Engineering, 2013, 5, 25-30
doi:10.4236/eng.2013.55B006 Published Online May 2013 (http://www.scirp.org/journal/eng)
Motion Classification Usi ng P ropos ed Pri nciple
Component Analysis Hybrid K-Means Clustering*
Ching Yee Yong, Rubita Sudirman, Nasrul Humaimi Mahmood, Kim Mey Chew
Infocomm Research Alliance, Faculty of Electrical Engineerin g, Universiti Teknologi Malaysia, Johor, Malaysia
Received 2013
ABSTRACT
This study investigates and acts as a trial clinical outcome for human motion and behaviour analysis in consensus of
health related quality of life in Malaysia. The proposed technique was developed to analyze and access the quality of
human motion that can be used in hospitals, clinics and human motion researches. It aims to establish how to wide-
spread the quality of life effects of human motio n. Reliability and v alidity are need ed to facilitate su bject outcomes. An
experiment was set up in a laboratory environment with conjunction of analyzing human motion and its behaviour. Five
classifiers and algorithms were used to recognize and classify the motion patterns. The proposed PCA-K-Means clus-
tering took 0.058 seconds for classification process. Resubstitution error for the proposed technique was 0.002 and
achieved 94.67% of true positive for total confusion matrix of the classification accuracy. The proposed clustering algo-
rithm achieved higher speed of processing, higher accuracy of performance and reliable cross validation error.
Keywords: Accelerometer; Gyroscope; Fuzzy; Bayes; Decision Tree
1. Introduction
This study focuses on investigating the human motion
and movement behavior through analyzing their jogging,
walking and throwing patterns, to come out with a better
solution for movement classification and nature behavior
analysis. The methodology of this research is to get the
motion pattern through few sensors attachment on skin
for processing and analysis. The reviews from previous
research on the requirement of experiment design and the
current trend of analysis act as guidance to develop a
good research fram e wor k.
The objective of this study is to investigate the human
motion and movement behavior in order to establish how
widespread the quality of life effects of motion are by
quantifying them. The expected results in terms of the
stability, design, efficient control for mobility will help
researchers to consider the outcomes of a human motion
and movement. This paper presents a novel motion signal
processing technique, and presents ideas for further de-
velopment and recognition, to give researchers ideas of
how they can use human movement in related field for
product development.
This paper is divided into six sections. The first sec-
tion mainly introduces the whole study. It provides the
general overview of the human motion analysis system.
The second section includes the objectives of this study,
which describes the aims that needed to be achieved. It
also discusses the background studies, literature review
and the basic concept in this study. Section 3 discusses
about the study implementation and a specification of the
experiment environment, thorough discussion on the de-
velopmental technique or algorithm and analysis on hu-
man motion. Finally, the last two sections provide the
conclusions, future developments and possible enhance-
ment and improvement.
2. Research Review
Accelerometer, gyroscope and compass sensors are the
most common devices used in movement detection and
analysis system [1]. Introduction of human actions into
digital domain is a primary driver for innovation of mo-
tion functionality. Human motion signal processing tech-
nique, which combines inertial measurement units with
digital signal processing , enables people readily incorpo-
rate motion [2,3]. Description in the next subsection pro-
vides readers with understanding of the sensors combina-
tions used in motion detection and analysis field [4,5].
2.1. Accelerometer
The primary usage of accelerometer is measuring linear
acceleration and tilt while velocity can be obtained by a
single integration and relative distance by a double inte-
gration. The benefit of a accelerometer is that it able di-
*Patterns classification using Proposed PCA-K-Means Classifier.

C. Y. YONG ET AL.
26
rectly measure tilt angle and linear distance based on
acceleration of gravity [6]. The main drawback of the
sensor is it unable to distinguish between acceleration
due to linear movement and acceleration due to gravity.
Problem can be solved by combining accelerometer with
gyroscope sensor.
2.2. Gyroscope
Gyroscope is mainly used to measure absolute rate of
rotation and relative angle by a single mathe matical inte-
gration. The performance is fast and accurate without
corrupted by linear acceleration or magnetic fields.
However, sometimes the integration may lead to errors
over time but it can be solved by combining with accel-
erometer.
2.3. Pattern Recognition
Pattern recognition consists of 6 stages: data collection,
pre-processing, feature extraction, training, classification
and cross validation [7-10].
Data are collected for training and training classifiers.
In order to achieve accurate result, data size should be
large enough to cover all relevant parameters. Training
and testing data should be differen t.
Pre-processing of signals aims to reduce noise and
normalize interests in a data set.
Feature extraction is a process to extract features that
characterize the region of interests. There are many types
of features such as histogram, shape information, texture
information, scale invariant feature transform and many
others. Principle Component Analysis (PCA) was used to
reduce the number of variables, from six to three; to re-
duce the complexity o f large data set.
Classifiers are trained by training data set. There are
two types of classifier: supervised and unsupervised
learning. Classifiers used in this study are Fuzzy, pro-
posed PCA-K-Means, C-Means, Naive Bayes and deci-
sion tree. Classifier should not just memorize class labels
of testing data. This gives 100% accuracy on training
data, but may not work on other unseen data. Hence,
novel testing data is used to test the performance of
trained classifier that able to generalize to novel data.
Cross validation aims to provide more thorough and
accurate evaluation for the classification process. The
large samples data are divided into a few subsets. The
first subset is left out for testing and the rest are used for
training. The process is repeated for each of the subset in
turn. Average accuracies is achieved from all the runs
and this is called N-fold cross validation.
3. Materials and Methods
3.1. Study Sample
Five healthy volunteers were selected inside university
campus for taking part in this study. Their ages are
around 20 years old with normal limbs movement and
significant mobility in everyday routine independent of
any walking aid.
3.2. Experimental Setup
For this study, experimental setup was done using a wire-
less 3-axis accelerometer. This device employs a YEI
3-Space Sensor breakout board for the tri-axial gyro-
scope, accelerometer, and compass sensors in conjunc-
tion with advanced processing and on-board quaternion-
based Kalman filtering algorithms to determine orienta-
tion relative to an absolute reference in real-time in an
enclosure measuring 60 mm × 35 mm × 15 mm. The
device was connected to a laptop using a standard USB
2.0 host system wireless asynchronous serial transmis-
sion.
The subjects wore a wearable sensor on above right
arm which employed of 3 sensors (gyroscope, acceler-
ometer and compass) inside the package. These sensors
were attached firmly on subjects’ skin with a special de-
signed holder.
3.3. Data Collection and Management
In the initial phase of the trial study, experiment was
conducted for three activities, there are jogging, walking
and throwing. Subjects were asked to perform a normal
walking with speed 3.7 ft/s and jogging with speed 6.5
ft/s on a treadmill with regular motor. Throwing was
performed by throwing a paper roll 1.50 m apart. These
activities were performed in a supervised and comfort-
able environment with presence of researcher for time-
stamping the start and end time of activities period.
Subjects were encouraged to perform the jogging, walk-
ing and throwing activities at their own pace and conven-
ience. The whole experiment setup place was ensuring a
relaxing and natural mood for the sake of subjects for
reflective of real world conditions.
The data were transmitted from sensors to the laptop
for further processin g .
3.4. Data Analysis
Data were collected through transmission of a mini wire-
less dongle from the sensors. They were pre-processing
for noise elimination and then ex tracted for classification
as below:
1) Fuzzy C-Means Clustering: This clustering is an it-
erative process. The parameters required for this cluster-
ing are numbers of cluster/class, exponent for the matrix
partition, maximum number of iteration and minimum of
improvement. First, an initial fuzzy partition matrix was
generated and the initial fuzzy cluster centers were cal-
culated. The cluster centers and the membership function

C. Y. YONG ET AL. 27
were updated during each step of iteration to minimize
the objective function for the best location for the clus-
ters. The process stopped either once the maximum
number of iterations was reached or the minimum
amount of improvement specified between two consecu-
tive objective fun c tions was achieved [11-14].
2) PCA-K-Means Clustering: K-Means Clustering is a
function partitions the N X P data matrix into K clus-
ters/classes through a fully vectorized algo rithm. N is the
number of data points while P is the number of variables.
In this project, the sum of all points to cluster centroid
was calculated using proposed Euclidean distances for all
clusters. The Euclidean distance d between two vectors x
and y is:
d = sum ((x-y)2)1/2 (1)
The function was returned the centroids locations for
all the clusters until the minimum sums were achieved.
PCA was hybrid with K-Means clustering to reduce the
variables dimensions.
3) Naïve Bayes: Naive Bayes performs classification
works based on diagonal covariance matrix estimations.
This classifier assumes the variables are conditionally
independent given the class label. It has been found to
work well in practice for many large or small data sets.
First the classifier was modelled using Gaussian distribu-
tion and assumed that multivariate data has normal
Gaussian distribution. Then, the classifier was enhanced
using kernel density estimation, which is a more flexible
nonparametric technique [15].
4) Decision Tree: Decision Tree is an algorithm that
following simple rules, such as “if the Y-axis gyroscope
reading is less than 0.5898, then classify the data as
Walking motion.” It is a nonparametric techniqu e since it
does not require any assumptions about the distribution
of the variables in each class. A set of rules was gener-
ated by training data. Decision Tree used this set of rules
to divide the plane and assign each data to each specific
class [16-18].
5) LDA and QDA: LDA is Linear Discriminant
Analysis and QDA is Quadratic Discriminant Analysis.
Data were classified using default LDA. Some data were
misclassified by the LDA function with drawing X
through the points. The function has separated the plane
into region divided by lines, and assigned different re-
gions to different classes. A grid was created for region
visualization. For some data as for this project, the vari-
ables were not separated well into the correct classes.
Hence, QDA was proposed for the data.
3.5. Instrument Revision
The preliminary set of outcome measures was shown in
this paper. There are 2 sensors used in this experimental
setup: gyroscope and accelerometer. The ability of the
classifiers in differentiating jogging, walking and throw-
ing patterns we re di st i ng ui shed in discussi on pa rt .
Accelerations due to jolting of the sensors if loosely
attached may add noise to the signal. The special de-
signed of sensor holder capable attached firmly to the
subject’s skin to avoid any disturbance.
4. Result and Discussion
Results of the project were displayed in tabular and
graphical form as Figure 1.
Table 1 shows the time consumption for every classify-
er used. Proposed PCA-K-Means clustering used 0.058
seconds for the whole pattern recognition process. It is
the fastest classification process compared with Fuzzy
C-Means (2.32 s), Naive Bayes (3.72 s), decision tree
(3.16 s) and normal K-Means statistical toolbox (0.56 s).
Table 2 shows the resubstitution and cross validation
errors. PCA-K-Means achieved 0.002 and 0.328 for re-
substitution and cross validation errors respectively.
Table 3 shows the confusion matrix and total true
positive for every algorithm used. PCA-K-Means and
decision tree achieved the highest true positive (accuracy)
percentage, 94.67% followed by Naive Bayes kernel
density 86%, QDA 77%, Naive Bayes Gaussian kernel
75% and LDA 64.33%.
-1 -0.500.5 1
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Accelerometer X-axis
Acceleromter Y-axis
Jogging
Walking
Thr ow ing
-5 05 10
-4
-3
-2
-1
0
1
2
3
4
5
6
Gyro sco p e X-axis
Gyroscope Y-axis
Jogging
Walking
Thro win g
Figure 1. Scatter plot for accelerometer and gyroscope data.
Table 1. Time consumption for every classifier used.
Classifier Processing Time (s)
Fuzzy C-Means 2.317
PCA-K-Means (Proposed) 0.058
K-Means Statistical Toolbox0.565
Naïve Bayes 3.723
Decision Tree 3.165

C. Y. YONG ET AL.
28
Table 2. Resubstitution and cross validation errors for every
algorithm.
Algorithm Resubstitution
Error (%) Cross Validation Error (%)
PCA-K-Means (Proposed) 0.002 0.328
Linear Discriminant Analysis 0.357 0.423
Quadratic Discriminant Analysis 0.230 0.253
Naïve Bayes Gaussian Kernel 0.250 0.260
Naïve Bayes Kernel Density 0.140 0.183
Decision Tree 0.053 0.323
Table 3. Confusion matrix for every algorithm used.
Algorithm Confusion Matrix True Positive (%)
100 0 0
11 89 0
PCA-K-Means
(Proposed) 0 5
95
94.67
54 26 20
42 49 9
Linear Discriminant
Analysis 4 6
90
64.33
51 33 16
13 86 1
Quadratic Discriminant
Analysis 5 1
94
77.00
51 32 17
11 81 8
Naïve Bayes Gaussian
Kernel 7 0
93
75.00
64 27 9
0 99 1
Naïve Bayes Kernel
Density 3 2
95
86.00
94 4 2
3 96 1 Decision Tree
6 0 94
94.67
4.1. Fuzzy C-Means Clustering
Figure 2 shows the initial and final fuzzy cluster centers.
The bold number represents the final centers after up-
dated from each iteration iteratively. Throwing always
has the higher peak then followed by jogging and walk-
ing for gyroscope values above 5g.
4.2. PCA-K-Means Clustering (Proposed)
Figure 3 shows the final clusters centroids for all of the
clusters. The bold “+” marks are the final updated cen-
troids locations for every cluster.
4.3. Naïve Bayes
Gaussian distribution and kernel density were hybrid
with Naive Bayes for classification. As shown in Table 2,
errors of resubstitutio n an d cro ss validation were redu ced
for kernel density due to its flexibility characteristic.
-1 -0.500.5 1
-1
-0.5
0
0.5
1
AX
AZ 1
2
3
1
23
-1 -0.500.5 1
-3
-2
-1
0
1
2
3
4
5
6
AX
GX
1 2
31
2
3
-1 -0.5 00.5 1
-4
-2
0
2
4
6
AX
GZ
1
2
31
2
3
-1 -0.500.5 1
-3
-2
-1
0
1
2
3
4
5
6
AZ
GX
1
2
3
1
2
3
-1 -0.500.5 1
-4
-2
0
2
4
6
AZ
GZ
1
2
3
1
2
3
-5 05 10
-4
-2
0
2
4
6
G
X
GZ
1
2
3
1
2
3
Figure 2. 2D initial and final fuzzy cluster centers for two
characteristics (AX-AZ, AX-GX, AX-GZ, AZ-GX, AZ-GZ,
GX-GZ) of the three types of motions (Walking in red, Jog-
ging in green and Throwing in blue).
-1-0.8 -0.6 -0.4 -0.200.2 0.4 0.6 0.81
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Jogging
Walking
Thr ow ing
Figure 3. 2D final cluste r plot with updated centroids of the
three types of motions.
4.4. Decision Tree
Initially, a full 12 level of pruning with 30 terminal nodes
cluttered-looking tree was generated after a series of
rules was applied to each data. The data were then per-
formed resubstitution and cross validation error to create
a simplest and smallest tree as in Fi gure 4.
4.5. LDA and QDA
For data from this study, LDA was not an appropriate
algorithm for classification. QDA perform a quadratic
analysis to the data and the resu bstitution error and cross
validation error were reduced.

C. Y. YONG ET AL. 29
Figure 4. Final pruned tr e e.
In order to obtain a clearer and cleaner approximation
plot, data received from the sensors were pre-processing
using averaging filter to eliminate DC noise and distur-
bance.
5. Discussion
Experimentally, K-Means statistical toolbox took short-
est time for classification process compared among other
classifiers, however, the proposed PCA-K-Means clus-
tering even shorten the process period by 10 times with
lowest resubstitution error and highest true positive for
confusion matrix by hybrid with PCA .
Decision tree classifier achieved 94.67% true positive
as accurate as proposed method, however, decision tree
took 3.165 s for 300 data sample size. The proposed
classifier is able to process data with 54 times faster than
decision tree. Furthermore, the proposed method achieved
31 times accuracy for resubstitution error. In short, deci-
sion tree is not appropriate for analyzing large sample
data.
LDA, QDA, Fuzzy C-Means, Naive Bayes Gaussian
and Naive Bayes kernel density classifiers took longer
time (more than one second) for processing, higher re-
substitution and cross validation error and lower true
positive percentage for confusion matrix. These rates are
very important for a testing in large dataset.
There is a lot to do with this study depending on the
imagination. One but not the only one straight forward
application for this research is motion recognition . It also
can be applied on the incredible thing likes gesture rec-
ognition, behavioural analysis and gait analysis.
There is also a possibility of in corp orating an EEG and
ECG into this study. ECG could involve the con dition of
human body while wearing sensors and EEG could in-
volve condition of human brain activity while performing
task. This data would be collected simultaneously using
the accelerometer.
6. Conclusions and Future Work
The sensor is capable to filter and normalized data using
Kalman filter. Results presenting in scatter plot success-
ful reveal information needed. The attachment of sensors
on subject’s skin was firm without significant distur-
bance. Overall this study completed the objectives from
attachment, detection, orientation, transmission, receiv-
ing, filtering, and analyzing.
The proposed PCA-K-Means classifier is successfully
recognized and classified all three motions data with
shortest period, higher accuracy and lower errors. The
classifier also appropriate for processing large data set
within period.
In order to fully realize this study, there are few things
that could be considered, the main feature of interest is
the data processing unit. All data are process under the
same platform without bias. Further approach need to be
taken in order to achieve a higher aim in this research.
As the initial, the study took place in a laboratory en-
vironment, it was considered appropriate for the initial
phase of the quantitativ e study to be conducted in a simi-
lar environment. Further work is planned to widen the
sample and to encompass different environments in both
the dynamic and transition activities.
7. Acknowledgements
A study of this magnitude depends on the hard work and
commitment of many professionals, and we are pleased
to acknowledge their contributions. The authors are
deeply indebted and would like to express our gratitude
to the Universiti Teknologi Malaysia for supporting and
funding this study under Research University Grant
(QJ13000. 2636.05J69) and MyPhD Scholarship Scheme
from Ministry of Higher Education (MOHE).
REFERENCES
[1] F. Tian, “Leveraging Psychophysical Data in Monitoring
and Analyzing the States of Badminton Players,” ACM.
Atlanta, Georgia, USA, 2010, pp. 930-935.
[2] Nasrul Humaimi Mahmood, Ching Yee Yong, Rubita
Sudirman, Camallil Omar and Kim Mey Chew, “Func-
tional And Health Related Analysis In The Discipline Of
Posthetics,” International Journal of Advances in Engi-
neering & Technology, 2011, Vol. 1, No. 3, pp. 171-179.
[3] C. Y. Yong, K. M. Chew, N. H. Mahmood, R. Sudirman
and C. Omar, “Development and Measurement Properties
of Prosthetics Users’ Survey”, 2011 IEEE Symposium on
Business, Engineering and Industrial Applications (IS-
BEIA2011), 25-29 September 2011, Langkawi, Malaysia,
pp. 570-575.
[4] R. Nalma and J. Canny, “The Berkeley Trocoder: Ambu-

C. Y. YONG ET AL.
30
latory Health Monitoring,” 2009 Sixth International
Workshop on Wearable and Implantable Body Sensor
Networks, 2009, pp. 53-58.
[5] U. Maurer, A. Smailagic and D. P. Siewiorek, “Activity
Recognition and Monitoring Using Multiple Sensors on
Different Body Positions,” International Workshop on
Wearable and Implantable Body Sensor Networks, 2006.
doi:10.1109/BSN.2006.6
[6] C. Y. Yong, Rubita Sudirman and Kim Mey Chew, “Mo-
tion Detection and Analysis with Four Different Detec-
tors,” 2011 Third International Conference on Computa-
tional Intelligence, Modelling & Simulation (CIMSim
2011), Langkawi, Malaysia, 20-22 September 2011, pp.
46-50.
[7] C. Ni Scanail, B. Ahearne and G. M. Lyons, “Long-term
Telemonitoring of Mobility Trends of Elderly People
Using SMS Messaging,” IEEE Trans Inform Tech Bio-
med, 2006, Vol. 10, pp. 34-37.
[8] A. Godfrey, K. M. Culhane and G. M. Lyons, “Compari-
son of the Performance of the Active PALTM Trio Profes-
sional Physical Activity Logger to a Discrete Acceler-
ometer-Based Activity Monitor,” Medical Engineering &
Physic, 2006.
[9] A. C. Bovik, M. Clark and W. S. Geisler, “Multichannel
Texture Analysis Using Localized Spatial Filters,” IEEE
Trans. On Pattern Analysis and Machine Intelligence,
1990, Vol. 12, No. 1, pp. 55-73. doi:10.1109/34.41384
[10] G. R. Cross and A. K. Jain, “Markov Random Field Tex-
ture Models,” IEEE Trans. On Pattern Analysis and Ma-
chine Intelligence, Vol. 5, 1983, pp. 25-39.
doi:10.1109/TPAMI.1983.4767341
[11] R. O. Duda, P. E. Hart and D. G. Stork, “Pattern Classifi-
cation,” John Wiley & Sons, 2 edition, 2001.
[12] D. G. Lowe, “Object Recognition from Local
Sclae-Invariant Features,” In Proc. Int. Conf. on Com-
puter Vision, Vol. 2, 1999, pp. 1150-1157.
[13] M. Vangelis, A. Ion and P. Geogios, “Spam Filtering with
Naive Bayes-Which Naive Bayes?” Third Conference on
Email and Anti-Spam, 2006.
[14] H. Spath, “Cluster Dissection and Analysis: Theory,”
FORTRAN Programs, Examples. Translated by J. Gold-
schmidt, New York: Halsted Press, 1985.
[15] J.-S. R. Jang and C.-T. Sun, “Neuro-Fuzzy Modeling and
Control,” Proceedings of the IEEE, 1995.
[16] L. Breiman, J. Friedman, R. Olshen and C. Stone, Classi-
fication and Regression Trees. Boca Raton, FL: CRC
Press, 1984.
[17] Ching Yee Yong, K. M. Chew, N. H. Mahmood and I.
Ariffin, “Image Processing Tools Package in Medical
Imaging in MATLAB,” International Journal of Educa-
tion and Information Technologies, North Atlantic Uni-
versity Union, NAUN, Vol. 6, No. 3, 2012, pp. 260 -268.
[18] Ching Yee Yong, R. Sudirman, N. H. Mahmood, K. M.
Chew, A. H. AB Rahim and M. N. H. Zainudin, “Time-Frequ
ency Domain and Spectrogram Distribution for Human
Motion and Movement Behaviour Analysis,” icbeb 2012
International Conference on Biomedical Engineering and
Biotechnology, 2012, pp. 943-946.