic sampling frequency which could be implemented by an
off-the-shelf activity monitor. No anti-aliasing filtering was applied to the acceleration data.
For each subject, data was collected with three activity monitors. These were attached to
waist (at the sacrum), the thigh (just above the knee) and the ankle (just above the lateral
maleollus). To secure each unit in place specialised bandage (FabriFoam®) was first
positioned around each of the body segments and the activity monitors, which were backed
with Velcro®, adhered to the underwrapped bandage. Once in position, additional bandage
was then wrapped over each sensor to ensure no movement could occur from overlying
clothing. This method of attachment has been illustrated in Figure 1 for the ankle and thigh
placement.

Frequently asked questions

Q1. What contributions have the authors mentioned in the paper "A comparison of feature extraction methods for the classification of dynamic activities from accelerometer data" ?

In this paper, the authors present a comparison of fourteen methods to extract classification features from accelerometer signals. Furthermore, the authors compared the classification accuracy for each feature set across different combinations of three different accelerometer placements. 

Q2. What are the future works mentioned in the paper "A comparison of feature extraction methods for the classification of dynamic activities from accelerometer data" ?

Further work is required to determine the most appropriate features sets for other subjects groups, such as the elderly or neurologically impaired. ~18~ 

Q3. What is the popular method of evaluating the accuracy of the kNN classifier?

Cross validation is a popular statistical resampling procedure [44] and the authors use it here to evaluate the accuracy of the kNN classifier for a given set of features. 

Q4. What is the way to collect and store accelerometer data?

With recent advances in miniaturised sensing technology, it is now possible to collect and store acceleration data from individual body segments over extended periods of time. 

Q5. What was used to secure the monitors in place?

To secure each unit in place specialised bandage (FabriFoam®) was first positioned around each of the body segments and the activity monitors, which were backed with Velcro®, adhered to the underwrapped bandage. 

Q6. What is the role of the accelerometer in the prediction of fall risk?

Ambulatory monitoring of various activities, including the time spent in sit-stand transitions have shown promise as predictors of fall-risk [8]. 

Q7. What is the role of portable systems in a ubiquitous computing scenario?

In addition to health-related applications, portable systems which can accurately identify the activity of the user have the potential to play a fundamental role in a ubiquitous computing scenario [9, 10]. 

Q8. How many steps did the subjects have to do to perform the first activity?

For the first of these two activities, subjects were instructed to perform a gentle jog over a 50m distance and for the second to perform a fast run over the same distance. 

Q9. What was the method used for the video problem?

The video method, used~17~in this study, was selected as it was believed to be more accurate than self observation by the subject. 

Q10. What is the role of self reported measures in epidemiological studies?

As self reported measures have been shown to be unreliable [4, 5], systems for activity profiling are beginning to play an important role in large-scale epidemiological studies in this area [6, 7]. 

Q11. What frequency bands were used to calculate the features?

Both Nyan et al. [24] and Sekine et al. [22] collected data at 256Hz, therefore as before, wavelet coefficients corresponding to appropriate frequency bands were used to calculate of each of the features. 

Q12. What is the highest classification accuracy for a single sensor?

The highest classification accuracy for a single sensor was obtained for the FFT component feature set and the ankle-mounted sensor. 

Q13. What is the method for calculating the classification accuracy of a single sensor?

for the three-activity problem, the highest classification accuracy for a single sensor (97±3%) was obtained using FFT components derived from the ankle-mounted unit. 

Q14. What is the way to interpret the accelerometer data?

Although this technology offers the ideal platform for monitoring daily activity patterns, effective algorithms are also required to interpret the accelerometer data in the context of~3~different activities.