A ROBUST PHYSICAL EXERCISE RECOGNITION SYSTEM

USING MACHINE LEARNING APPROACH

Abstract: Modern life is becoming more linked to our devices and work is being done
in a more regulated way. As life became more complicated, it is becoming challenging
to keep track of one’s health and fitness which may lead to unexpected illness and
diseases. Moreover, lack of activity monitoring and corresponding reminders is
preventing the adoption of a healthier lifestyle. This research provides an effective
approach for identifying Human Activity by using accelerometer data obtained from
wearable devices. The model automatically finds pattern among 33 different physical
exercises such as running, rowing, cycling, jogging etc. and correctly identifies them.
Principal component Analysis was used on the statistical features in order to make the
system more robust. Classification of the physical exercise was performed on the
reduced features using WEKA. An overall accuracy of 85.51% was obtained using 10-
Fold Cross Validation method and K nearest Neighbor Algorithm while 84% accuracy
for Random Forest. The accuracy obtained was better than previous models and could
improve recognition systems in monitoring user activity more precisely.

Keywords: Physical Activity Recognition, RealDisp Activity Recognition

Wearable Device

1. INTRODUCTION of keeping count in one’s mind technology has

Nowadays, people live and interact in ways that
have lack proper physical activity and these
have brought adverse changes in their health.
Not only it prevented a better lifestyle but it has
become the leading cause of many common
chronic diseases such as obesity, heart failure
and even premature death [1]. In addition, less
serious diseases are contributing to loss of
muscle strength and immense fatigue [2]. In
order to mitigate this issue, 3-5 hours of intense
aerobic exercise is recommended by the United
States Department of health [3]. Furthermore,
with the rapid rise in the number of smart
sensing devices, such as smartphones and smart
watches various high detailed applications
linked to monitoring of personal healthcare,
management of obesity, interactive and
experiential gaming etc., have continually been
made better to meet human standards [4]. But,
in general, both are in short of meeting this
standard regularly and have various significant
causes.
One of the apparent causes is the failure to
monitor where and how many times activity
occurs and time spent on each activity. Instead
accomplished this task using small sensors
inside IOT devices like smart watches and
smartphones. If smart-watches can employ a
faster and efficient recognition system, then it
will reduce both data consumed and time taken.
The recognition system in our devices performs
activity in different ways. When Human
performs an activity, they require a combination
of several basic movements repeatedly. Running
of a child is very different from those of
walking. In practice each activity takes only a
few second to complete and record. Each basic
movement could be involved over any given
duration. So, the study of HAR is very crucial to
make advancements in academics.
The goal of this study is to find a better way
through feature extraction and feature reduction
procedure to produces accurate classification
across a wide range of physical exercises. The
context that was explored is in a gym setting
where people performed 33 different physical
exercises such as walking, running, cycling etc.
Through using adaptive time series
segmentation features were extracted, followed
by reduction into lower dimension using PCA.
Computer Science and Engineering Research Journal

Instead of using a regular fixed sized window, 90.5% for 10 daily activities. In [11] the task of
adaptive segment was used which ensures more inferring activities when travelling by metro was
likely to detect an activity in those respective explored. It showed when two features could be
segments. Reduction Algorithm makes the selected based on feature selection method the
system computational efficient and robust in model can obtain good accuracy.
many scenarios. Finally, 2 classifiers, one of
Our model improves the existing system by
them is Lazy Algorithm while the other
using adaptive sliding window on preprocessed
Ensemble Tree was used to evaluate
accelerometer signals and then normalizing
performance.
those features. Next using pca the reduced
This research can be pointed down to six features were classified using Machine Learning
sections. Section 2 reviews the previous models.
research related works. Section 3 introduces our
3. PROPOSED METHODOLOGY
proposed model and discusses the experimental
steps that were involved. Section 4 portrays the
findings of our study. Lastly, Section 5 gives the
conclusion of the study and addresses future
works.

2. LITERATURE REVIEW
In order to track and count human activity there
are two approach: The wearable sensor
approach and the other one is the smartphone
based. Vu Ngoc in [5] collected five activities
data from smartphone and was recognized
where they obtained an accuracy of 74% for
kNN and 75.3% for ANN respectively. Activity
and sensor was limited and there were some
ambiguity in the data collected. In [6] Rui
proposed a modified full CNN based algorithm
which predicted human activity sequences on
the self-collected opportunist and hospital
dataset. In [7], CNN was applied directly on 16
lower limb activities using 5 sensors and it
showed comparison to using 1 single sensor.
The Researchers in [8] used a tenfold random-
partitioning cross-validation evaluation to
evaluate the system performance and the
number of features were only limited to 3. For
capturing repeated periods in activities Banos et
al used a time window to segment the discrete
signal. The model failed to demonstrate reliable
accuracy for accelerometer data when many
activities were involved. The obtained accuracy
was a little more than 80%. Tuan Le in [9]
extracted related feature from raw data of 30
volunteer performing 6 activity and finally used
IB3 method to improve accuracy by reducing
dimension. 15% improved accuracy was
obtained through Naïve Bayes (91.85%) and
Decision Tree (96%). In [10] Yu-Liang Hsu
used NWFE algorithm on two sensor data
located at two places which are wrist and ankle.
Moreover, the model ensured an accuracy of
An adaptive time series technique [12] of
segment length 200 was used on each of the
sensor magnitude for all the subjects from which
feature were extracted.
3.1 Datasets
To conduct our experiment, we have collected
REALDISP Activity Recognition dataset [8]
from UCI Machine Learning Repository. The
dataset consists recordings from 17 subjects, ten
males and seven females. The recordings
include 13 inertial signals obtained from sensors
located at different part of the body. A
multivariate time series were obtained from the
accelerometer sensors which are sampled at
regular intervals. The experiment consisted of
subjects performing 33 physical Activities in 3
scenarios. The exercise lasted about 15-20 min.
It is important to note that there are about 1900
instances where the user weren’t moving The
count rate of each of the activity after removing
the irrelevant null activity is as follows
Computer Science and Engineering Research Journal
Fig. 2. Total Frequency of each Activity
3.2 Feature Extraction
The obtained initial signal generated from the
accelerometer signals were grouped together
into segments of fixed length [12]. Then the
feature vector was measured having following
features:
Considering the sample 𝑥1, 𝑥2, 𝑥3, 𝑥4, 𝑥𝑛 the
arithmetic mean x is the sum of the sampled
values divided by the total number of items.
𝑥̅ √ ∑ 𝑥 (1)
The uppercase Σ and the lowercase letter σ are
used to represent standard deviation. The
formula for the sample standard deviation is
∑ ( )
√ (2)
( )
The quantity of distribution for numerical data
can be measured using histogram.
∑ (𝑥 ) (3)
An amplitude RMS can be expressed as
√ ∑ 𝑥 (4)
To obtain how spread the values are in the
dataset we are to calculate the average distance
between each data value and their mean.
∑|
3.3 Z score- normalization
Following by the arrangement of similar
features by combining all the 17 subjects, the
feature was normalized to create a normal
distribution upon the entire set of features. It is
done before reduction phase in order to
maximize the variance of the components. The
equations are
(6)
Where,
√ ∑ (𝑥 ) (7)
∑(𝑥 ) (8)
3.4 Principal Component Analysis
PCA is a useful dimension reduction algorithm
where orthogonal transformation of the data
converts correlated variables into linearly
uncorrelated features called main components.
Through reduction of original data features, the
amount of noise is reduced, and in addition it
diminishes the training data to let the model
become computationally faster. If number of
observations made with p is n, then number of
separate main components is min (n 1, p). The
resulting vector contains n observations having
different variances.[13] The linear combination
of the variables 1, 2, is the first principal
components Y1 represented by the equation
below
𝑌1 = 𝑎11𝑍11 + 𝑎12 12 + ⋯+ 𝑎𝑝𝑍 (9)
The first principal component is calculated by
choosing large values for weights a11, a12, alp
once could make the variance of Y1 as large as
possible. In addition, it is the maximum
variance. So, to limit the size of variance
weights are calculated with sum of squares as 1.
𝑎21𝑋21 + 𝑎22𝑋12 + ⋯+ 𝑎22 22 = 1 (10)
Similarly, the second principal component is
calculated. Collectively, all these original
|
(5)
variables are transformed into the main
components.
Y = XA (11)
3.5 K nearest Neighbor
Computer Science and Engineering Research Journal

KNN is a non-parametric supervised learning

algorithm that is used in both classification and
regression problems. It is also known as lazy
learning algorithm because it does not classify
based on the training data points. The purpose of
this algorithm is to classify a new sample based
on its features and labeled training samples.
Given a query point 𝑥 , k training points closest
in distance (Euclidean distance) to 𝑥 are found.
Based on most of the neighbors found, the new
query is classified to its cluster. Any ties in
voting are broken at random. So, no specific
training model is used for classification. KNN
algorithm works based on nearest neighbor’s
classification and works best on smaller dataset
with fewer features.
The Euclidean distance [14] of two n-
dimensional vectors, x and y, is defined as:
( ) √(𝑥 𝑥 ) (𝑥 𝑥 )
and Manhattan (or city block) distance [15] is
defined as:
The ratio between the true positive values and
the sum of the true positive and false positive
values is the precision
Precision = (16)
The ratio between the true positive values and
the sum of the true positive and false positive
values.
Recall = (17)
F1 score helps to regulate the balance between
precision and recall. It is the harmonic average
of precision and recall. It is formulated as:
F1 Score = (18)
The majority of the activity is where the user
isn’t performing any movement. First, our
⋯ (𝑥 𝑥 ) (12)
model was tested along with this activity as Null
Activity. The results are as follows
Table 4.1 Result with Null Class

( ) √|𝑥 𝑥 | |𝑥 𝑥 | ⋯ |𝑥 𝑥 | (13) Method CA F1 Precision Recall

KNN 92.99% 94.55% 93.77% 91.45%
Random 86.73% 87.04% 86.73% 86.11%
3.6 Random Forest
Forest
Random forest is an ensemble algorithm that Considering a round robin strategy nine subset was
works well for models with low bias (high used training and one subset was used for testing.
feature target relevant relationship) and high Averaging the accuracy obtained from these 10
variance (spread of data). For classification it experiments a final cross-validation result was
joins more than one or similar kind of learning obtained. It is same cross validation method used in
algorithms and predicts labels with much more paper [8]. K-nearest algorithm gives an accuracy of
accuracy. From randomly chosen training set 85.51% while Random Forest Algorithm gives an
decision trees are created. After totaling the accuracy of 84.005%. Experiment was performed on
votes produced the final class of the dataset WEKA [17]. It is observed the K nearest algorithm
performs better to predict Activity recognition on the
tested is decided [16]. Finally computing the
entire dataset for ideal placement scenarios.
average of single tree prediction for each new
object is made. If ̅ is the number of trees used
for bagging then, Method CA F1 Precision Recall
KNN 85.51% 85.3% 85.7% 85.5%
̂ ̅
∑ 𝑝 ( ) (14) Random 84.00% 83.2% 84.3% 84.0%
Forest
4. Experimental Results and Evaluation
TABLE 4.2 Cross validation results
We measured and evaluated our performance as
follows: The bar chart below shows the accuracy of 3
Accuracy is a measure, which determines the different machine Learning Models outperforms
probability that how many results are correctly previous models.
classified
Accuracy = (15)
Computer Science and Engineering Research Journal
Fig 3. Bar chart of 3 Algorithms
Table 4.3 Comparisons with other paper
Name Method Accuracy
Propose Model KNN(reduced 85.51%
feature)
Banos et al. [8] KNN 85%
Vu Ngoc [5] ANN 75.3%
The results obtained with a success rate of
85.51% are better compared with the paper [8]
in Table III. In this study, the evaluated data
includes 33 diverse activity types which are
mostly dynamic in nature.
5. CONCLUSION
To classify Human activity there are many
uncertainties for the best algorithm. Different
combination of signals vibrational features gives
different activity detection rate. Moreover, the
sheer amount of data created by these devices
causes significant amount of strain. Researchers
still debates what features influences the
recognition of Human Activity in Signals. In our
research, a model for successful detection of
Human Activity was proposed. About five
features were used and normalized using PCA.
Classification was done with 2 classifiers. KNN
shows the best Accuracy. Further experiments
with huge volume of irrelevant data where no
activity was identified data shows significant
improvement in result. Possibly results may be
improved using RFE to remove the weakest
feature to reach desired set of features.
Moreover combining smart watch features with
those of smartphone or any wearable sensor can
provide a much better recognition.
REFERENCES
1. F. W. Booth, C. K. Roberts, and M. J.
Laye, "Lack of exercise is a major cause
of chronic diseases," Comprehensive
Physiology, vol. 2, 2012.
2. P. C. Hallal, C. G. Victora, M. R.
Azevedo, and J. C. Wells, "Adolescent
physical activity and health," Sports
medicine, vol. 36, 2006.
3. A. Bulling, U. Blanke, and B. Schiele,
"A tutorial on human activity
recognition using body-worn inertial
sensors," ACM Computing Surveys,
vol. 46, 2014.
4. L. C. Jatoba, U. v. Großmann, C.
Kunze, J. Ottenbacher, and W. Stork,
"Context-aware mobile health
monitoring: Evaluation of different
pattern recognition methods for
classification of physical activity," 2008
30th Annual International Conference
of the IEEE Engineering in Medicine
and Biology Society, pp. 5250-5253,
2008.
5. V. N. T. Sang, N. D. Thang, V. Van
Toi, N. D. Hoang, and T. Q. D. Khoa,
"Human Activity Recognition and
Monitoring Using Smartphones," 2015.
6. Yao R, Lin G, Shi Q, et al. Efficient
Dense Labelling of Human Activity
Sequences from Wearables using Fully
Convolutional Networks[J]. Pattern
Recognition, 2017.
7. A. Bevilacqua, K. MacDonald, A.
Rangarej, V. Widjaya, B. Caulfield, and
T. Kechadi, “Human Activity
Recognition with Convolutional Neural
Networks,” Lecture Notes in Computer
Science, 2019.
8. O. Banos, M. A. Toth, M. Damas, H.
Pomares, and I. Rojas, "Dealing with
the effects of sensor displacement in
wearable activity recognition," Sensors
(Basel), 2014.
9. T. D. Le and C. V. Nguyen, "Human
activity recognition by smartphone,"
2015 2nd National Foundation for
Science and Technology Development
Conference on Information and
Computer Science (NICS), 2015.
10. Y. Hsu, S. Lin, P. Chou, H. Lai, H.
Chang, and S. Yang, "Application of
Computer Science and Engineering Research Journal

nonparametric weighted feature

extraction for an inertial-signal-based
human activity recognition system," in
2017 International Conference on
Applied System Innovation (ICASI),
2017.
11. P. Nurmi, Przybilski, M., Lindén, G.,
Floréen, P., "A Framework for
Distributed Activity Recognition in
Ubiquitous Systems" in: IC-AI.
Citeseer, 2005.
12. Andrey D. Ignatov and Vadim V.
Strijov, “Human activity recognition
using quasiperiodic time series collected
from a single tri-axial accelerometer”,
Multimedia Tools and Applications”,
2015.
13. B. Yuan and J. Herbert, “Context-
aware Hybrid Reasoning Framework
for Pervasive Healthcare,”Pervasive
Ubiquitous Computing., vol. 18, no. 4,
pp. 865–881, 2013.
14. G. Guo, H. Wang, D. Bell, Y. Bi, and
K. Greer, "KNN Model-Based
Approach in Classification," Berlin,
Heidelberg, 2003.
15. L.-Y. Hu, M.-W. Huang, S.-W. Ke, and
C.-F. Tsai, "The distance function effect
on k-nearest neighbor classification for
medical datasets," SpringerPlus, 2016.
16. J. Han and M. Kamber, "Data mining:
concepts and techniques morgan
kaufmann," vol. 54, 2006.
17. Weka 3: Data Mining Software in Java,
Machine Learning Group at the
University of Waikato, Official Web:
http://www.cs.waikato.ac.nz/ml/Weka/i
ndex.html/, accessed on 10th October
2019