Human Activity Recognition from Triaxial Accelerometer Data

Feature Extraction and Selection Methods for Clustering of Physical Activities

es Machado

, Ricardo Gomes

, Hugo Gamboa

1,2

and V

ıtor Paix

CEFITEC, Physics Department, FCT-UNL, Lisbon, Portugal

PLUX - Wireless Biosignals, Lisbon, Portugal

Champalimaud Foundation, Lisbon, Portugal

Keywords:

Physical Activity Recognition, Signal Processing, Feature Extraction, Feature Selection, Unsupervised

Learning.

Abstract:

The demand for objectivity in clinical diagnosis has been one of the greatest challenges in Biomedical Engi-

neering. The study, development and implementation of solutions that may serve as ground truth in physical

activity recognition and in medical diagnosis of chronic motor diseases is ever more imperative. This paper

describes a human activity recognition framework based on feature extraction and feature selection techniques

where a set of time, statistical and frequency domain features taken from 3-dimensional accelerometer sensors

are extracted. In this paper, unsupervised learning is applied to the feature representation of accelerometer

data to discover the activities performed by different subjects. A feature selection framework is developed in

order to improve the clustering accuracy and reduce computational costs. The features which best distinguish

a particular set of activities are selected from a 180

- dimensional feature vector through machine learning

algorithms. The implemented framework achieved very encouraging results in human activity recognition: an

average person-dependent Adjusted Rand Index (ARI) of 99.29% ± 0.5% and a person-independent ARI of

88.57% ± 4.0% were reached.

1 INTRODUCTION

The constant concern with the human physical and

psychological well-being has been the drive for re-

search studies that have led to a promising evolution

of medicine and engineering. The study, development

and implementation of solutions that may serve as

ground truth in physical activity recognition and in

medical diagnosis of chronic motor diseases is ever

more imperative. In this paper, a Human Activity

Recognition (HAR) framework is developed using a

wearable 3-dimensional accelerometer sensor. The

main focus of this paper is to understanding the sig-

nals produced by a Triaxial Accelerometer (TA), in-

terpreting them in the context of human movement

and identifying relevant parameters from the data.

The versatility of the algorithm enables the identiﬁca-

tion of relevant features able to recognize simple daily

activities. We obtain a 180

- dimensional feature

vector from statistical, time and frequency domains.

The dimensionality of the feature vector should be as

small as possible by reducing the amount of irrele-

vant and redundant information in the data, not only

to reduce the computation complexity, but also to ob-

tain better clustering performance. The remainder of

the paper is organized as follows: Section 2 describes

the background and related work. The importance of

objective monitoring human movement is discussed.

That section also presents an overview on other stud-

ies about HAR with wearable sensors. Section 3 ex-

plains the composition of the TA signal. The signal

is made up of several components, and each of these

is examined. The difﬁculties in distinguish between

the different signal components are discussed. Sec-

tion 4 describes the proposed methodology used in

this work to extract and select features based on mo-

tion data. Section 5 describes the architecture of the

acquisition system and the obtained results. Section 6

presents the conclusions obtained from the investiga-

tion and some future research directions.

2 BACKGROUND

In recent decades, there has been an increasing in-

terest in the use of Accelerometry (ACC) to moni-

155

Machado I., Gomes R., Gamboa H. and Paixão V..

Human Activity Recognition from Triaxial Accelerometer Data - Feature Extraction and Selection Methods for Clustering of Physical Activities.

DOI: 10.5220/0004749801550162

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2014), pages 155-162

ISBN: 978-989-758-011-6

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

tor human behaviour. Monitoring human movement

can provide valuable information on a patient and

some parameters of movement can provide informa-

tion of health status, rate of rehabilitation and other

potentially useful clinical data. The advance of tech-

nology has helped the development of accelerome-

ters of small size and low cost, making them a very

convenient tool for monitoring subjects. One of the

key point is the diversity of areas where ACC has

been used in the past. The most studied have be-

ing: metabolic energy expenditure, Physical Activity

(PA), balance and postural sway, sit-to-stand transfers

(which is an important indicator for postural instabil-

ity) and detection of falls. The use of accelerometers

has also allowed to help on diagnose of a number of

diseases such as Parkinson’s Disease (Palmerini et al.,

2013), Autism Spectrum Disorder, (Bandini et al.,

2013) and Depression (Phillips and McAuley, 2013).

3 TRIAXIAL ACCELEROMETER

SIGNAL

The signal measured by each ﬁxed-body ac-

celerometer is a linear sum of, approximately, three

components (Mathie, 2003):

• Body Acceleration Component: acceleration re-

sulting from body movement;

• Gravitational Acceleration Component: accelera-

tion resulting from gravity;

• Noise intrinsic to the measurement system.

The ﬁrst two components provide different infor-

mation about the wearer of the device: the Gravita-

tional Acceleration (GA) provides information about

the space orientation of the device, and the Body Ac-

celeration (BA) provides information about the move-

ment of the device. The separation of the information

regarding the movement of the device - Body Accel-

eration Component - is important, however these two

components have overlapping frequency spectra. The

BA component ranges from above 0 Hz to possibly

up 20 Hz, but is mostly contained in the range above

0 and below 3 Hz. This range overlaps the area cov-

ered by the GA component, which goes from 0 to

several Hertz. It is possible to approximately sepa-

rate the BA and the GA components with some ﬁlter-

ing. A wide range of different ﬁlters types with differ-

ent characteristics and different windowing percent-

ages were tested in previous studies, as in (Mathie,

2003), in order to determine their ability to differenti-

ate the components of the acceleration signal. In the

presented study, a cut-off frequency of 0.25 Hz was

chosen, as it is consistent with the frequencies used in

other research works. (Smeja and Muller, 1997) and

(Foerster and Fahrenberg, 2000) choose to use 0.5 Hz,

while (Khan et al., 2010) choose 0.1 Hz. In the pre-

sented study, in order to isolate the BA component,

a second-order Butterworth High-Pass ﬁlter with cut-

off frequency of 0.25 Hz is used. Figure 1 illustrates

each component of a typical recording from the ac-

celerometer showing seven minutes of motion data

where the subject is asked to perform seven speciﬁc

tasks.

The placement of the accelerometer is another im-

portant point of discussion. A device that is to be

worn over extended periods must be designed to be

as simple to put on and comfortable to wear in order

to encourage compliance of patients. General body

motion can be measured with a single accelerometer

placed close to the body’s center of mass, which is lo-

cated within the pelvis (Liu, 2013). The advantage of

placing the accelerometer attached to the waist is that

it allows the monitoring of accelerations near the cen-

ter of mass. Any movement of the body will cause the

center of mass to shift. This study aims to develop a

HAR framework, for a waist mounted accelerometer

based system.

Figure 1: Body and Gravitational Acceleration for each axis

of accelerometer sensor.

4 PROPOSED METHODS

Different segmentation methods can be applied to

time-series data which enhance signal behaviour and

enable the gather of useful information from contin-

uous stream of data such as timing and sliding win-

dows. For activity recognition, where accelerome-

ter data is windowed, the choice of the number of

frames is guided by a trade-off between information

and resolution. The accelerometer data was collected,

cleaned, and preprocessed to extract features that

characterize different samples data windows. Cluster-

BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

156

ing mechanisms separate and organize unlabeled data

into different groups whose members are similar to

each other in some metric. Different approaches gen-

erally lead to different clusters. Even for the same al-

gorithm, the parameter identiﬁcation or the sequence

of input patterns may affect the ﬁnal results. These

assessments should be unbiased. In this work, the

K-Means Clustering Algorithm (Lloyd, 1982) and a

squared Euclidean distance metric were used.

4.1 Feature Design

The HAR strategy depends essentially on the set of

features that are extracted from the signal. TA are

made up of three separated accelerometer data time

series, one time series for acceleration on each orthog-

onal axis ACC

, ACC

and ACC

. Complementary to

the three axes data, an additional time series, ACC

tot

have been obtained by computing the magnitude of

the acceleration, Equation 1:

ACC

tot

ACC

+ ACC

(1)

Each time series ACC

, with i = x, y, z has been ﬁl-

tered with a second-order Butterworth High-Pass ﬁl-

ter with cut-off frequency of 0.25 Hz in order to sepa-

rate low frequencies component and high frequencies

component as suggested in (Mathie, 2003) and (Man-

nini and Sabatini, 2010). This way, for each time se-

ries, three extra time series BA

are obtained, with i =

x, y, z, representing the time series with body acceler-

ation component. Finally, features from each one of

the time series are extracted.

4.2 Accelerometer Signal Annotation

In unsupervised learning, the motion data has to be

annotated to compute the performance of the algo-

rithm. If true class labels are known, the validity of

a clustering can be veriﬁed by comparing the pre-

dicted labels and the true labels. An aspect of activ-

ity recognition that has been greatly explored is the

method of annotating sample data that can be used

to compute the performance of the clustering method.

Many experiments use unsupervised learning meth-

ods and apply manually annotated test data to eval-

uate their performances. In other cases, the experi-

menters told the participants in which order the spec-

iﬁed activities should be performed, so the correct

activity labels were identiﬁed before the sensor data

was even collected. Still in other studies, the raw

sensor data is manually inspected in order to anno-

tate it with a corresponding activity label (Wren and

Tapia, 2006). In the presented study, participants were

continuously observed during experiments and an ob-

server was stating starting/ending time of each activ-

ity. The subjects know in which order the speciﬁed

activities should be performed and latter, raw sensor

data was manually inspected in order to annotate it

with a corresponding activity label. For each signal,

an annotation, in JavaScript Object Notation (JSON)

(Crockford, 2006) is created, with i the number of ac-

tivities, Scheme 2:

”Labels” : [l

, , . . . , l

”Initial Times” : [init

, . . . , init

]

”End Times” : [end

, . . . , end

]

(2)

The dictionary has information about the number

and label of the movements that took place and the

time intervals that delimit them. Each label corre-

sponds to one, and only one, activity, regardless of the

subject. The input is an array with the initial and ﬁ-

nal times of each activity. It also receives as input the

window size and the considered overlap percentage.

4.3 Feature Extraction

Recognizing human activities depends directly on

the features extracted for motion analysis. A set of

features, which will most efﬁciently and meaning-

fully represent the information that is important for

analysis and the clustering process, is performed. In

this section, tests were made in order to assess the

following parameters:

• The inﬂuence of the signal’s window size on the

clustering performance.

• The inﬂuence of the free parameters in that same

performance.

• The best feature combination that leads to a better

performance of the implemented algorithm.

A dictionary of the extracted features from the

motion data, was created, in a JSON format (Crock-

ford, 2006). For each feature, the following informa-

tion was collected: Description, Imports, Use, Metric,

Free Parameters, Parameters, Number of Features,

Function, Source and Reference. Table 1 shows the

high level list of features considered in the presented

study. The implemented dictionary divides the fea-

tures into statistical, temporal and spectral domains.

By manipulating this dictionary, the clustering algo-

rithm can be easily tested with a different combina-

tion of features. To compute the feature vector the

following inputs are needed: motion data, window

length of the signal, sampling frequency of the data

HumanActivityRecognitionfromTriaxialAccelerometerData-FeatureExtractionandSelectionMethodsforClustering

ofPhysicalActivities

157

Table 1: Statistical, Temporal and Spectral Domain Fea-

tures.

Statistical Domain

Kurtosis

Skewness

Mean

Standard Deviation

Interquartile Range

Histogram

Root Mean Square

Median Absolute Deviation

Temporal Domain

Zero Crossing Rate

Pairwise Correlation

Autocorrelation

Spectral Domain

Maximum Frequency

Median Frequency

Cepstral Coefﬁcients

Power Spectrum

MFCC

Fundamental Frequency

Power Bandwidth

acquisition, a feature’s dictionary, a matrix of free

parameter combinations and the considered overlap

percentage. For each ACC axis, this function goes

through each window of the signal, with the consid-

ered window length and overlap percentage and com-

putes a feature matrix with n-samples by m-features

dimension. For each signal, three new ﬁles were cre-

ated: one with the features information per window,

one with the names of the features that were extracted

for the respective clustering test and another with the

label of the activity corresponding to each window.

The sensor acceleration signal is made up of three

separated accelerometer data time series and comple-

mentary to the three axes data, an additional time se-

ries have been obtained by computing the magnitude

of the acceleration, so four signal vectors are con-

sidered. From each window, a vector of features is

obtained by calculating features from the statistical,

time and frequency domain. This way, a 180

- di-

mensional feature vector is obtained: from each one

of the four signal vectors, we compute ﬁfteen features

with only one output and three features (histogram,

cepstral coefﬁcients and mel-frequency cepstral coef-

ﬁcients (MFCC)) with ten outputs each.

Because the scale factors and units of the features

described above are different, all the features must

be normalized to zero mean and unit variance, before

proceed to the feature selection stage.

4.4 Feature Selection for Motion Data

A large number of features can usually be measured

in many pattern recognition applications. However,

not all features are equally important for a speciﬁc

task. For each signal, different combinations of fea-

tures, free parameters of these features and window

size of the signal can be tested, in order to evalu-

ate the performance of the implemented clustering al-

gorithm. Optimal features are identiﬁed depending

on the resulting clustering accuracies for each feature

subset.

4.4.1 Free Parameters of Features Set

In order to make the implemented code versatile and

the least subjective as possible, a matrix with the val-

ues of all the possible combinations that these param-

eters can take, was created. No window size value

was stipulated, but a combination of different values

from a growing logarithmic scale can be tested. Ac-

cording to Table 2, tests were made in which the win-

dow size ranged from 1000 to 4000 samples, in a log

scale. For each window size, different performances

were obtained. Tests were made to determine the free

parameters in each activity, that allow a better activity

recognition performance. Examples of free parame-

ters are the number of bins or the range of the im-

plemented histogram. The values given to these pa-

rameters will dictate the performance obtained by the

clustering algorithm. In this way, a 486-dimensional

free parameter combinations vector was obtained.

Table 2: Possible combinations of free parameters and win-

dow size values.

Free Parameter Range Combinations

Window Size [1000 ; 4000] 3

Bins of Histogram [10 ; 20] 3

Range of Histogram [1 ; 3] 2

Cepstral C. [1 ; 11] 3

MFCC [10 ; 20] 3

Power Bandwidth [10 ; 20] 3

4.4.2 Graphical Perception of Features

Visualizations

A technique for the visualization of time series data

and evaluate their effect in value comparison tasks

was described in (Heer et al., 2009). In order to visu-

ally analyse each feature’s behaviour throughout dif-

ferent activities, horizon graphs are used. This pro-

cedure ensures a visual perception of the features that

better separate certain activities, those which do not

change their value between activities and those which

only add redundant information. Figure 2 shows an

example of a horizon graph generated for a matrix of

features, resulting from an ACC signal composed by

seven distinct activities. Each activity lasts about one

BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

158

minute and we consider 4000 samples for the win-

dow size of the signal. It is possible to quantitatively

compare the behaviour of each feature in each activ-

ity. First, the area between data curve and zero y-

axis is ﬁlled in so that dark reds are very negative and

dark blues are very positive. Then, negative values

are ﬂipped and coloured red, cutting the chart height

by half. Finally, the chart is divided into bands and

overlaid, again halving the height.

4.5 Unsupervised Learning

Machine learning algorithms based on the feature rep-

resentation of accelerometer data have become the

most widely used approaches in PA prediction (I. H.

Witten, E. Frank and Hall, 2011). In this work, unsu-

pervised learning is used to distinguish different ac-

tivities. Clustering mechanisms separate and orga-

nize unlabeled data into different groups whose mem-

bers are similar to each other in some metric. This

method receives the number of clusters to form as

well as the number of centroids to generate. In the

presented study, the number of clusters was deﬁned,

a priori, a priori, from the designed protocol of the

performed activities. A good clustering methodology

will produce clusters in which the intra-class similar-

ity is high and the inter-class similarity is low. The

K-Means Clustering Algorithm (Lloyd, 1982) gives a

single set of clusters, with no particular organization

or structure within them.

5 DATA ACQUISITION AND

RESULTS

The experiments have been carried out with a group

of 8 volunteers within an age range of 16-44 years.

The test consists in performing of a gym circuit. Each

person performs seven activities in sequence lasting

about one minute each - standing, sitting, walking,

running, lying down (belly up), lying down (right side

down) and lying down (left side down), wearing an

accelerometer on the waist. Using this system, data

with 3-axial acceleration at a constant rate of 800 Hz

and 12 bits of resolution was acquired. The data ac-

quisition was performed with OpenSignals platform

(Gomes et al., 2012) and saved in a h5 format. The

collected data was processed ofﬂine using Python

Programming Language (Oliphant, 2006). Clustering

tests are performed, individually, for each subject and

with the respectively concatenated data: in a subject-

dependent and a subject-independent context. To

evaluate the subject-dependent accuracy of the pro-

posed algorithm, the K-Means Clustering Algorithm

(Lloyd, 1982) was performed for each subject data.

Given the knowledge of the ground truth class assign-

ments (labels true) and the clustering algorithm as-

signments of the same samples (predicted labels), the

adjusted Rand index (ARI) is a function that measures

the similarity of the two assignments, ignoring per-

mutations and with chance normalization. The ARI

was calculated to obtain the performance of the clus-

tering method. An average person-dependent accu-

racy of 99.29% and standard deviation of 0.5% were

obtained, with a window size of 4000 samples and the

best set of features: mean, autocorrelation, root mean

square and MFCC. High accuracies are reached for

all subjects. The subject-independent performance

was also evaluated with K-Means Clustering Algo-

rithm (Lloyd, 1982). A person-independent accuracy

of 88.57% and standard deviation of 4.0% were ob-

tained, with a window size of 4000 samples and the

best set of features: mean, autocorrelation, root mean

square and MFCC.

Table 3: Clustering Performance (mean value) as a function

of different window length extracted from the best set of

features.

Window Size Adjusted Rand Index (%)

1000 samples 89.73% ±0.4%

2000 samples 97.42% ±0.9%

4000 samples 99.29% ±0.5%

Table 3 shows the obtained performance for each

value of window size, considering the best imple-

mented set of features: mean, autocorrelation, root

mean square and MFCC. An average of the perfor-

mances obtained for the 8 subjects was calculated.

Based on these results, the HAR system reaches

an accuracy between 89.73% ±0.4% and 99.29%

±0.5%, with 1000 and 4000 samples, respectively.

5.1 Classiﬁcation-based Evaluation:

Proposed Metric

A new metric for assessing the obtained results

from unsupervised techniques, a classiﬁcation-based

evaluation metric, was developed. Initially, a confu-

sion matrix that contains information about true and

predicted labels done by a clustering method was

constructed. Once the clustering algorithm randomly

associates the clustering results to non-annotated

groups, the Algorithm, Best Cluster Permutation,

that links these groups to their corresponded activity,

was implemented. The presented Algorithm receives

the confusion matrix with a random assignment and

goes through each row of the matrix and stores the

HumanActivityRecognitionfromTriaxialAccelerometerData-FeatureExtractionandSelectionMethodsforClustering

ofPhysicalActivities

159

Figure 2: Horizon Graph - Time Series Visualization Technique.

index that contains the maximum value of each row.

Algorithm: Best Cluster Permutation.

Input: Input: confusion matrix with a

random assignment.

Output: confusion matrix correctly

assigned.

It is checked whether the index is unique through-

out the matrix. If the index is unique, it makes the

direct correspondence between the vector of true and

predicted labels. Otherwise, it checks the index with

the maximum value, and assigns it. The process is re-

cursively repeated. After obtaining the swap vector,

the matrix with the labels already associated is recon-

structed. Table 4 shows the confusion matrix for this

study where label i, with i = {1, 2, ...,7}, corresponds

respectively to: standing, sitting, walking, running,

lying down (belly up), lying down (right side down)

and lying down (left side down). For the concatenated

data, the algorithm successfully distinguish all activi-

ties.

6 CONCLUSIONS AND FUTURE

WORK

The continuously need to obtain more information,

more efﬁcient, more quickly and with less interven-

tion from an expert has led to a growing application

of signal processing techniques to motion data.

During the experiment, acceleration signals were

collected from a waist mounted accelerometer based

framework. In the presented study, a methodology to

search for the best features able to classify different

physical activities was presented. The techniques

that operate on the statistical, time and frequency

domains, as well as on data representations that can

be used to discriminate between user activities such

as Horizon Plot were described. The obtained results

in clustering accuracy of HAR were very encour-

aging: an average person-dependent ARI (Santos

and Embrechts, 2009) of 99.29% and a person-

independent ARI of 88.57% were reached. The

major achievements of the current work, compared to

the state of the art are: the presented study performs

tests in intra and inter subject context; a set of 180

features was implemented, which are easily selected

to test different groups of subjects and different

activities and the implemented algorithm does not

stipulate, a priori, any value for window length of the

signal or overlap percentage, but performs a search

to ﬁnd the best parameters that deﬁne the speciﬁc

data. A clustering metric based on the construction

of the data confusion matrix was also proposed. The

presented research leaves a few opened questions, to

be explored in the future:

• Bigger Timespan. Week Long Acquisitions of

Movement.

• More Data. Increase the Number of Subjects and

Applications.

• More Computing Power. Use Parallel Comput-

ing Infrastructures on the Data Collected.

• More Discoveries. Detect the Behaviour changes

and annotate those changes.

In the future, this framework should be tested on

other intensity varying activities and across more sub-

jects. For example, test it on individuals running and

walking at a greater range of intensity levels. ACC

data obtained from wearable accelerometers can be

synchronized with the activity of daily living data

recorded by such monitoring systems to better de-

scribe the information of human mobility, behavioural

pattern and functional ability that encompass the im-

portant parameters regarding the overall health status

of an individual.

BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

160

Table 4: Confusion Matrix, in percentage, for concatenated data, where Lying Down

(1)

is lying down (belly up), Lying

Down

(2)

is lying down (right side down) and Lying Down

(3)

is lying down (left side down).

Standing Sitting Walking Running Lying

Down

(1)

Lying

Down

(2)

Lying

Down

(3)

Standing 92.1±3.2 0.0±0.0 0.0±0.0 0.0±0.0 5.4± 2.3 1.3±0.9 1.1±0.8

Sitting 28.3±6.9 68.0±5.9 1.1±0.6 0.3±0.7 0.1± 0.3 1.6±0.7 0.6±1.3

Walking 0.0±0.0 0.4±0.5 99.5±0.5 0.1±0.3 0.0±0.0 0.0±0.0 0.0±0.0

Running 0.0±0.0 0.0±0.0 0.3±0.4 99.4±0.7 0.3±0.4 0.1±0.3 0.0±0.0

Lying Down

(1)

0.9±0.6 2.0±1.1 0.1±0.3 0.0±0.0 82.1±1.9 7.5±1.4 7.4±1.3

Lying Down

(2)

0.0±0.0 0.0±0.0 0.1±0.3 0.5±1.0 1.1±0.0 90.4±0.9 8.0±1.3

Lying Down

(3)

0.0±0.0 0.0±0.0 0.0±0.0 0.0±0.0 0.1±0.3 0.4±0.5 99.5±0.5

The main challenge for future work in this area

will be the development of features and recognition

strategies that can work in an ambient assisted living

under a wide variety of environmental conditions.

REFERENCES

Ankerst, M., Breunig, M. M., Kriegel, H.-P., and Sander, J.

(1999). Optics: ordering points to identify the cluster-

ing structure. SIGMOD Rec., 28(2):49–60.

Bandini, L. G., Gleason, J., Curtin, C., Lividini, K., An-

derson, S. E., Cermak, S. A., Maslin, M., and Must,

A. (2013). Comparison of physical activity between

children with autism spectrum disorders and typically

developing children. Autism, 17(1):44–54.

Bouten, C. V., Koekkoek, K. T., Verduin, M., Kodde, R.,

and Janssen, J. D. (1997). A triaxial accelerometer

and portable data processing unit for the assessment of

daily physical activity. IEEE Transactions on Biomed-

ical Engineering, 44(3):136–147.

Bugan

e, F., Benedetti, M. G., Casadio, G., Attala, S., Biagi,

F., Manca, M., and Leardini, A. (2012). Estimation

of spatial-temporal gait parameters in level walking

based on a single accelerometer: Validation on normal

subjects by standard gait analysis. Comput. Methods

Prog. Biomed., 108(1):129–137.

Crockford, D. (2006). The application/json media type for

javascript object notation (json). RFC 4627, IETF.

Donalek, C. (2011). Supervised and Unsupervised Learn-

ing. (April).

Ester, M., peter Kriegel, H., S, J., and Xu, X. (1996).

A density-based algorithm for discovering clusters in

large spatial databases with noise. pages 226–231.

AAAI Press.

Foerster, F. and Fahrenberg, J. (2000). Motion pattern and

posture: correctly assessed by calibrated accelerome-

ters, page 32(3):450457.

Frey, B. J. and Dueck, D. (2007). Clustering by

passing messages between data points. Science,

315(5814):972–976.

Gomes, R., Nunes, N., Sousa, J., and Gamboa, H. (2012).

Long term biosignals visualization and processing. In

BIOSIGNALS, pages 402–405.

Heer, J., Kong, N., and Agrawala, M. (2009). Sizing the

horizon: The effects of chart size and layering on the

graphical perception of time series visualizations. In

ACM Human Factors in Computing Systems (CHI).

I. H. Witten, E. Frank and Hall, M. A. (2011). Data Mining:

Practical Machine Learning Tools and Techniques.

Khan, A. M., Lee, Y.-K., Lee, S. Y., and Kim, T.-

S. (2010). A triaxial accelerometer-based physical-

activity recognition via augmented-signal features and

a hierarchical recognizer. Trans. Info. Tech. Biomed.,

14(5):1166–1172.

Liao, L., Fox, D., and Kautz, H. (2005). Location-based ac-

tivity recognition using relational markov networks.

In Proceedings of the 19th international joint confer-

ence on Artiﬁcial intelligence, IJCAI’05, pages 773–

778, San Francisco, CA, USA. Morgan Kaufmann

Publishers Inc.

Liu, A. D. (2013). The Effect of Sensor Mass, Sensor Lo-

cation, and Delamination Location Of. PhD thesis,

California Polytechnic State University.

Lloyd, S. P. (1982). Least squares quantization in pcm.

IEEE Transactions on Information Theory, 28:129–

137.

Mannini, A., Intille, S. S., Rosenberger, M., Sabatini, A. M.,

and Haskell, W. (2013). Activity recognition using a

single accelerometer placed at the wrist or ankle. Med

Sci Sports Exerc.

Mannini, A. and Sabatini, A. (2010). Machine learn-

ing methods for classifying human physical activities

from on-body sensors.

Mathie, M. (2003). Monitoring and interpreting human

movement patterns using a triaxial accelerometer.

Oliphant, T. (2006). Guide to Numpy. Tregol Publishing.

Palmerini, L., Mellone, S., Avanzolini, G., Valzania, F., and

Chiari, L. (2013). Classiﬁcation of early-mild subjects

with parkinson’s disease by using sensor-based mea-

sures of posture, gait, and transitions. In AIME, pages

176–180.

HumanActivityRecognitionfromTriaxialAccelerometerData-FeatureExtractionandSelectionMethodsforClustering

ofPhysicalActivities

161

Peeters, G. (2004). A large set of audio features for

sound description (similarity and classiﬁcation) in the

CUIDADO project.

Phillips, S. M. and McAuley, E. (2013). Physical activ-

ity and fatigue in breast cancer survivors: A panel

model examining the role of self-efﬁcacy and depres-

sion. Cancer Epidemiology Biomarkers & Prevention,

22(5):773–781.

Pour, G., Troped, P., and Evans, J. (2013). Environment

feature extraction and classiﬁcation for context aware

physical activity monitoring. In Sensors Applications

Symposium (SAS), 2013 IEEE, pages 123–128.

Santos, J. M. and Embrechts, M. (2009). On the use of the

adjusted rand index as a metric for evaluating super-

vised classiﬁcation. In Alippi, C., Polycarpou, M. M.,

Panayiotou, C. G., and Ellinas, G., editors, ICANN (2),

volume 5769 of Lecture Notes in Computer Science,

pages 175–184. Springer.

Simon, B. J., Errico, J. P., and Rafﬂe, J. T. (2013). Non-

invasive magnetic or electrical nerve stimulation to

treat or prevent autism spectrum disorders and other

disorders of psychological development. US Patent

App. 13/783,319.

Smeja, J. F. F. F. M. and Muller, W. (1997). Assessment

of posture and motion by multichannel piezoresistive

accelerometer recordings. Psychophysiology, page

34(5):607612.

Song, M. R., Lee, Y.-S., Baek, J.-D., and Miller, M. (2012).

Physical activity status in adults with depression in

the national health and nutrition examination survey,

2005–2006. Public Health Nursing, 29(3):208–217.

Wren, C. R. and Tapia, E. M. (2006). Toward scalable ac-

tivity recognition for sensor networks. In In Lecture

Notes in Computer Science, pages 168–185. Springer.

Z. Ghahramani (2004). Unsupervised learning. Gatsby

Computational Neuroscience Unit, University College

London.

Zheng, Y., Wong, W.-K., Guan, X., and Trost, S. (2013).

Physical activity recognition from accelerometer data

using a multi-scale ensemble method.

BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

162