IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 14, No. 2, April 2025, pp. 1461~1470

ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i2.pp1461-1470  1461

Enhancing fall detection and classification using

Jarratt‐butterfly optimization algorithm with deep learning

Kakirala Durga Bhavani, Melkias Ferni Ukrit

Department of Computational Intelligence, SRM Institute of Science and Technology, Kattankulathur, India

Article Info ABSTRACT

Article history: Falls pose significant risk to the health and safety of individuals, specifically
for vulnerable populations as the elderly and those with specific medical
Received Apr 6, 2024 conditions. The repercussions of falls can be severe, leading to injuries, loss
Revised Nov 3, 2024 of independence, and increased healthcare costs. Consequently, the
Accepted Nov 14, 2024 development of effective fall detection systems is crucial for providing
timely assistance and enhancing the overall well-being of affected
individuals. Recent advancements in deep learning (DL) have opened new
Keywords: avenues for automating fall detection through the analysis of sensor data and
video footage. DL algorithms are especially well-suited for this task because
Computer vision they can automatically learn complex features and patterns from raw data,
Deep learning eliminating the need for extensive manual feature engineering. This article
Fall detection introduces a novel approach to fall detection and classification, termed the
Machine learning fall detection and classification using Jarratt‐butterfly optimization algorithm
Metaheuristics with deep learning (FDC-JBOADL) algorithm. The FDC-JBOADL
technique employs a median filtering (MF) method to mitigate noise and
utilizes the EfficientNet model for robust feature extraction, capturing both
motion patterns and appearance characteristics of individuals. Furthermore,
the classification of fall events is achieved through a long short-term
memory (LSTM) classifier, with hyperparameter optimization facilitated by
Jarratt‐butterfly optimization algorithm (JBOA). Through a comprehensive
series of experiments, the efficacy of FDC-JBOADL technique is validated,
demonstrating superior performance compared to existing methodologies in
the domain of fall detection.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Melkias Ferni Ukrit
Department of Computational Intelligence, SRM Institute of Science and Technology
Kattankulathur-603203, Chennai, India
Email: [email protected]

1. INTRODUCTION
The increasing aging of the population, particularly in developing nations, poses a significant
challenge to the sustainability of medical treatments [1]. The proportion of people of working age (15 to 64)
among the total population in European nations is projected to decline from 65.16% in 2016 to 56.15% by
2070, while life expectancy at birth is anticipated to rise by an additional 7 years for both women and men
over the same timeframe [2]. In this instance, falls have been a significant cause of loss of autonomy and
accidents among the elderly. The World Health Organization (WHO) research indicates that the annual fall
rate for those aged 64 to 70 is around 28 to 35% and 32 to 40%, respectively [3]. Despite the lack of
significant injuries, 47% of individuals who fall are unable to rise alone post-fall [4]. Moreover, prolonged
periods of lying on the ground prior to falling are significantly associated with comorbidities and pressure
sores, which increase the likelihood of mortality within six months to 50% [5]. From this viewpoint, prompt

Journal homepage: http://ijai.iaescore.com

1462  ISSN: 2252-8938

response after a fall is crucial for mitigating the physical and psychological effects (fear of falling (FoF)
syndrome), which undermine the well-being of elderly individuals and their confidence in living
independently and self-sufficiently [6].
Fall detection system (FDS), is proficient in discriminative falls from activities of daily living
(ADL) thereby an alarm to remote monitoring point has been automatically produced immediately, and the
user or patient under observation is suspected to have fallen [7]. The traditional machine learning (ML)
approaches endure the shortage of labelled training databases and greatly depend on the extracted features by
humans which creates it hard to utilize on massive platforms [8]. Deep learning (DL) has a new paradigm in
the ML domain primarily determined by utilizing artificial neural networks (ANNs) and greater performance
than the other standard ML approaches. The DL includes different networks namely recurrent neural network
(RNN), restricted Boltzmann machines (RBMs), convolutional neural network (CNN), deep belief network
(DBN), which have various features and abilities [9]. These networks can perform the learning process in
unsupervised, semi-supervised, or supervised behaviors. Also, it advantages from the hierarchical layers
targeted for finding appropriate higher-level features from the raw input data in place of utilizing manual
features [10].
This article presents a new fall detection and classification using Jarratt‐butterfly optimizer
algorithm with deep learning (FDC-JBOADL) technique. The FDC-JBOADL technique applies median
filtering (MF) approach to remove the noise. In addition, the FDC-JBOADL technique makes use of
EfficientNet model for the extraction of relevant features from both motion patterns and appearance
characteristics of individuals. Moreover, the classification of fall events occurs using long short-term
memory (LSTM) network. Finally, the Jarratt‐butterfly optimization algorithm (JBOA) can be employed for
the optima hyperparameter choice of the LSTM model. A wide range of experiments was performed to
validate the superior recognition results of the FDC-JBOADL technique.

2. RELATED WORKS
In recent years, the integration of innovative DL techniques and advanced sensor technologies has
significantly enhanced the capabilities of classification and monitoring systems in various domains. For
instance, Li et al. [11] introduced a novel DL framework that integrates temporal convolution networks
(TCN) with gated recurrent units (GRU), aimed at extracting higher-level features for improved classification
accuracy. Their research involved a comparative analysis against 2 extensively utilized ML classifiers and six
existing DL approaches, leveraging 2 well-established open-source datasets collected from inertial sensors.
Concurrently, Raeve et al. [12] proposed an innovative fault detection and alert system tailored for care
centers, which utilizes bluetooth low energy (BLE) for wireless communication. This study also emphasizes
the development of a real-time data filtering method to enhance the accuracy of measurements. Additionally,
the exploration of millimeter-wave (mmWave) radar technology for unobtrusive human fall detection has
been highlighted in [13]. This research involved collecting data from healthy young volunteers, with radar
systems strategically positioned either on the side wall or overhead within a designated area. To address the
underlying fault detection challenges, a CNN-based DL approach was also developed. Collectively, these
studies underscore the transformative potential of integrating advanced DL methodologies with cutting-edge
sensor technologies to improve monitoring and classification tasks across various applications.
This study also develops a CNN based approach to address underlying FD challenges. Building on
these advancements, Yao et al. [14] proposed an efficient FD technique utilizing a joint motion map
constructed from two parallel CNNs, innovatively employing the red, green, and blue (RGB) channels of
pixels to capture relative motion data. Their method predicts the limits of stability and accurately identifies
the initial and final key frames preceding a potential fall. Furthermore, the hybrid deep CNN model known as
squeeze and excitation (SE)-Deep ConvNet, designed by Mekruksavanich et al. [15], enhances fall detection
capabilities with the implementation of squeeze and excitation techniques. Collectively, these studies
underscore the transformative potential of DL and advanced sensor technologies in improving FDS,
ultimately contributing to enhanced safety and monitoring in various environments.
In recent years, the advancement of fault diagnosis techniques has gained significant attention in
various fields, mostly in the context of high voltage direct current (HVDC) models and fall detection for the
elderly. Jawad and Abid [16] proposed a novel approach for HVDC fault diagnosis that integrates a
probabilistic generative algorithm based on feature selection (FS) and wavelet transform methods. Their
methodology involves the extraction of noise from both non-fault and fault signals, followed by the
application of ant colony optimization (ACO) to eliminate irrelevant attributes within the feature vectors. The
refined features are subsequently utilized for training an ANN to effectively distinguish between non-fault
and fault conditions. Concurrently, Lee et al. [17] introduced a dual verification strategy for fall detection in
older adults, employing a combination of RGB cameras and inertial measurement unit-location (IMU-L)

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1461-1470

Int J Artif Intell ISSN: 2252-8938  1463

sensors. This approach is particularly innovative as it leverages wearable technology to monitor falls, with
the IMU-L sensor providing real-time detection capabilities. To enhance the accuracy of fall classification, a
DL method utilizing RNN is employed, marking a significant step forward in the consistency of FDS. This
research highlights the intersection of advanced signal processing techniques and ML methodologies in
addressing critical safety concerns in both electrical engineering and geriatric care.

3. THE PROPOSED MODEL

Falls represent a major health risk, particularly among the elderly population, leading to severe injuries
and increased mortality rates. As such, the progress of automated systems for fall detection and classification
has garnered considerable attention over recent years. This research introduces the FDC-JBOADL technique, a
novel approach designed to develop the accuracy and effectiveness of fall event identification and classification
through the integration of DL models. The FDC-JBOADL technique employs a multifaceted methodology that
includes noise removal using a MF, feature extraction leveraging the EfficientNet architecture, fall detection
using LSTM, and hyperparameter optimization through the JBOA algorithm. The workflow of the
FDC-JBOADL method is exemplified in Figure 1, demonstrating the systematic process involved in
recognizing and categorizing fall events. This innovative technique aims to assist the progress of automated
FDS, advanced enhancement the safety and well-being of vulnerable populations.

Figure 1. Workflow of FDC-JBOADL approach

3.1. Image pre-processing

To pre-process the input images, the MF technique is exploited in this study. It is a nonlinear digital
image processing method used for preserving edges and reducing noise in images. It is very efficient at
eliminating salt-and-pepper noises, where random white and black pixels appear throughout the image.
The MF method includes substituting the value of all the pixels with median value of its neighboring pixels
within the given kernel or window.

3.2. Feature extraction using EfficientNet model

For effectual identification of the feature vectors, the EfficientNet method can be employed.
EfficientNet is a family of deep neural network (DNN) architecture that has been intended to accomplish
remarkable performance while being computationally effective [18]. The basic concept behind EfficientNet
is to simultaneously balance the model's depth, width, and resolution to attain best outcomes with less
computational parameters and. Traditional model scaling techniques focus mainly on increasing the
dimension (for example: width or depth), resulting in suboptimum performance. EfficientNet makes use of a
compound coefficient to uniformly scale the 3D, which is derived from a series of experiments. The
compound scaling coefficient is used for scaling the resolution, depth (number of layers), and width (number
of channels) of the networks. The EfficientNet model has accomplished outstanding performances on
computer vision task while being more effective than other DenseNet and ResNet architectures. They are
Enhancing fall detection and classification using jarratt‐butterfly optimization … (Kakirala Durga Bhavani)
1464  ISSN: 2252-8938

widely applied for the tasks including object detection, image classification, and segmentation. There exist
numerous variants of EfficientNet, namely EfficientNet-B0-B7, with various levels of performance and
complexity. B0 is the simplest and smallest version, whereas B7 is the most complex and largest one. Based
on the computational resource available and the task requirement, users can select the suitable variant.

3.3. Fall detection using long short-term memory

In this work, the LSTM framework can be utilized for the identification and classification of fall
events. The LSTM network is an improved method of an RNN [19], [20]. Different CNNs namely multilayer
perceptron (MLP) and RNNs could not be restricted to a unidirectional flow of data. It is loop done many
layers and temporarily memorizes data which is employed later. In the meantime, an easy RNN is vulnerable
to gradient disappearing problems, and the GRU and LSTM are established for solving the problem. The
LSTM learns long‐term dependencies, enduring appropriate to classify sequential data like credit card
information. LSTM network comprises memory cell 𝑐𝑡 , with input gate 𝑖𝑡 , output gate 𝑜𝑡 , and forget gate 𝑓𝑡 .
The 3 gates control that the data has been managed and employed. Figure 2 illustrates the structure of LSTM.
The subsequent mathematical equations define the data flow in the LSTM layers as in (1) to (6).

𝑖𝑡 = 𝜎(𝑉𝑖 𝑥𝑡 + 𝑊𝑖 ℎ𝑡−1 + 𝑏𝑖 ) (1)

𝑓𝑡 = 𝜎(𝑉𝑓 𝑥𝑡 + 𝑊𝑓 ℎ(𝑡−1) + 𝑏𝑓 ) (2)

𝑐̃𝑡 = 𝑡𝑎𝑛ℎ(𝑉𝑐 𝑥𝑡 + 𝑊𝑐 ℎ(𝑡−1) + 𝑏𝑐 ) (3)

𝑐𝑡 = 𝑓𝑡 ⊗ 𝑐(𝑡−1) + 𝑖𝑡 ⊗ 𝑐̃𝑡 (4)

𝑜𝑡 = 𝜎(𝑉𝑜 𝑥𝑡 + 𝑊𝑜 ℎ(𝑡−1) + 𝑏𝑜 ) (5)

ℎ𝑡 = 𝑜𝑡 ⊗ 𝑡𝑎𝑛ℎ(𝑐𝑡 ) (6)

Whereas 𝑉∗ , 𝑊∗ , and 𝑏∗ signifies the learnable parameter, ℎ∗ denotes the hidden layer, which, ∗ is employed
in place of 𝑓, 𝑖, 𝑜, or 𝑐 to signify the provided memory cell and gates. In the meantime, ⊗ denotes the
element‐by-element product; 𝑡𝑎𝑛ℎ and 𝜎 denote the tanh activation and sigmoid functions.

Figure 2. LSTM architecture

3.4. Parameter tuning using Jarratt‐butterfly optimization algorithm

In recent years, the optimization of LSTM has garnered significant attention because of o their
efficacy in handling sequential data. One of the promising strategies for enhancing the performance of LSTM
networks is the integration of optimization algorithms, such as the butterfly optimization algorithm (BOA).
While BOA has revealed that a powerful tool for various applications, it is not without its challenges,
particularly with issues of divergence and the propensity to become trapped in local optima during the
resolution of nonlinear systems of equations (NSE). To mitigate these limitations, the Jarratt's model is
incorporated into the BOA framework, resulting in the development of the JBOA. This hybrid approach
leverages the strengths of both algorithms, significantly improving the accuracy and convergence rate in
solving NSE. Specifically, Jarratt's technique is employed iteratively within the BOA process, enhancing the
candidate butterfly positions identified by BOA and ensuring that the most optimal solutions are selected
based on fitness criteria. The integration of Jarratt's method not only accelerates the convergence process but

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1461-1470

Int J Artif Intell ISSN: 2252-8938  1465

also enhances the overall effectiveness of JBOA in resolving complex optimization problems. The following
sections will detail the algorithmic framework of JBOA, illustrated through its pseudocode in Algorithm 1,
and discuss its implications for parameter tuning in LSTM networks.

Algorithm 1. Pseudocode of JBOA

Objective function 𝑓(𝑥), 𝑥 = (𝑥1 , 𝑥2 , … , 𝑥𝑑𝑖𝑚 ), 𝑑𝑖𝑚 = No. of dimensions
Create population initialization of n Butterflies 𝑥𝑖 = (1,2, … , 𝑛)
Stimulus Intensity 𝐼𝑖 at 𝑥𝑖 is determined 𝑓(𝑥𝑖 )
Describe switching probability 𝑝, sensor modality 𝑐 and power exponent 𝑎
While ending condition is not met do
For all the butterflies 𝑏𝑓 in the population do
Evaluate fragrance for 𝑏𝑓
End for
Find the better 𝑏𝑓
for every butterfly 𝑏𝑓 in the population do
Generate a modem 𝑟 from [0,1]
If 𝑟 < 𝑝 then
Move toward solution or butterfly
Else
Move randomly
End if
End for
Upgrade the value of a
end while
Compute Jarratt’s location 𝑥𝑛+1 using 𝑏𝑓
Evaluate the fitness of 𝑥𝑛+1 and 𝑥𝑏𝑓
If Fitness (𝑥𝑛+1 ) < Fitness(𝑥𝑏𝑓 ) then
𝑥𝑏𝑓 = 𝑥𝑛+1
End if
Output the better solution found (𝑋𝑏𝑓 )

JBOA uses the modification given in the red box at the iteration end. Based on fitness value, this
comparison was made between the BOA butterfly’s location (𝑋𝑏𝑓 ) and Jarratt’s technique’s location (𝑋𝑛+1 ).
At last, the best position that evaluates better fitness is chosen as an optimum solution.
The fitness choice is an important element in the JBOA classifier. Encoder performance can be
executed for measuring the goodness of candidate outcomes. The accuracy value is the basic premise
engaged for developing a fitness function (FF).

𝐹𝑖𝑡𝑛𝑒𝑠𝑠 = max (𝑃) (7)

𝑇𝑃
𝑃 = 𝑇𝑃+𝐹𝑃 (8)

Where 𝑇𝑃 and 𝐹𝑃 represent the true positive ratio and the false positive ratio.

4. RESULTS AND DISCUSSION

This section evaluates the performance of the FDC-JBOADL methodology utilizing the KFall and
CAUCAFall datasets. The KFall dataset [21], [22] includes various classes such as forward fall while
attempting to sit down (20), backward fall while attempting to sit down (21), lateral fall while trying to sit
down (22), forward fall while trying to get up (23), lateral fall while getting up (24), forward fall while sitting
due to fainting (25), lateral fall while sitting due to fainting (26), Backward fall due to fainting (27), vertical
(forward) fall while walking due to fainting (28), fall while walking with hands used to soften the impact due
to fainting (29), forward fall while walking due to tripping (30), forward fall while jogging due to tripping
(31), forward fall while walking due to slipping (32), lateral fall while walking due to slipping (33), and
backward fall while walking due to slipping (34). Meanwhile, the CAUCAFall dataset [23] consists of
13,581 ADL labeled as "nofall" and 6,421 labeled as "fall". Sample images are illustrated in Figure 3.
Table 1 and Figure 4 demonstrates the overall classifier results of the FDC-JBOADL technique on
the KFall dataset. The outcomes indicate that the FDC-JBOADL technique reaches effectual outcomes on
both training set (TRS) and testing set (TSS). On the applied TRS, the FDC-JBOADL technique offers
𝑎𝑐𝑐𝑢𝑦 , 𝑝𝑟𝑒𝑐𝑛 , 𝑟𝑒𝑐𝑎𝑙 , 𝐹𝑠𝑐𝑜𝑟𝑒 , and Matthews correlation coefficient (MCC) of 99.39, 99.39, 99.41, 99.30, and
99.03% respectively. At the same time, on the applied TSS, the FDC-JBOADL method provides 𝑎𝑐𝑐𝑢𝑦 ,
𝑝𝑟𝑒𝑐𝑛 , 𝑟𝑒𝑐𝑎𝑙 , 𝐹𝑠𝑐𝑜𝑟𝑒 , and MCC of 99.16, 99.32, 99.38, 99.25, and 99.01% correspondingly.

Enhancing fall detection and classification using jarratt‐butterfly optimization … (Kakirala Durga Bhavani)
1466  ISSN: 2252-8938

Figure 3. Sample images

Table 1. Classifier outcome of FDC-JBOADL algorithm on KFall dataset

Metrics Training set (%) Testing set (%)
Accuracy 99.39 99.16
Precision 99.39 99.32
Recall 99.41 99.38
F-score 99.30 99.25
MCC 99.03 99.01

Figure 4. Classifier outcome of FDC-JBOADL approach on KFall database

The performance evaluation of ML methodologies is crucial for understanding their effectiveness in

various applications. This study focuses on the FDC-JBOADL method, specifically its application to the
KFall dataset, to analyze training and validation accuracies as well as loss metrics. As illustrated in Figure 5,
the training accuracy (TR_accu_y) and validation accuracy (VL_accu_y) exhibit a positive correlation with
the number of training epochs, indicating that increased epochs contribute to enhanced model efficacy on
both the training and testing datasets. Furthermore, Figure 6 presents the trends in training loss (TR_loss) and
validation loss (VR_loss) associated with the FDC-JBOADL approach. The results reveal a consistent
decrease in both TR_loss and VR_loss as the epochs progress, underscoring the model's capability to
minimize prediction discrepancies and improve classification precision. Collectively, these findings affirm
the FDC-JBOADL method's potential in effectively identifying patterns and relationships within data,
thereby establishing its significance in the realm of predictive analytics.

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1461-1470

Int J Artif Intell ISSN: 2252-8938  1467

Figure 5. 𝐴𝑐𝑐𝑢𝑦 curve of FDC-JBOADL algorithm on KFall database

Figure 6. Loss curve of FDC-JBOADL algorithm on KFall dataset

Table 2 and Figure 7 signifies the classification outcomes of the FDC-JBOADL method on the
CAUCAFall dataset. The outcomes specify that the FDC-JBOADL technique reaches effectual outcomes on
both TRS and TSS. On the applied TRS, the FDC-JBOADL technique offers 𝑎𝑐𝑐𝑢𝑦 , 𝑝𝑟𝑒𝑐𝑛 , 𝑟𝑒𝑐𝑎𝑙 , 𝐹𝑠𝑐𝑜𝑟𝑒 ,
and MCC of 98.81, 98.72, 98.54, 98.50, and 98.18% respectively. At the same time, on the applied TSS, the
FDC-JBOADL technique offers 𝑎𝑐𝑐𝑢𝑦 , 𝑝𝑟𝑒𝑐𝑛 , 𝑟𝑒𝑐𝑎𝑙 , 𝐹𝑠𝑐𝑜𝑟𝑒 , and MCC of 98.33, 98.27, 98.12, 98.20, and
98.04% respectively.
The evaluation of ML methodologies often hinges on their ability to accurately classify and predict
outcomes based on training and validation datasets. In this context, the FDC-JBOADL method has been
applied to the CAUCAFall dataset, revealing significant insights into its performance metrics. As illustrated
in Figure 8, both TR_accu_y and VL_accu_y exhibits a positive correlation with the number of training
epochs, suggesting that prolonged training enhances the model's efficacy on both the training (TR) and
testing (TS) datasets. This trend underscores the importance of epoch duration in optimizing the efficiency of
ML approaches. Moreover, Figure 9 presents the loss metrics associated with the FDC-JBOADL approach,
specifically the TR_loss and VR_loss. These metrics provide a quantitative measure of the discrepancy
among predicted outcomes and actual outcomes, with findings indicating a consistent decline in both
TR_loss and VR_loss as epochs progress. This reduction in loss values further corroborates the model's
increasing proficiency in identifying underlying patterns and relationships within the data. Collectively, these

Enhancing fall detection and classification using jarratt‐butterfly optimization … (Kakirala Durga Bhavani)
1468  ISSN: 2252-8938

outcomes emphasize the solution of the FDC-JBOADL method in achieving precise classifications, thereby
contributing to the broader discourse on the optimization of ML techniques in complex datasets.
The comparison study of the FDC-JBOADL technique in existing approaches take place in Table 3.
The results indicate that the FDC-JBOADL method achieves enriched performance over other models
[24], [25]. Based on 𝑎𝑐𝑐𝑢𝑦 , the FDC-JBOADL technique accomplishes higher 𝑎𝑐𝑐𝑢𝑦 of 99.39% but the
CNN classifier, LSTM algorithm, CNN-LSTM approach, and FDSNeXt models attain minimal 𝑎𝑐𝑐𝑢𝑦 values
of 85.69, 90.12, 84.04, and 91.87% respectively. In addition, based on 𝑝𝑟𝑒𝑐𝑛 , the FDC-JBOADL technique
accomplishes higher 𝑝𝑟𝑒𝑐𝑛 of 91.44% whereas the CNN classifier, LSTM algorithm, CNN-LSTM approach,
and FDSNeXt algorithm attain lower 𝑝𝑟𝑒𝑐𝑛 values of 91.28, 89.90, 91.07, and 99.39% respectively. Next to
that, based on 𝑟𝑒𝑐𝑎𝑙 , the FDC-JBOADL method accomplishes higher 𝑟𝑒𝑐𝑎𝑙 of 89.72% while the CNN
classifier, LSTM algorithm, CNN-LSTM approach, and FDSNeXt system attain reduce 𝑟𝑒𝑐𝑎𝑙 values of
90.67, 89, 90.05, and 99.41% correspondingly. At last, based on 𝐹𝑠𝑐𝑜𝑟𝑒 , the FDC-JBOADL technique
accomplishes higher 𝐹𝑠𝑐𝑜𝑟𝑒 of 89.69% while the CNN classifier, LSTM algorithm, CNN-LSTM approach,
and FDSNeXt methodology attain lesser 𝐹𝑠𝑐𝑜𝑟𝑒 values of 91.39, 90.44, 90.75, and 99.30% correspondingly.
These performances guaranteed the excellent solution of the FDC-JBOADL technique.

Table 2. Classifier outcome of FDC-JBOADL algorithm on CAUCAFall database

Metrics Training set (%) Testing set (%)
Accuracy 98.81 98.33
Precision 98.72 98.27
Recall 98.54 98.12
F-Score 98.50 98.20
MCC 98.18 98.04

Figure 7. Classifier outcome of FDC-JBOADL algorithm on CAUCAFall dataset

Figure 8. 𝐴𝑐𝑐𝑢𝑦 curve of FDC-JBOADL algorithm on CAUCAFall dataset

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1461-1470

Int J Artif Intell ISSN: 2252-8938  1469

Figure 9. Loss curve of FDC-JBOADL algorithm on CAUCAFall dataset

Table 3. Comparative outcome of FDC-JBOADL algorithm with other methods

Model 𝐴𝑐𝑐𝑢𝑦 (%) 𝑃𝑟𝑒𝑐𝑛 (%) 𝑅𝑒𝑐𝑎𝑙 (%) 𝐹𝑆𝑐𝑜𝑟𝑒 (%)
CNN algorithm 85.69 91.44 89.72 89.69
LSTM algorithm 90.12 91.28 90.67 91.39
CNN-LSTM 84.04 89.90 89.00 90.44
FDSNeXt 91.87 91.07 90.05 90.75
FDC-JBOADL 99.39 99.39 99.41 99.30

5. CONCLUSION
This research highlights the efficacy of the FDC-JBOADL technique in automating the recognition
and classification of fall events through the integration of advanced DL models. By employing a multifaceted
approach that includes MF-based noise removal, EfficientNet for feature extraction, LSTM for fall detection,
and JBOA for hyperparameter optimization, the FDC-JBOADL technique demonstrates significant
improvements in recognizing and classifying fall incidents. The methodology not only leverages the strengths
of EfficientNet in capturing both motion patterns and individual appearance characteristics but also ensures
optimal performance through meticulous hyperparameter tuning. The comprehensive experimental validation
underscores the excellent solution of the FDC-JBOADL technique compared to other systems, reinforcing its
potential as a robust solution for fall detection in various applications. Future work concentrates on further
refining the technique and exploring its applicability in real-time monitoring systems, ultimately contributing
to enhanced safety and well-being for individuals at risk of falls.

REFERENCES
[1] X. Jiang, L. Zhang, and L. Li, “Multi-task learning radar transformer (MLRT): a personal identification and fall detection network
based on IR-UWB radar,” Sensors, vol. 23, no. 12, 2023, doi: 10.3390/s23125632.
[2] H. Sadreazami, M. Bolic, and S. Rajan, “Contactless fall detection using time-frequency analysis and convolutional neural
networks,” IEEE Transactions on Industrial Informatics, vol. 17, no. 10, pp. 6842–6851, 2021, doi: 10.1109/TII.2021.3049342.
[3] A. De, A. Saha, P. Kumar, and G. Pal, “Fall detection approach based on combined two-channel body activity classification for
innovative indoor environment,” Journal of Ambient Intelligence and Humanized Computing, vol. 14, no. 9, pp. 11407–11418,
2023, doi: 10.1007/s12652-022-03714-2.
[4] S. Mobsite, N. Alaoui, and M. Boulmalf, “A framework for elders fall detection using deep learning,” in Colloquium in
Information Science and Technology, CIST, 2020, pp. 69–74, doi: 10.1109/CiSt49399.2021.9357184.
[5] D. Kraft, K. Srinivasan, and G. Bieber, “Deep learning based fall detection algorithms for embedded systems, smartwatches, and
IoT devices using accelerometers,” Technologies, vol. 8, no. 4, 2020, doi: 10.3390/technologies8040072.
[6] K. C. Liu, K. H. Hung, C. Y. Hsieh, H. Y. Huang, C. T. Chan, and Y. Tsao, “Deep-learning-based signal enhancement of low-
resolution accelerometer for fall detection systems,” IEEE Transactions on Cognitive and Developmental Systems, vol. 14, no. 3,
pp. 1270–1281, 2022, doi: 10.1109/TCDS.2021.3116228.
[7] R. Delgado-Escaño, F. M. Castro, J. R. Cózar, M. J. Marín-Jiménez, N. Guil, and E. Casilari, “A cross-dataset deep learning-
based classifier for people fall detection and identification,” Computer Methods and Programs in Biomedicine, vol. 184, 2020,
doi: 10.1016/j.cmpb.2019.105265.
[8] X. Cai, X. Liu, M. An, and G. Han, “Vision-based fall detection using dense block with multi-channel convolutional fusion
strategy,” IEEE Access, vol. 9, pp. 18318–18325, 2021, doi: 10.1109/ACCESS.2021.3054469.

Enhancing fall detection and classification using jarratt‐butterfly optimization … (Kakirala Durga Bhavani)
1470  ISSN: 2252-8938

[9] C. Yao, J. Hu, W. Min, Z. Deng, S. Zou, and W. Min, “A novel real-time fall detection method based on head segmentation and
convolutional neural network,” Journal of Real-Time Image Processing, vol. 17, no. 6, pp. 1939–1949, 2020, doi:
10.1007/s11554-020-00982-z.
[10] G. V. Leite, G. P. da Silva, and H. Pedrini, “Three-stream convolutional neural network for human fall detection,” Advances in
Intelligent Systems and Computing, vol. 1232, pp. 49–80, 2021, doi: 10.1007/978-981-15-6759-9_3.
[11] Y. Li, Z. Zuo, and J. Pan, “Sensor-based fall detection using a combination model of a temporal convolutional network and a
gated recurrent unit,” Future Generation Computer Systems, vol. 139, pp. 53–63, 2023, doi: 10.1016/j.future.2022.09.011.
[12] N. D. Raeve et al., “Bluetooth-low-energy-based fall detection and warning system for elderly people in nursing homes,” Journal
of Sensors, vol. 2022, 2022, doi: 10.1155/2022/9930681.
[13] A. Rezaei et al., “Unobtrusive human fall detection system using mmwave radar and data driven methods,” IEEE Sensors
Journal, vol. 23, no. 7, pp. 7968–7976, 2023, doi: 10.1109/JSEN.2023.3245063.
[14] L. Yao, W. Yang, and W. Huang, “A fall detection method based on a joint motion map using double convolutional neural
networks,” Multimedia Tools and Applications, vol. 81, no. 4, pp. 4551–4568, 2022, doi: 10.1007/s11042-020-09181-1.
[15] S. Mekruksavanich, P. Jantawong, A. Charoenphol, and A. Jitpattanakul, “Fall detection from smart wearable sensors using deep
convolutional neural network with squeeze-and-excitation module,” in ICSEC 2021 - 25th International Computer Science and
Engineering Conference, 2021, pp. 448–453, doi: 10.1109/ICSEC53205.2021.9684626.
[16] R. S. Jawad and H. Abid, “HVDC fault detection and classification with artificial neural network based on ACO-DWT method,”
Energies, vol. 16, no. 3, 2023, doi: 10.3390/en16031064.
[17] D. W. Lee, K. Jun, K. Naheem, and M. S. Kim, “Deep neural network–based double-check method for fall detection using IMU-L
sensor and RGB camera data,” IEEE Access, vol. 9, pp. 48064–48079, 2021, doi: 10.1109/ACCESS.2021.3065105.
[18] R. Chaganti, V. Ravi, and T. D. Pham, “Image-based malware representation approach with EfficientNet convolutional neural
networks for effective malware classification,” Journal of Information Security and Applications, vol. 69, 2022, doi:
10.1016/j.jisa.2022.103306.
[19] I. D. Mienye and Y. Sun, “A deep learning ensemble with data resampling for credit card fraud detection,” IEEE Access, vol. 11,
pp. 30628–30638, 2023, doi: 10.1109/ACCESS.2023.3262020.
[20] R. Sihwail, O. Said Solaiman, and K. A. Zainol Ariffin, “New robust hybrid Jarratt-Butterfly optimization algorithm for nonlinear
models,” Journal of King Saud University - Computer and Information Sciences, vol. 34, no. 10, pp. 8207–8220, 2022, doi:
10.1016/j.jksuci.2022.08.004.
[21] X. Yu, J. Jang, and S. Xiong, “KFall: A comprehensive motion dataset to detect pre-impact fall for the elderly based on wearable
inertial sensors,” KFall Dataset. 2021. [Online]. Available: https://sites.google.com/view/kfalldataset
[22] X. Yu, J. Jang, and S. Xiong, “A large-scale open motion dataset (KFall) and benchmark algorithms for detecting pre-impact fall
of the elderly using wearable inertial sensors,” Frontiers in Aging Neuroscience, vol. 13, 2021, doi: 10.3389/fnagi.2021.692865.
[23] J. C. E. Guerrero, E. M. España, M. M. Añasco, and J. E. P. Lopera, “Dataset for human fall recognition in an uncontrolled
environment,” Data in Brief, vol. 45, 2022, doi: 10.1016/j.dib.2022.108610.
[24] N. Hnoohom, S. Mekruksavanich, and A. Jitpattanakul, “Pre-impact and impact fall detection based on a multimodal sensor using
a deep residual network,” Intelligent Automation and Soft Computing, vol. 36, no. 3, pp. 3371–3385, 2023, doi:
10.32604/iasc.2023.036551.
[25] M. K. Yi, K. Han, and S. O. Hwang, “Fall detection of the elderly using denoising LSTM-based convolutional variant
autoencoder,” IEEE Sensors Journal, vol. 24, no. 11, pp. 18556–18567, 2024, doi: 10.1109/JSEN.2024.3388478.

BIOGRAPHIES OF AUTHORS

Kakirala Durga Bhavani is doing research on deep learning in Department of

Computer Science and Engineering at SRM Institute of Science and Technology, India. She
did her post-graduation in computer science and engineering at Amrita Vishwa Vidyapeetham
Coimbatore. She has published few papers in top journals. She can be contacted at email:
[email protected].

Melkias Ferni Ukrit holds a Ph.D. in computer science and engineering from
Sathyabama Institute of Science and Technology Chennai, India. She is working as Associate
Professor in SRM Institute of Science and Technology, Kattankulathur. Her main area of
research includes image processing, machine learning, deep learning, and IoT. She is a life
member of the Indian Society for Technical Education (ISTE). She has published several
papers in well-known peer-reviewed journals. She can be contacted at email:
[email protected].

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1461-1470