ISSN 2277-2685

IJESR/Sep. 2023/ Vol-13/Issue-3/1-15

Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

Clustering-classification method for human activity recognition using smart

phone dataset
Dr T. BENARJI, 2 P. RENUKA

1
Professor, Department of CSE, Indur Institute of Engineering and Technology, Siddipet,
Telangana, Hyderabad. Email: [email protected]

2
Associate Professor, Department of ECE, Indur Institute of Engineering and Technology, Siddipet,
Telangana, Hyderabad. Email: [email protected]

Abstract: The study of human activities is known as Human Activity Recognition (HAR). A few
examples of the many real-world uses of HAR include healthcare systems, rehabilitation, and
monitoring patients on a regular basis to determine how they are doing in terms of lowering the
risk associated with activities of daily life. Traditional HAR algorithms abound, but they have a
long way to go before they meet modern standards in areas like privacy and accuracy. Two main
types of HAR exist: those that rely on vision, such surveillance footage and images, and those that
rely on sensors, like smartphones, smart watches, and wearable gadgets. Concerns about privacy,
data storage capacity, expense, infrastructure, and accuracy arise when using external devices for
data collection in vision-based HAR. Data is collected by the sensor-based HAR using on-body
smart devices worn in the bell, ankle, wrist, chest, and abdomen. The sensors built into these
gadgets collect data that is both useful and sensitive to user privacy. This study proposes GSCV-
RF and GSCV-SVM methods based on Random Forest (RF) and Support Vector Machine
Classifier (SVMC), respectively, with the goal of addressing all facets of sensor-based human
activity recognition, including datasets, pre-processing techniques, optimization techniques,
classification models, and prediction accuracy. If we compare these approaches to more traditional
models, we see that they provide far more accurate predictions. The following terms are used
interchangeably: sensors, classification, machine learning, optimization, prediction, and human
activity recognition.
I. INTRODUCTION

One way to anticipate complex movements is with the help of Human Activity Recognition, which
uses video and sensor data gathered from security cameras, wearable devices, and the
accelerometer, magnetometer, and gyroscope sensors in smartphones, smartwatches, and smart
bands. Care for the elderly, assisted living, rehabilitation, sports activity tracking, and smart health
care are just a few of the many HAR uses. Machine learning models use sensor data to forecast a
person's physical actions, which aids in patient rehabilitation and enhances athletic performance.
While the gyroscope detects rotation, the accelerometer detects x, y, and z-axis orientation (Figure
1.a and 1.(b) respectively).

1
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

This study proposes a new optimization model that uses machine learning to forecast people's
physical activity levels. A literature overview on machine learning methods for human activity
identification was provided in section 2. Human activity recognition prediction model is detailed
in Section 3. We covered the experimental findings that were acquired using Python program in
section 4. Section 5 concludes with the results and planned efforts.

II. LITERATURE REVIEW

Information on data sources, data kinds, pre-processing, optimization methods, classification and
prediction models, and metrics for fitting the model are all covered in this area of the health care
literature, with a focus on the identification of human physical activities. Collecting data about
people's physical activity is possible with the use of wearable and body sensors. Machine learning
optimization methods like Adaptive Learning rate, Stochastic Gradient descent, Gradient descent,
conjugate gradient method, derivative free optimization, zeroth order optimization, meta learning,
and other population algorithms like Genetic Algorithm and Particle Swarm Optimization (PSO)
are commonly employed for parameter optimization in classification models.

2.1 Pre-processing techniques B. A.

Several classifiers, including Naïve Bayes, Random Forest (RF), Deep learning, and k-Nearest
Neighbor, use the quick feature dimensionality reduction approach that Mohammed Hashim and
R. Amutha[1] suggested for identifying human physical activities. Results from experiments with
the UCI machine learning repository's human activity dataset demonstrate a random forest
classifier's high classification accuracy of 98.72%.

2.2 Classification models

Decision trees, neural networks, Bayesian classification, and lazy learners are some of the
classification models used in machine learning. With the use of accelerometer sensors included in
smartwatches, Min-Cheol Kwon and Sunwoong Choi[3] demonstrated a 95% success rate in
classifying 11 different human physical activities. In order to enhance the classification
performance for smartphone-based human activity identification, Mekruksavanich S. and
Jitpattanakul A.[4] suggested a CNN-LSTM hybrid model. When fine-tuning hyperparameters,
the Bayesian optimization method is used. Classification accuracy in the CNN-LSTM hybrid

2
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

model may reach 2.24%. For their presentation of the Ensemble Learning Algorithm with
Smartphone Sensor Data using fully connected DNN, Tan T-H, Wu J-Y, Liu S-H, and Gochoo M.
[5] demonstrated a prediction accuracy of 96.7% as determined by the f1 measure. To identify
qualitative human activity, Ysenllari, E., Ottenbacher, J., and McLennan[6] presented a 2D-
convolutional neural network (2D-CNN) and achieved an accuracy ranging from 96.57% to
99.28%. From 0.96 to 0.99, they were able to derive Cohen's k-value. The CNN-GRU model,
suggested by Dua, N. et.al. [7], obtained 96.20 percent accuracy on the UCI-HAR dataset, 97.21%
on the WISDM dataset, and 95.27% on the PAMAP2 dataset, respectively. For the UCI-HAR
dataset, Mutegeki R. and Han DS.[8] obtained an accuracy of 92.13 using a CNNLSTM method
for human activity prediction.In order to identify human activities, Prasad et al.[9] suggested a
convolutional neural network (CNN) method. They used the WISDM dataset, which includes data
from the accelerometers in mobile phones, and achieved an accuracy rate of 89.67% in their
predictions. An optimization strategy for crop suitability prediction was proposed by
Vasanthanageswari, Prabhu P. [11]. Using publicly available datasets, the model is shown to be
inferior to more traditional models. One well-designed k-nearest neighbour (KNN) method for
human activity identification was given by Mohsen, S et.al. [11]. A higher level of classification
accuracy is achieved by fine-tuning the parameters of this method. Experimental analysis is
performed on this method using real-world datasets. Classifiers were found to attain a
classification accuracy of 90.37 percent when tested using conventional performance criteria.
Using a cardiac dataset, Sivakami.M. and Prabhu P.[12] used machine learning methods to
enhance the precision of testing and classification. When tested against more traditional models,
the model proves to be superior. Using a frequent item set and fuzzy logic, Prabhu P et.al.[13–14]
suggested business intelligence models for an e-commerce recommender system. If you own an
online store, this model can help you choose what products to sell. When it comes to prediction
accuracy, sparsity, and scalability, traditional classifiers fall short. The increased reliance on
sensors and other devices in these models raises concerns about patient privacy and potential
adverse consequences. As a result, better health care approaches and parameter-optimized accurate
prediction should be made available via mobile devices.

There are approaches that can solve the HAR problem, such as those based on decision trees,
Bayesian statistics, and neural networks. To make reliable predictions about physical activities,
this paper proposes to employ optimization approaches before adjusting categorization. According
to this research, prediction models still have room for improvement in order to reach high accuracy
and go beyond problems with sensitivity and scalability. The health care applications field is
particularly in need of improved categorization.

3
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

III. METHODOLOGY

Optimal classification model for human activity recognition is described in this section. Figure 2
depicts the design of the created model.

Preprocessing, modeling, and assessment are the three main parts of this model. As an initial step,
the suggested model loads the accelerometer and gyroscope data used for human activity
identification. Label encoding and dimension reduction are the next steps in getting this dataset
ready for modeling. Train data and test data are subsets of the input dataset. Applying an
optimization strategy to determine the optimal values for the prediction model's parameters is the
next stage in the suggested model. The rest of this section goes into more depth on the suggested
prediction model. 3.1 preparatory Inconsistencies, missing values, and irrelevant data are checked
for during the pre-processing of the dataset. When it comes to lowering the amount of
characteristics, the dimensionality reduction approach is crucial for eliminating the least

4
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

significant ones. The input for the model is a list of features extracted from the dataset using the
highly accurate and interpretable Random Forest feature selection approach. Separate "train" and
"test" datasets are used for model testing. Part 3.2: Model Training. Training the model is the next
stage of this planned project. A classifier's prediction accuracy may be greatly improved by
hyperparameter tuning, which involves identifying the optimal values for these parameters. To
fine-tune the hyperparameters, you may use either the GridSearchCV or RandomizedSearchCV
methods. While RandomizedSearchCV defines combinations at random, GridSearchCV allows
the user to specify all potential values. Because it uses a random selection process, the
RandomizedSearchCV model reduces the amount of time required to find a value when there are
many possible combinations of parameters.

3.2.1 GridSearchCV optimization

To determine the optimal classifier parameters from the provided list, this study used the
GridsearchCV method, a cross validation tool. It tests several values for each hyperparameter and
their combinations to see which one works best, then picks the best one. Hyperparameters for
random forest classifiers include forest size, maximum tree levels, feature count, and impurity
criterion (Gini and entropy) for decision tree generation. The optimal parameters have been
determined to be criterion='entropy,' max_depth=11, n_estimators=200, and random_state=42.
Regularization parameter, gamma, and kernel are the hyperparameters that support VGG19
classifiers. We found that the SVM approach works best with the following parameters: C=100,
gamma=0.001, kernel='rbf'.

3.3 Activity Prediction and Evaluation

Predicting the appropriate class for the new input data (i.e., test data) is the final step in this model.
After receiving the test data, the model compares it to the train data and produces a result. Applying
the optimal parameters to the classifier model allows it to make predictions on test data once they
have been determined via an optimization approach.

5
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

3.3.1 Evaluation Measures

We put the prediction model through its paces in a Python environment using real-world datasets
and mathematical metrics to ensure its quality. A variety of metrics are used to assess the
performance of the suggested prediction model. These include recall (sensitivity), precision, F-
measure, macro average, and weighted average derived from the total number of TP, FP, and FN
that were detected. A measure of accuracy is the proportion of results that are really positive out
of all possible outcomes. The F-measure finds the harmonic mean of the recall and accuracy
values, therefore balancing them.

6
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

IV. RESULTS AND DISCUSSION

Here we describe the experimental setup and the outcomes of the prediction models that were
constructed for human activity recognition.

4.1 Experimental setup

4.1.1 Datasets description

The HAR datasets may be sourced from publicly accessible online databases as well as real-world
sources. collection of machine learning models housed at UCI, Nottingham Trent University,
WISDM[15,16], Groupware, Wrist Sensor Dataset, Chest Sensor Dataset, and Human Activity
Recognition Using Smartphones Data Set. The following datasets are often made accessible to the
public: HAR Single Chest-Mounted Accelerometer, SisFall, Cornell Activity Dataset, MobiSense,
MobiAct, and PAMAP2. The purpose of this study is to evaluate the model's efficacy in human
activity identification using a dataset extracted from smartphones and stored in the machine
learning repository at the University of California, Irvine [14]. Thirty participants participated in
six different physical activities, including 0-laying, 1-sitting, 2-standing, 3-walking, 4-walk
downstairs, and 5-walk upstairs, and their raw data is included in this dataset. It contains data from
the temporal and frequency domains and is comprised of 102,299 samples with 561 properties.
Subject identifiers, target classes, time domain variables, triaxial angular velocities, and
accelerations are all part of each sample. Using a dataset consisting of activities recognized by
cellphones, figure 3(a) displays the distribution of these activities from the UCI Machine Learning
library.

Figure 3.(b) shows the sample of Accelerometer and gyroscope signals of six activities.

7
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

The Public UCI datasets [15] collected from website can also be used to test the performance of
this prediction model.

4.1.2 Feature selection Results

4.1.2.1 Random forest feature selection

The selected features using Random Forest feature selection method are [76 72 54 475 509 38 74
10 503 51 559 508 348 288 202 41 215 57 42 75 394 84 182 354 100 87 53 390 382]. Figure 3.(c)
shows the features selected using random forest feature selection method.

4.1.2.2 Best fit feature selection

Selected attributes Using Best fit method: [11,24,38,39,42,43,45,51,52,53,54,55,58,67,69,71,7

3,76,79,80,81,105,106,109,122,159,160,161,181,20
4,232,298,304,305,309,312,368,370,371,383,388,4
13,450,451,452,453,492,506,520,539,558,559,560, 561 : 54].

8
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

4.1.2.3 Greedy stepwise feature selection

Selected attributes using greedy stepwise method:

[11,24,38,39,42,43,45,51,52,53,54,55,58,67,69,71,7
3,76,79,80,81,105,106,109,122,159,160,161,181,20
4,232,298,304,305,309,312,368,370,371,383,388,4
13,450,451,452,453,492,506,520,539,558,559,560, 561 : 54]

4.1.2 Software implementation description

Python 3.10 for Windows, an Intel(R) Core (TM) i7-8565U CPU running at 1.80 GHz with a turbo
boost to 1.99 GHz, and 8 GB of RAM were used to execute this study. Created with the help of
Scilab, Numpy, Pandas, Tensorflow, sklearn, matplotlib, and seaborn, the optimization-based
prediction model for Human Activity Recognition is well named.

4.2 Results and discussion

4.2.1 Performance comparison

To test how well a prediction model works, one often looks at its accuracy, recall, F measure,
sensitivity, specificity, ROC curve, MAE, and confusion matrix. The methods for implementing a
HAR system to identify potential health hazards using two kinds of sensor data were laid forth by
Abdulhamit Subasi et al., [2]. The experimental findings of analyzing the performance of several
machine learning approaches in recognizing critical conditions are shown and analyzed.

Table 1 shows the accuracy comparison of various conventional classifiers calculated using HAR
dataset.

9
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

Table 2 shows the accuracy comparison of various improved conventional classifiers calculated
using HAR dataset. The Logistic Regression (LR) Classifier gives high f-measure accuracy of
98.12 when compared with other conventional models.

The 2D-CNN[6] based method provides high accuracy of classification up to 99.28%. The figure
4 shows the ROC curve shows false positive rate vs. True positive rate value of 0.998 of sitting
activity classified using Random Forest Method.

The confusion matrix is a performance evaluation metrics using for visualizing the performance
of classifiers.

Figure 5 shows the confusion matrix of proposed GridSearchCV + SVM classifier using UCI
machine learning HAR smartphone dataset.

10
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

Table 3 shows the activity wise precision, recall and f1 score accuracy of SVM classifier calculated
using HAR smartphone dataset. The Walk-downstairs activity gets high f1 score of 0.9950. The
overall accuracy obtained for all activities is 0.9754.

Table.4 shows the activity wise precision, recall and f1 measure accuracy of proposed
GridSearchCV+ SVM classifier calculated using HAR dataset. The Walk-downstairs activity gets
high f1 score of 0.9981. The overall accuracy obtained is 0.9764.

11
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

Figure 6 shows the accuracy of activities calculated using GridSearchCV+SVM classifier. The
Walk-downstairs activity gets high f1 score of 0.9981 in the bar chart. The overall accuracy
obtained is 0.9764.

Table.5 shows the accuracy comparison of various conventional classifiers with proposed
GridSearchCV classifiers calculated using HAR dataset.

Figure.7 shows the accuracy comparison of various conventional classifiers with proposed
GridSearchCV classifiers calculated using HAR dataset.

Using Random Forest and SVM classifier, we assess the performance of the proposed enhanced
grid search optimization model. With grid search optimization, the accuracy of the Random Forest
model went up from 97.67% to 97.72%, and that of the SVM model went up from 97.67% to

12
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

98.70%. GridSearchCV (GSCV) improves the accuracy of SVM models. It is possible to evaluate
this model's efficacy by testing it with real-world data acquired from hospitals, although there are
still some challenges and unanswered research questions. This work is in the subject of health care
systems, thus there may be a need for greater improvement in prediction accuracy. Finding the
best settings for several classifiers is the goal of this proposed study, which employs three feature
selection philosophies and optimization techniques including grid search and random forest.

CONCLUSION:

Care for the elderly and other medical conditions presents significant challenges to the research of
human activity recognition. Predicting human physical activity is the focus of this research, which
proposes a machine learning approach based on grid search optimization. We also identify and
examine the accuracy of different classifiers. The paper's meat and potatoes are the recommended
model for predicting people's physical activity from data gathered by cellphones' sensors. When it
comes to predicting human actions, the experimental findings reveal that the suggested model
works better. For the purpose of making predictions, the enhanced classifier employs optimum
parameters. This approach offers superior accuracy when compared with other traditional
optimization methods. Additional optimization methods may be used to test this prediction model.
It is possible to suggest models using deep neural networks, which are based on neural networks.
The performance benchmark may be tested with datasets retrieved from various public repositories
or actual data from the real world.

REFERENCES

[1] Mohammed Hashim, B.A., Amutha, R. Human activity recognition based on smartphone using
fast feature dimensionality reduction technique. J Ambient Intell Human Comput 12, 2365–2374
(2021). https://doi.org/10.1007/s12652-020- 02351-x.

[2] Abdulhamit Subasi, Kholoud Khateeb, Tayeb Brahimi and Akila Sarirete, Human activity
recognition using machine learning methods in a smart healthcare environment, Innovation in
Health Informatics (2020) Elsevier Inc. pp123- 144.

[3] Min-Cheol Kwon and Sunwoong Choi, Recognition of Daily Human Activity Using an
Artificial Neural Network and Smartwatch, Hindawi Wireless Communications and Mobile
Computing Volume 2018, Article ID 2618045, https://doi.org/10.1155/2018/2618045.

[4] Mekruksavanich S, Jitpattanakul A., Biometric User Identification Based on Human Activity
Recognition Using Wearable Sensors: An Experiment Using Deep Learning Models. Electronics.
2021; 10(3):308. https://doi.org/10.3390/electronics10030308.

13
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

[5] Tan T-H, Wu J-Y, Liu S-H, Gochoo M. Human Activity Recognition Using an Ensemble
Learning Algorithm with Smartphone Sensor Data. Electronics. 2022; 11(3):322.
https://doi.org/10.3390/electronics1103032.

[6] ysenllari, E., Ottenbacher, J. & McLennan, D. Validation of human activity recognition using
a convolutional neural network on accelerometer and gyroscope data. Ger J Exerc Sport Res 52,
248–252 (2022). https://doi.org/10.1007/s12662- 022-00817-y.

[7] Dua, N., Singh, S.N. & Semwal, V.B. Multi-input CNN-GRU based human activity recognition
using wearable sensors. Computing 103, 1461– 1478 (2021). https://doi.org/10.1007/s00607-021-
00928-8.

[8] R. Mutegeki and D. S. Han, "A CNN-LSTM Approach to Human Activity Recognition," 2020
International Conference on Artificial Intelligence in Information and Communication (ICAIIC),
2020, pp. 362-366, doi: 10.1109/ICAIIC48513.2020.9065078.

[9] Prasad, A.; Tyagi, A.K.;Althobaiti, M.M.; Almulihi, A.;Mansour, R.F.; Mahmoud, A.M.
Human Activity Recognition Using Cell PhoneBased Accelerometer and Convolutional Neural
Network. Appl.Sci. 2021, 11, 12099. https://doi.org/10.3390/app112412099.

[10] Vasanthanageswari, Prabhu P, Improving Svm Classifier Model Using Tree Structured Parzen
Estimator Optimization for Crop Prediction, Journal of Theoretical and Applied Information
Technology, 2022, 100(22), pp. 6808–6818

[11] Mohsen, S., Elkaseer, A., Scholz, S.G. (2022). Human Activity Recognition Using K-Nearest
Neighbor Machine Learning Algorithm. In: Scholz, S.G., Howlett, R.J., Setchi, R. (eds)
Sustainable Design and Manufacturing. KESSDM 2021. Smart Innovation, Systems and
Technologies, vol 262. Springer, Singapore. https://doi.org/10.1007/978-981-16-6128-0_29.

[12] M. Sivakami, P. Prabhu, Classification of Algorithms Supported Factual Knowledge

Recovery from Cardiac Data Set, International Journal of Current Research and Review, vol.13.
issue 6. March 2021 pp161-166.

[13] Prabhu P., Anbazhagan N. (2013) FI-FCM Algorithm for Business Intelligence. In: Prasath
R., Kathirvalavakumar T. (eds) Mining Intelligence and Knowledge Exploration. Lecture Notes in
Computer Science, vol 8284. Springer, Cham, pp 518-528.

[14] Paulraj P., Neelamegam A. (2014) Improving Business Intelligence Based on Frequent
Itemsets Using k-Means Clustering Algorithm. In: Meghanathan N., Nagamalai D., Rajasekaran
S. (eds) Networks and Communications (NetCom2013). Lecture Notes in Electrical Engineering,
vol 284. Springer, Cham, DOI : 10.1007/978-3-319-03692-2_19, pp 243-254 .

14
*Corresponding Author www.ijesr.org
 ISSN 2277-2685
IJESR/Sep. 2023/ Vol-13/Issue-3/1-15
Dr T. BENARJI et. al., / International Journal of Engineering & Science Research

[15] UCI machine learning repository: https://archive.ics.uci.edu/ml/index.php. [16] UCI machine

learning repository WISDM dataset: https://archive.ics.uci.edu/ml/datasets/WISDM+
Smartphone+and+Smartwatch+Activity+and+Bi ometrics+Dataset

15
*Corresponding Author www.ijesr.org