biosensors

Article
The Applications of Metaheuristics for Human Activity
Recognition and Fall Detection Using Wearable Sensors:
A Comprehensive Analysis
Mohammed A. A. Al-qaness 1, * , Ahmed M. Helmi 2,3 , Abdelghani Dahou 4,5 and Mohamed Abd Elaziz 6,7,8,9
1 College of Physics and Electronic Information Engineering, Zhejiang Normal University, Jinhua 321004, China
2 College of Engineering and Information Technology, Buraydah Private Colleges,
Buraydah 51418, Saudi Arabia
3 Computer and Systems Engineering Department, Faculty of Engineering, Zagazig University,
Zagazig 44519, Egypt
4 Mathematics and Computer Science Department, University of Ahmed DRAIA, Adrar 01000, Algeria
5 LDDI Laboratory, Faculty of Science and Technology, University of Ahmed DRAIA, Adrar 01000, Algeria
6 Department of Mathematics, Faculty of Science, Zagazig University, Zagazig 44519, Egypt
7 Artificial Intelligence Research Center (AIRC), College of Engineering and Information Technology,
Ajman University, Ajman 346, United Arab Emirates
8 Faculty of Computer Science and Engineering, Galala University, Suez 435611, Egypt
9 Department of Electrical and Computer Engineering, Lebanese American University, Byblos 13-5053, Lebanon
* Correspondence: [email protected]

Abstract: In this paper, we study the applications of metaheuristics (MH) optimization algorithms
in human activity recognition (HAR) and fall detection based on sensor data. It is known that MH
algorithms have been utilized in complex engineering and optimization problems, including feature
selection (FS). Thus, in this regard, this paper used nine MH algorithms as FS methods to boost
the classification accuracy of the HAR and fall detection applications. The applied MH were the
Aquila optimizer (AO), arithmetic optimization algorithm (AOA), marine predators algorithm (MPA),





 artificial bee colony (ABC) algorithm, genetic algorithm (GA), slime mold algorithm (SMA), grey
Citation: Al-qaness, M.A.A.; Helmi, wolf optimizer (GWO), whale optimization algorithm (WOA), and particle swarm optimization
A.M.; Dahou, A.; Abd Elaziz, M. The algorithm (PSO). First, we applied efficient prepossessing and segmentation methods to reveal the
Applications of Metaheuristics for motion patterns and reduce the time complexities. Second, we developed a light feature extraction
Human Activity Recognition and Fall technique using advanced deep learning approaches. The developed model was ResRNN and was
Detection Using Wearable Sensors: A composed of several building blocks from deep learning networks including convolution neural
Comprehensive Analysis. Biosensors
networks (CNN), residual networks, and bidirectional recurrent neural networks (BiRNN). Third, we
2022, 12, 821. https://doi.org/
applied the mentioned MH algorithms to select the optimal features and boost classification accuracy.
10.3390/bios12100821
Finally, the support vector machine and random forest classifiers were employed to classify each
Received: 31 July 2022 activity in the case of multi-classification and to detect fall and non-fall actions in the case of binary
Accepted: 28 September 2022 classification. We used seven different and complex datasets for the multi-classification case: the
Published: 3 October 2022 PAMMP2, Sis-Fall, UniMiB SHAR, OPPORTUNITY, WISDM, UCI-HAR, and KU-HAR datasets. In
Publisher’s Note: MDPI stays neutral addition, we used the Sis-Fall dataset for the binary classification (fall detection). We compared the
with regard to jurisdictional claims in results of the nine MH optimization methods using different performance indicators. We concluded
published maps and institutional affil- that MH optimization algorithms had promising performance in HAR and fall detection applications.
iations.

Keywords: fall detection; human activity recognition wearable sensors; deep learning (DL); convolution
neural network

Copyright: © 2022 by the authors.

Licensee MDPI, Basel, Switzerland.
This article is an open access article
1. Introduction
distributed under the terms and
1.1. Motivation
conditions of the Creative Commons
Attribution (CC BY) license (https:// Many context-awareness services allow computers to track and recognize human
creativecommons.org/licenses/by/ motion and activities. It is clear that smart environments can be investigated due to the
4.0/). increased usage of smart devices in homes as well as Internet of things (IoT)-supported

Biosensors 2022, 12, 821. https://doi.org/10.3390/bios12100821 https://www.mdpi.com/journal/biosensors

Biosensors 2022, 12, 821 2 of 21

devices [1]. Human activity recognition (HAR) can be defined as the process of identifying
a person’s actions from a series of measurements captured by different mechanisms such
as cameras (computer vision mechanism) [2], interior sensors [3,4], radars [5], wireless
signals [6], and others.
This study focuses only on the HAR applications that use sensor data. Sensor-based
HAR mechanisms have the advantage of being able to collect sensing data at any time and
in any location as well as obtain information that is unique to each user. The disadvantage
is that each user must own the sensor equipment; however, the widespread popularity of
smartphones and smartwatches has solved this problem. In addition, there are still some
issues, such as different measurement conditions, e.g., device kind; the installed position
of the sensors; the wearing technique; and the measuring applications, which differ from
one user to the next and from one measurement date to the next. To solve these challenges,
in this study, we use different and complex datasets that cover all types of sensor data
proposed for HAR applications.
The classification process in HAR applications is a challenge due to the complexity
of the sensors’ datasets. Thus, the feature selection process plays a significant role in
HAR applications to reduce computation time and select only the optimal feature set.
The selected MH optimization algorithms have been widely employed for feature selection
in different domains. For example, the AO was utilized as an FS technique to enhance the
intrusion detection system in the IoT and cloud environments [7,8]. In [9], the AO was
employed as an FS for COVID-19 image classification. The AOA was utilized in various FS
tasks such as osteosarcoma detection [10] and our previous study of HAR [11]. The MPA
was also utilized in different applications, such as COVID-19 CT image classification [12],
breast cancer classification [13], and wind power forecasting [14]. The SMA was adopted in
different FS applications such as medical data classification [15], parameter identification
of photovoltaic systems [16], and crude oil forecasting. The GA was also utilized for
different applications such as lung cancer classification [17], data mining [18], and credit
risk assessment [19]. The GWO was adopted in advanced applications such as mammogram
image analysis [20], Parkinson’s disease diagnosis [21], and chronic disease prediction [22].
The WOA was applied in several FA applications such as email spam detection [23],
software fault prediction [24], and medical data classification [25]. The ABC was adopted
in different FS tasks such as colon cancer detection [26], medical image classification [27],
and IDS [28]. The PSO is one of earlier methods that was employed for different FS such as
data mining in the oil industry [29], breast cancer recurrence prediction [30], and intrusion
detection system (IDS) [31].

1.2. Paper—Main Contributions

In this paper, we propose an efficient HAR approach that can be utilized for multi-
classification (for different daily activities) and binary classification (for fall detection,
including fall or non-fall action). The developed approach depends on two main processes.
The first is feature extraction in which a light deep learning (DL) approach called ResRNN
is developed to extract a subset of features that represent human motion from the sensor
data. The ResRNN is built based on several building blocks from DL networks includ-
ing convolution neural networks (CNN), residual networks, and bidirectional recurrent
neural network (BiRNN). The second approach is to leverage recent advances in MH
optimization algorithms in feature selection. We tested nine MH algorithms in feature
selection to build an efficient HAR system, namely the Aquila optimizer (AO), arithmetic
optimization algorithm (AOA), marine predators algorithm (MPA), artificial bee colony
(ABC) algorithm, genetic algorithm (GA), slime mold algorithm (SMA), grey wolf optimizer
(GWO), whale optimization algorithm (WOA), and particle swarm optimization algorithm
(PSO). The main problems of sensor-based HAR approaches are the different positions
of sensors in the human body, the types of motions, the complexities of human activities,
the number of activities, and the number of tested users (the people who implemented
the tested activities). To solve those challenges and comprehensively analyze the MH
algorithms’ applications in HAR applications, we used seven datasets that covered all
Biosensors 2022, 12, 821 3 of 21

the mentioned challenges, namely, UCI-HAR, WISDM, UniMiB SHAR, OPPORTUNITY,

KU-HAR, Sis-Fall, and PAMMP2. To sum up, we present the following contributions to the
field of HAR:
• We studied the impacts of metaheuristic (MH) optimization algorithms on human
activity recognition (HAR) and fall detection using body-attached sensor data. We
tested nine MH algorithms and compared their performances.
• We developed a light feature extraction approach called ResRNN using several deep
learning models, such as convolution neural networks (CNN), residual networks,
and bidirectional recurrent neural network (BiRNN), to expose the related features
from the collected signal data.
• We examined the suggested feature selection methods based on MH algorithms using
different and complex datasets that covered all the aspects of sensor data for HAR and
fall-detection applications.

1.3. Paper—Organization
In Section 2, we describe several recently published studies for sensor-based human
activity recognition using different datasets. In Section 3, we describe the preliminaries of
the applied methods including the nine MH optimization algorithms, the basic definitions
of the seven datasets, the developed feature extraction method, and the feature optimiza-
tion process. Moreover, the evaluation experiments’ settings, evaluation, comparison,
and results are presented in Section 4. We conclude the paper in Section 6.

2. Related Work
In this section, we present a quick review of the previous studies proposed for senor-
based HAR applications using different public datasets. Deep learning models have been
widely utilized in different fields [32] including HAR and fall detection.
In [33], the authors proposed a new deep learning method using a hybrid gated
recurrent unit (GRU) and LSTM -based RNN model for HAR. They used the TRECVID
dataset to test the performance of the proposed deep learning model, which showed
significant performance. Wang et al. [34] applied the traditional CNN model with an
attention mechanism to extract sensor data features. They used the well-known UCI-HAR
dataset to assess the performance of the CNN-based HAR model. Xia et al. [35] applied a
hybrid LSTM and CNN to classify different human activities using three public datasets:
OPPORTUNITY, UCI-HAR, and WISDM. The main goal of this combined model was to
automatically extract features from the sensor data with fewer parameters to reduce the
computation complexity. Sikder et al. [36] employed a sequential deep learning method to
classify human activities using three public datasets: UCI-HAR, KU-HAR, and WISDM.
They applied a specific preprocessing method using a specific matrix formulated from
sensor data. The classification outcomes achieved high accuracy rates for the three datasets.
Kumar and Suresh [37] proposed a new HAR model called DeepTransHHAR using deep
learning with heterogeneous deep transfer learning. They used two well-known datasets,
KU-HAR and HHAR, to evaluate the developed DeepTransHHAR, achieving acceptable
classification results. Dua et al. [38] presented a combined deep learning model using the
CNN and GUR to extract features automatically and recognize implemented activities.
They evaluated the combined CNN-GRU with three public datasets, PAMAP2, WISDM,
and UCI-HAR. Khatun et al. [39] proposed a combined deep learning model (CNN-LSTM)
with a self-attention mechanism for HAR smartphone applications. They assessed the
quality of the combined model using two benchmark datasets, UCI-HAR and MHEALTH.
Ghate et al. [40] applied a hybrid of deep CNN with the random forest classifier to classify
human activities using two datasets, WISDM and UCI-HAR. It was compared to other
models, such as LSTM, GRU, and CNN, and they found that CNN with RF obtained the
best classification accuracy. Ronald et al. [41] developed a deep leering model based on
Inception-ResNet architecture called iSPLInception to classify different human activities.
They assessed the quality of the proposed model with four well-known HAR datasets, UCI
Biosensors 2022, 12, 821 4 of 21

HAR, OPPORTUNITY, PAMAP2, and Daphnet freezing of the gait dataset. Tufek et al. [42]
studied the capability of the LSTM and CNN models to classify human activities based
on sensor data. They used the UCI-HAR dataset to evaluate both the LSTM and CNN
models. They found that the LSTM model had better classification accuracy than the
CNN model. Gao et al. [43] developed a new HAR model based on a multi-branch CNN
architecture using a selective-kernel mechanism. They employed four datasets to test the
performance of the developed CNN model, namely UCI-HAR, WISDM, UniMiB SHAR,
OPPORTUNITY, and PAMAP2. Huang et al. [44] employed a shallow CNN model with
channel selectivity for HAR applications. They used five benchmark datasets to verify the
performance of the CNN-based HAR model called OPPORTUNITY, namely the UCI-HAR,
PAMAP2, WISDM, and UniMiB SHAR datasets. Gao et al. [45] suggested a dual attention
deep learning approach, namely DanHAR, for HAR applications using sensor data. This
model can blend channel and temporal attention on the residual network to enhance feature
representation capability. The WISDM, UniMiB SHAR, PAMAP2, and OPPORTUNITY
datasets were utilized to assess the classification capability of the developed DanHAR
model, which obtained higher accuracy rates than several other models. Tang et al. [46]
presented a CNN-based HAR method. The main idea was to boost the multi-scale feature
representation capability using one feature layer. The UniMiB SHAR, PAMAP2, UCI-HAR,
and WISDM datasets were utilized to assess the method performance with comparisons to
several CNN models. The outcomes showed that the applied CNN-based feature extraction
approach boosted the classification accuracy of all datasets.
It is worth mentioning that MH optimization algorithms including SI methods have
limited applications in HAR systems. Almost all of the MH and SI method applications
were in computer vision-based HAR applications such as the PSO [47,48] and genetic
algorithm (GA) [49]. For sensor data, we carried out a simple implementation for the
arithmetic optimization algorithm (AOA) with the KU-HAR, UCI-HAR, and WISDM
datasets from [11] and the grey wolf optimizer using WISDM and UCI-HAR [50].
Unlike previous approaches, this paper presents a comprehensive analysis of MH
optimization algorithms in HAR applications using sensor data. This is the first com-
prehensive study that investigates nine MH algorithms in HAR applications with seven
public datasets, including comprehensive as well as normal and complex daily activities.
The paper opens a new direction for further investigating MH optimization algorithms to
boost HAR applications based on smart devices installed with embedded sensors.

3. Materials and Methods

In this section, we describe the preliminaries and the backgrounds of the used HAR
datasets (KU-HAR, UCI-HAR, WISDM, PAMAP2, OPPORTUNITY, SiS-Fall, and UniMiB
SHAR) and the applied feature selection methods based on different MH optimization
algorithms, namely the Aquila optimizer (AO), arithmetic optimization algorithm (AOA),
marine predators algorithm (MPA), artificial bee colony (ABC) algorithm, genetic algorithm
(GA), slime mold algorithm (SMA), grey wolf optimizer (GWO), whale optimization
algorithm (WOA), and particle swarm optimization algorithm (PSO). The main workflow
of the HAR approach described in this paper can be seen in Figure 1. It consists of four
main stages. The first is the data collection from the body-attached sensors. In this study,
we used seven public datasets containing comprehensive and complex activities in different
environments and were collected by different people in different countries. The second
stage is data prepossessing. We used different cleaning and filtering methods to obtain
clean data that can expose the impact of human activities on the collected signals. The third
stage is the feature extraction stage. In this paper, we developed a new deep learning model
to extract relevant features from the prepossessed datasets. The model is called ResRNN.
It consists of several building blocks using different deep learning structures. (BiRNN).
The fourth stage is the classification stage. In this stage, MH optimization methods are
employed to select optimal features, reduce computation complexity, and optimize the
classification process. Then, a classifier can be employed to recognize the implemented
Biosensors 2022, 12, 821 5 of 21

activities. This study tested and compared two well-known classifiers, the SVM and RF
classification models.

Figure 1. The main workflow of HAR application using the integration of deep learning and MH
optimization algorithms.

3.1. Experimental HAR Datasets

3.1.1. KU-HAR
Skider and Nahid [51] collected and built a new dataset for human daily activities
and actions, called the KU-HAR dataset. They used the gyroscopes and accelerometers
of smartphones (See Figure 2a). They collected 1945 raw samples for 18 different classes
(human activities) that were implemented by 90 users aged between 18 and 34. The collected
samples contained 20,750 subsamples, each with a period of 3 s.

Figure 2. Sensor placement on the subject’s body. A waist-mounted smartphone was used for
KUHAR, UCI-HAR, and WISDOM. WISDOM was collected using additional smartwatches. IMU
units were used to collect the signals for OPPO (positions: back, right/left upper/lateral parts of the
arm and right/left shoe ), PAMAP2 (positions: chest, wrist, and ankle), and Sis-Fall (position: waist).

3.1.2. OPPORTUNITY (Oppo)

The OPPORTUNITY dataset was collected by [52] using wearable inertia measurement
units (IMUs) that were placed on human users’ bodies at seven different positions (see
Figure 2b). Four human users implemented 18 daily motions and activities in a simulation
room, such as “close/open doors/fridges/drawers, clear tables, toggle switches, and sip
from cups,” and others. The collected samples had 77-dimensional attribute columns that
characterized the activity signals and have been sampled at a frequency of 30 Hz. It is
worth mentioning that the “emphNULL” class represents 72% of all collected samples,
which represents a non-relevant activity. More details can be found in [52].
Biosensors 2022, 12, 821 6 of 21

3.1.3. PAMAP2
The PAMAP2 dataset was collected by [53] with nine human users who implemented
different daily activities. The users were asked to wear three IMUs at the ankle, chest,
and wrist positions, as clarified in Figure 2c. During data collection, for each activity (one
motion), the magnetometer, gyroscope, and accelerometer reported 27-dimensional signal
spaces at a frequency rate of 100 Hz. A total of 2,844,868 raw samples for 12 different
human activities and motions were collected during the data collection period.

3.1.4. Sis-Fall
The SiS-Fall dataset was collected by [54] using an embedded unit with a Kinets
MKL25Z128VLK4 microcontroller, which had two accelerometers (MMA8451Q and
ADXL345) and one ITG3200 gyroscope (see Figure 2d). It contained two types of hu-
man motions; the first was called activities of daily living (ADLs) and the second was called
fall actions. The ADL type contained 19 activities: jogging, walking, walking downstairs
and upstairs, standing, jumping, sitting, stumbling while walking, and different sitting
motions. The fall type contained 15 fall actions such as fall forward, fall backward, lateral
fall, fall while jogging, fall while getting up, fall while sitting down, and others. A total
of 38 human users implemented the activities, with 15 elderly users (7 female and 8 male)
and 23 young people (12 female and 11 male).

3.1.5. UCI-HAR
In the collection stage of the UCI-HAR dataset, a Samsung Galaxy SII smartphone
was used [55] (see Figure 2e). A total of 30 users implemented six daily activities: walking,
walking upstairs, walking downstairs, sitting, standing, and lying. The gyroscope and
accelerometer records were collected separately. Thus, a total of 10,299 samples were
collected at a rate of 50 Hz, with a sliding window of 2.56 s.

3.1.6. UniMiB SHAR

In [56], a new HAR dataset was collected based on the accelerometers of smartphones.
During the data collection stage, the users were asked to put the smartphone in their
pockets (right and left), as shown in Figure 2f. The UniMiB SHAR had two types of human
motions: daily activities and fall actions. The first contained nine activities, (“Running,
Walking, Standing UPFS (Standing up from sitting), Jumping, Standing UPFL (Standing
up from lying), Going Ups (Going upstairs), Going Downs (Going downstairs), sitting
down, and Lying DownFS (Lying down from standing)”). The second type had eight fall
actions (“Hitting Obstacle, Falling right, Falling Back, Falling Left, Falling BackSC (Falling
backward sitting chair), Falling with protection strategy (Falling withPS), Falling Forward,
and Syncope”). A total of 30 human users were asked to implement the mentioned motions,
and 11,771 raw samples were collected during the experiments.

3.1.7. WISDM
The wireless sensor data mining (WISDM) dataset was collected by [57] using Android
smart devices (i.e., smartwatches and smartphones; see Figure 2g) ). A total of 36 users
implemented six physical activities, such as jogging, walking, sitting, upstairs, downstairs,
and standing. A total of 1,098,207 raw samples were collected during the experiments,
where each sample indicated an accelerometer data measurement (at 20 Hz).

3.2. Applied Metaheuristic Optimization Algorithms

3.2.1. Aquila Optimizer (AO)
The AO algorithm was developed by [58]. It is a population-based MH optimization
method. The main idea of the AO is to mimic aquilas’ natural behavior in catching prey in
the wild. The workflow and detailed description of the AO can be found in [58].
Biosensors 2022, 12, 821 7 of 21

3.2.2. Arithmetic Optimization Algorithm (AOA)

The AOA optimization method was developed by [59] and was inspired by basic
mathematical operations (i.e., −, +, ∗, and /). The basic steps and detailed description of
the workflow of the AOA can be found in [59].

3.2.3. Marine Predators Algorithm (MPA)

The MPA was developed by [60] and was inspired by the natural behavior of predators
and prey. The prey and predators can be considered search agents, whereas a predator
searches for prey and a prey searches for food. Similar to other MH techniques, it starts
by generating a set of solutions (agents) as an initialization. After that, the agents can be
modified depending on the main workflow of the algorithm. More details and mathematical
descriptions can be found in [60].

3.2.4. Slime Mold Algorithm

The SMA was developed by [61] as a natural-inspired MH technique and belongs to
swarm intelligence-inspired algorithms. The main idea of the SMA is to mimic the natural
behaviors of slime mold oscillations and their propagation wave feedback depending on
the bio-oscillator. It generates the optimum routes to connect food. More details can be
found in [61].

3.2.5. Whale Optimization Algorithm (WOA)

The WOA was proposed by [62] as a natural swarm intelligence method. It was
inspired by the behavior of humpback whales in nature. In contrast, the whale’s position
indicates the agent’s solution to a problem and it can be updated depending on the whale’s
behavior when attacking prey. Two attacking techniques are used in the mathematical
definition of the WOA. More details can be found in [62].

3.2.6. Artificial Bee Colony (ABC) Algorithm

The ABC algorithm is a swarm intelligence-based MH method developed by Karaboga
in 2005 [63] to solve complex optimization, numerical, and engineering problems. The
ABC was inspired by the intelligent foraging behavior of honey bees. The ABC generally
depends on the model developed by [64] based on the foraging behavior of honey bee
colonies. This model has three phases, employed and unemployed foraging bees and food
sources. The employed and unemployed bees are searching for food sources. This behavior
can be represented in the ABC algorithm by generating a population that has agents. Thus,
an agent (a colony of artificial forager bees) searches for a good solution (a good food
source) (good solutions for a given problem). More details about the ABC algorithm can be
found in [65]

3.2.7. Grey Wolf Optimizer (GWO)

The GWO is a swarm intelligence-based method inspired by the natural behavior of
Canis lupus (grey wolves). It was developed by [66] and received wide attention for solving
different optimization problems. It simulates the leadership hierarchy’s natural behavior
and grey wolves’ hunting processes. Four kinds of grey wolves (including alpha, beta,
omega, and delta) are applied to simulate the leadership hierarchy. Additionally, there are
three main steps in the hunting process: searching for prey, encircling prey, and attacking
prey. The mathematical definition and a more detailed description can be found in [66].

3.2.8. Genetic Algorithm

The GA was presented by [67] as a population-based MH algorithm. It can be consid-
ered an evolution-based algorithm. In the definition of the GA, each individual (agent) in
the population is referred to as the solution. It has three phases that can be utilized for up-
dating the solutions (agents) called the selection, crossover, and mutation processes. In the
first phase, two agents can be randomly selected to boost the diversity of the population.
Biosensors 2022, 12, 821 8 of 21

The second phase, the crossover mechanism, can generate new agents (individuals) from
the selected agents (parents). After that, the mutation process can be utilized to replace
randomly selected agents with random values belonging to the search domain. Lastly,
the current population can be updated based on the fitness values of the newly initialized
agents and their parents. Furthermore, the population can be updated using the three
processes mentioned (selection, crossover, and mutation) until the stop criteria are met.
The mathematical definition and a detailed description of the GA can be found in [68].

3.2.9. Particle Swarm Optimization (PSO)

The PSO is a population-based intelligence method proposed by [69]. It is one of the
earliest swarm intelligence methods and has received much attention in previous decades.
The PSO’s main workflow and mathematical details can be found in [69].

3.3. Data Cleaning, Filtration, and Segmentation

In general, HAR applications are considered real-time applications in which the
recognition process must be performed immediately. Thus, the time window is generally
set between 2 and 10 s [70–72]. In this section, we briefly describe the data preparation for
all seven employed datasets.
KU-HAR. The KU-HAR dataset samples contain 300 data points; each segment is 3 s
with non-overlapping, as described in the original study of the KU-HAR dataset [51]. At the
same time, only one accelerometer was used to collect the tracing human activity samples.
The Butterworth lowpass filter was applied to remove noise and filter the collected signals.
OPPORTUNITY (Oppo). In the Oppo dataset, there are many NAN values. The col-
lected samples were gathered using wearable IMUs at a low-frequency of 30 Hz. Thus,
a light linear interpolation was performed to prepare the collected samples. Follow-
ing [45,73], the time window was set to 2.133 s, which had 64 data points. Furthermore, over-
lapping was employed for the segmentation process at a rate of 50%.
PAMAP2. To process the PAMAP2 collected samples, a lowpass filter called the
Butterworth filter was used with a cut-off frequency of 20 Hz. The time window for
activity segmentation was set to 5 s [45,74]. Overlapping also was used at a rate of 50% for
data segmentation.
SisFall. Following previous studies such as [75], for this dataset, the time window was
set to 3 s, in which 600 points were collected for each activity sample. A lowpass filter was
also utilized to remove noise, as suggested by the original study of the Sis-Fall dataset [54].
UCI-HAR. The samples of the UCI-HAR dataset were segmented with a time window
of 2.56 s, with a frequency of 50 Hz. Also, overlapping with 50% was performed. More
details can be found in the original study [55].
UniMiB SHAR. The segmentation of the UniMiB SHAR dataset was set by the original
study [56]. A time window of 3 s was used, with 150 data points for each activity sample.
WISDM. Similar to the other datasets, the Butterworth lowpass filter was employed
to remove the noise from the collected samples, with a cut–off frequency of 10 Hz. For the
collected samples, a time window with a length of 128 s was applied. Also, overlapping
with 50% was implemented to train the model, as carried out by [71,76].
It is worth mentioning that we performed standardization on the raw data before
feeding it to the DL model, where the input signals were scaled to unit variance along with
removing the mean as in Equation (1).

( x − u)
z= (1)
s
Where the mean and standard deviation of sample data, x, are represented by u and s,
respectively. And thus z becomes the corresponding standard score.
Biosensors 2022, 12, 821 9 of 21

3.4. Feature Extraction

In this section, we describe the DL model used to extract features from preprocessed
input data to learn better representations of the raw input data. The proposed DL model
named ResRNN was composed of several building blocks from various DL networks in-
cluding convolution neural networks (CNN), residual networks, and bidirectional recurrent
neural networks (BiRNN). The ResRNN used parallel and sequential alignment of the
components based on their structures. For instance, the CNN layers aligned parallel to
extract different feature maps from the input data and learn activity nature-related features.
The BiRNN layers were aligned in sequence to learn more complex and temporal-related
features based on the previous layer’s output (CNN output). The input data were com-
posed of triaxial signals (X, Y, Z), which represented each data sample based on the collected
IMU signals from sensors such as the accelerometer, gyroscope, and magnetometer. At the
beginning of the ResRNN, a set of convolution blocks aligned in parallel received the input
data and extracted the feature maps. Based on the IMU signals collected from each sensor,
the signals were distributed over the parallel convolution block with the following structure:
(Conv → BN → ReLU → RC → Max − pooling). The Conv represented a convolution
layer with a kernel size of (1 × 3), a stride of 1, padding of 1, and several filters of 64, 128,
and 256 for the triaxial signals. The BN layer was a batch normalization layer. The ReLU
layer was a Rectified Linear Unit activation layer. The RC layer was a residual connection
similar to the skip connection mechanism in residual networks. The max-pooling layer
was a max-pooling layer with a kernel size of 2 and a stride of 2. The learned feature maps
from all parallel CNN blocks with different output channels were concatenated and fed
to a BiRNN block with gated recurrent units (GRU) as the RNN structure and attention
mechanism. The ResRNN benefited from the DL techniques as it could learn more complex
representations and lower the computation complexity, thus generating a small model
size, overcoming over-fitting and reducing the training time. The learned features from
the BiGRU with the attention block were concatenated and fed to a fully connected layer
(FC) with 128 neurons, which served as our feature extraction layer. Thus, the fine-tuned
features in the FC layer were extracted and inputted into the feature selection algorithm.
Each input sample was represented by a feature vector of size 128 extracted from the FC
layer. The ResRNN’s weight was optimized for the activity classification task using the
Adam optimizer with a 1 × 10−4 learning rate. A softmax layer was placed at the top of the
model to perform the classification. In addition, the ResRNN was trained for 350 epochs
with dropout (0.3) and early stopping set to monitor the validation loss. Figure 3 shows the
architecture of the ResRNN.
Skip connection
Max-pooling
BN + ReLU
Conv (1x3)

FC + Softmax

xN
FC (128)

...
Concatenation

BiGRU

Input
IMU1 ... IMUN Acitvity
Skip connection
Max-pooling
BN + ReLU
Conv (1x3)

Figure 3. The proposed ResRNN model for feature extraction.

3.5. Feature Optimization

Let F represent the feature vector extracted by the proposed DL model, then
Equation (2) can be applied to characterize the model performance using the optimized
feature vector S according to both the classification rate improvement and feature set
reduction.
|S|
 
↓ f (S) = αCES + (1 − α) , (2)
|F |
Biosensors 2022, 12, 821 10 of 21

where CES represents model classification error for S, α ∈ [0, 1] is a commonly

  applied
|S|
balance factor to control the effect of CES and the feature set optimization |F | , where |F |
denotes the feature set cardinality. Thus, a lower value, i.e., f → 0, in Equation (2) refers to
the higher performance of the working model and vise-versa. In this work, CES is set as
the model classification accuracy defined by Equation (3).

TP
Accuracy(Acc) = (3)
TP + TN + FP + FN
where TP, TN, FP, and FN are the model rates of true positive, true negative, false positive,
and false negative, respectively. Algorithm 1 illustrates the designed cost function to work
under the applied MH algorithms here, where α is set to 0.99 and thr = 0.

Algorithm 1 Cost Function (S , Ttr , Ttst , α, thr)

1: Input: Solution S , Train Ttr and test Ttst sets, control parameter α and threshold thr.
2: Output: Solution cost (f )
3: f ←∞
4: Obtain the equivalent binary solution S B ← S > thr
5: Select the features indicated by ones in S B in each of Ttr and Ttst
6: Train the classification layer in the model using the selected features in Ttr
7: Calculate the classification predictions of the trained model for Ttst
8: Calculate f using Equation (2)

4. Experiments
4.1. Experiments Setup
Table 1 presents the parameter settings of each MH algorithm applied in this work.
For efficient computational overheads, the search settings were set to a population size
of 20 and a total number of iterations of 20. To characterize the average performance of
each applied MH algorithm, each experiment was repeated for 20 independent runs. Both
binary and multi-class classification tasks were carried out using the SVM algorithm [77]
and RF algorithm [78] under the ’Bag’ learning method, where the size of decision trees
was set to 50.

Table 1. Parameter settings of used MH algorithms.

Algorithm Parameters
ABC a=1
AO α = 0.1, δ = 0.1, u = 0.0265, r0 = 10, ω = 0.005, φ0 = 3π/2,
AOA MOPMax = 1, MOPMin = 0.2, α = 2.5, µ = 0.1
GA Crossover = 40%, Mutation = 15%
GWO a:2→0
MPA FADs = 0.2, P = 0.5
PSO Inertia weight (w) = 1, wdamp = 0.99, c1 = 2, c2 = 2
SMA z = 0.03
WOA a : 2 → 0, a2: −1 → −2, b = 1

In addition to Acc, precision, recall, and the F1 measure defined in Equations (3)–(6)
were used here to characterize the performance of the proposed model.

TP
Precision(Prec) = (4)
( TP + FP)

TP
Recall(Rec) = (5)
( TP + FN )
Prec × Rec
F1-measure(F1-m) = 2 × (6)
( Prec + Rec)
Biosensors 2022, 12, 821 11 of 21

The classification results were validated based on previous studies where the KU-HAR,
UCI-HAR, UniMiB SHAR, and WISDM datasets were randomly split into 70% training
and 30% testing subsets. For the OPPORTUNITY dataset, the data records of the second,
third, and fourth trials of participants 1, 2, and 3 were used only for testing. For PAMAP2,
the testing set was also selected for the data records of subjects 5 and 6, which were never
introduced to the model during training. For Sis-Fall, 10-fold cross-validation was applied.

4.2. Results
In this section, the performances of all the applied MH methods are discussed and
analyzed. The comparison results are given in Tables 2–12. In addition, Table 2 illustrates
the average of the classification accuracy obtained using RF and SVM as classifiers to
assess the different MH techniques. Additionally, the convergence curve for compared MH
optimization methods using all datasets are displayed in Figure 4.

10-3
0.128 0.0665 0.088 5
ABC
AO 0.066 0.086
0.126 AOA 4
GA 0.0655
0.084
GWO
Cost Function

Cost Function

Cost Function

Cost Function
MPA 0.065
3
0.124 0.082
PSO
SMA 0.0645
WOA 0.08
2
0.064
0.122
0.078
0.0635
1
0.063 0.076
0.12

0.0625 0.074 0
2 4 6 8 10 12 14 16 18 20 5 10 15 20 2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20
Iteration Iteration Iteration Iteration

(a) KUHAR-RF (b) OPPO-RF (c) PAMAP2-RF (d) SisFallB-RF

0.108 0.058 0.145 0.016

0.056
0.144
0.107 0.015
0.054
0.143
Cost Function

Cost Function

Cost Function

Cost Function

0.014
0.106 0.052
0.142
0.05
0.013
0.105 0.141
0.048
0.012
0.14
0.046
0.104

0.044 0.139 0.011

5 10 15 20 2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20
Iteration Iteration Iteration Iteration

(e) SisFallM-RF (f) UCI-HAR-RF (g) UniMiB-RF (h) WISDM-RF

10-3
0.143 0.067 0.095 5

0.142 0.0665
0.09 4
0.066
0.141
Cost Function

Cost Function

Cost Function

Cost Function

0.0655
0.085 3
0.14
0.065
0.139
0.08 2
0.0645
0.138
0.064
0.075 1
0.137 0.0635

0.136 0.063 0.07 0

2 4 6 8 10 12 14 16 18 20 5 10 15 20 2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20
Iteration Iteration Iteration Iteration

(i) KUHAR-SVM (j) OPPO-SVM (k) PAMAP2-SVM (l) SisFallB-SVM

Figure 4. Cont.
Biosensors 2022, 12, 821 12 of 21

0.113 0.058 0.149 0.017

0.148
0.112 0.056 0.016
0.147
0.111 0.054 0.015
Cost Function

Cost Function

Cost Function

Cost Function
0.146

0.11 0.052 0.145 0.014

0.144
0.109 0.05 0.013
0.143
0.108 0.048 0.012
0.142

0.107 0.046 0.141 0.011

2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20
Iteration Iteration Iteration Iteration

(m) SisFallM-SVM (n) UCI-HAR-SVM (o) UniMiB-SVM (p) WISDM-SVM

Figure 4. Convergence curve for applied MH optimizers for each studied dataset using RF and
SVM classifiers.

It can be seen in Table 2 that the WOA had the highest accuracy for PAMAP2, SisFallB
(binary classification for fall detection), UCI-HAR, and WISDM. This was followed by the
GWO, which had the best accuracy using three datasets Oppo, UniMiB SHAR, and WISDM.
Meanwhile, the AO and PSO algorithms achieved the highest accuracy on two datasets,
and the MPA and SMA provided better accuracy in one dataset only. In addition, we
observed that RF provided better accuracy than SVM among the tested datasets and
competitive algorithms.

Table 2. Average classification accuracy (%) of applied MH optimizers.

KUHAR OPPO PAMAP2 SisFallB SisFallM UCI-HAR UniMiB WISDM

RF SVM RF SVM RF SVM RF SVM RF SVM RF SVM RF SVM RF SVM
ABC 88.03 86.61 93.73 93.84 92.06 92.13 99.97 99.97 89.59 89.50 94.85 94.76 85.74 86.02 98.83 98.83
AO 88.3 86.83 93.81 93.76 92.44 92.2 99.98 99.97 89.92 89.33 95.37 95.07 86.17 85.75 98.95 98.87
AOA 87.89 86.66 93.64 93.61 92.23 91.44 99.97 99.97 89.82 89.26 95.12 94.9 86 85.69 98.79 98.85
GA 87.95 86.7 93.74 93.89 91.67 92.25 99.97 99.97 89.47 89.62 95.15 94.75 85.87 86.1 98.85 98.82
GWO 88.4 86.69 93.90 93.9 92.5 92.31 99.97 99.97 89.93 89.65 95.13 94.77 86.25 86.1 98.95 98.83
MPA 88.41 86.6 93.88 93.86 92.58 92.02 99.97 99.97 89.96 89.50 95.27 94.75 86.22 86.02 98.94 98.82
PSO 88.32 86.71 93.87 93.91 92.58 92.1 99.97 99.97 90.00 89.54 95.21 94.78 86.21 86.05 98.95 98.82
SMA 87.77 86.38 93.72 93.76 91.63 91.34 99.97 99.97 89.52 89.37 94.62 94.82 85.71 85.82 98.85 98.8
WOA 88.31 86.7 93.81 93.75 92.57 92.63 99.98 99.97 89.91 89.33 95.51 95.28 86.14 85.85 98.95 98.87

Furthermore, Table 3 depicts the accuracy of the classification obtained using the
best model (i.e., RF or SVM). Highest accuracy for KUHAR, OPPO, PAMAP2, SisFallB
(binary classification), SisFallM (multi-classification), UCI-HAR, UniMiB, and WISDM was
achieved with MPA, GWO, WOA, AO/WOA, AO, WOA, AO, and AO/WOA, respectively.
Moreover, from these results, we can conclude that RF still performs better than SVM.

Table 3. Classification accuracy (%) of the best model for each dataset.

KUHAR OPPO PAMAP2 SisFallB SisFallM UCI-HAR UniMiB WISDM

RF RF/SVM SVM RF RF RF RF RF
ABC 88.37 93.79 92.36 99.97 89.92 95.11 85.96 98.85
AO 88.53 93.92 92.48 100 90.13 95.49 86.44 98.99
AOA 88.09 93.73 91.57 99.97 89.86 95.22 86.13 98.81
GA 88.07 93.78 92.38 99.97 89.59 95.39 86.3 98.91
GWO 88.49 93.95 92.42 99.97 89.98 95.42 86.33 98.97
MPA 88.52 93.91 92.3 99.97 90.04 95.39 86.35 98.97
PSO 88.44 93.91 92.32 99.97 90.1 95.25 86.33 98.97
SMA 88.06 93.86 91.71 99.97 89.68 94.74 86.04 98.87
WOA 88.42 93.83 92.95 100 89.98 95.59 86.18 98.99

Table 4 illustrates the average reduction rate of the features obtained by each MH
technique and both classifiers (i.e., RF and SVM). From these values, it can be seen that
Biosensors 2022, 12, 821 13 of 21

using the AO with RF and the AOA with SVM was better than other techniques on the KU-
HAR dataset and the PSO with RF and the MOA with SVM were the better MH techniques
on the Oppo dataset. In addition, the GWO and PSO provided better results than the
other models on the PAMAP2 dataset using RF; however, by using SVM, the WOA and
MPA were the best techniques. The SMA reduced the number of selected features better
than the others on the SisFallB dataset using either RF or SVM. From the results of the
algorithms on the SisFallM dataset, one can see that the PSO was better than the compared
MH algorithms when the RF classifier was used, but by using SVM, we can see that the
SMA was the best MH algorithm. For the last three datasets, we can see that the SMA
had the smallest number of features on the UCI-HAR dataset with RF/SVM, UniMiB with
SVM, and WISDM using SVM. However, the PSO with RF and the WOA with RF were the
better algorithms according to their feature reduction rates on the UniMiB and WISDM
datasets, respectively.
Table 5 lists the different classification metrics used to assess the best performing MH
and classification models using the KU-HAR dataset. The AO and RF techniques showed
the best performance among the other techniques on the KU-HAR dataset. As shown
in Table 5, most of the activities scored more than a 90% classification score in terms of
the F1 score. Activities such as Standing, Jumping, Pushing-up, and Walking-backward
scored less than 86% in terms of the F1 score due to the data imbalance. Table 6 lists the
results of classifying the Oppo dataset activities using the GWO and RF (SVM), where the
NULL class had the majority of samples dominate the results with the highest F1 score
of 96.9%. The F1 scores for the Close Drawer 1 and Close Drawer 2 activities were the
lowest among all the activities and were 33.33% and 33.96%, respectively. Table 7 lists the
results using the WOA and SVM on the PAMAP2 dataset. The Vacuuming activity scored
lower than the other activities in F1 score and Recall. Other activity scores were close to
each other due to the balanced nature of the dataset. Table 8 lists the AO and RF results on
the SisFallB dataset, which is a binary classification task. The RF model showed excellent
results, reaching an F1 score of 100% on both classes, Fall and No Fall. In terms of multiclass
classifications of the SisFall dataset, Table 9 shows the results of SisFallM using the AO
and RF on all regular activities and the Fall activity. The RF classifiers with fewer selected
features using the AO compared to the original amount of features performed well on the
classification of regular activities and scored 99.44% in terms of the F1 score on Fall activity.
Table 10 lists the results of the WAO and RF on the UCI-HAR dataset, where the Sitting
and Standing activities were less accurately classified compared to the other activities due
to the similarity between the two activity signals. Table 11 lists the AO and RF results
on the UniMiB SHAR dataset, where most falling activities were classified with high F1
scores. It can be seen that the Falling Rightward activity, Falling with protection, Falling
backward-sitting chair, and Syncope fall had lower F1 values due to the difficult nature
of the activities. Table 12 lists the WISDM dataset results using the AO and RF, which
shows almost perfect scores on all classification metrics, where the RF classifier accurately
differentiated between activities.
Biosensors 2022, 12, 821 14 of 21

Table 4. Average percentage of feature reduction (%) of applied MH optimizers.

KUHAR OPPO PAMAP2 SisFallB SisFallM UCI-HAR UniMiB WISDM

RF SVM RF SVM RF SVM RF SVM RF SVM RF SVM RF SVM RF SVM
51.41 ± 50 ± 60.31 ± 60.94 ± 57.66 ± 61.72 ± 73.13 ± 73.75 ± 51.56 ± 53.13 ± 56.72 ± 66.09 ± 54.22 ± 55.78 ± 63.91 ± 69.22 ±
ABC
4.6 6.04 5.07 3.87 4.44 3.39 2.51 2.97 3.67 8.15 3.58 4.04 7.5 7.83 5.63 2.07
39.69 ± 17.97 ± 85.47 ± 87.19 ± 63.13 ± 82.97 ± 98.44 ± 98.75 ± 51.09 ± 53.91 ± 90.78 ± 93.28 ± 66.41 ± 69.84 ± 92.81 ± 92.5 ±
AO
14.7 14.53 8.96 4.93 13.33 8.44 1.22 0.55 3.91 12.59 5.02 2.07 16.4 13.85 2.95 3.44
51.88 ± 8.59 ± 85.31 ± 90.47 ± 55.31 ± 62.34 ± 98.91 ± 98.44 ± 52.66 ± 51.88 ± 76.25 ± 78.13 ± 64.69 ± 63.44 ± 93.44 ± 95.31 ±
AOA
8.44 24.6 10.71 2.95 2.59 24.16 0.55 1.41 3.21 5.41 27.52 28.52 13.33 12.95 4.1 1.87
48.75 ± 46.56 ± 59.53 ± 62.66 ± 54.06 ± 58.59 ± 74.06 ± 73.75 ± 56.25 ± 51.41 ± 55.63 ± 62.66 ± 53.44 ± 56.72 ± 65.63 ± 67.34 ±
GA
5.13 4.45 2.39 3.83 4.97 4.47 2.05 4.28 3.61 4.32 8.7 5.54 4.72 4.51 3.94 2.39
50.47 ± 46.25 ± 64.53 ± 62.34 ± 51.88 ± 59.69 ± 71.88 ± 73.59 ± 55.47 ± 51.72 ± 59.84 ± 61.72 ± 52.5 ± 53.13 ± 67.5 ± 69.06 ±
GWO
3.78 2.59 4.77 3.42 6.91 2.3 3.46 4.27 4.9 4.76 7.13 4.06 2.17 4.47 2.19 6.35
47.81 ± 45.31 ± 59.38 ± 57.5 ± 53.28 ± 56.56 ± 70 ± 69.38 ± 55.16 ± 51.72 ± 57.81 ± 62.97 ± 53.13 ± 54.69 ± 60.31 ± 65.47 ±
MPA
1.3 4.3 4.64 5.64 11.23 2.41 5.37 2.17 2.79 5.76 3.24 3.21 1.87 4.53 3.7 1.64
46.88 ± 44.69 ± 57.5 ± 61.09 ± 51.88 ± 59.84 ± 73.13 ± 72.81 ± 50.78 ± 51.72 ± 58.28 ± 64.22 ± 50 ± 56.88 ± 61.72 ± 70.16 ±
PSO
3.67 6.61 8.44 5.63 2.19 4.83 4.28 1.64 11.11 5.93 6.73 4.15 5.96 3.11 2.74 5.59
46.88 ± 46.09 ± 58.59 ± 59.69 ± 55.31 ± 57.97 ± 63.75 ± 62.19 ± 53.75 ± 50.47 ± 53.59 ± 56.09 ± 55.78 ± 53.13 ± 60.62 ±
SMA 62.5 ± 3
2.74 5.52 4.3 3.44 5.5 6.3 2.51 1.52 3.27 5.59 4.34 10.47 5.32 5.79 2.7
47.81 ± 45.31 ± 59.38 ± 57.5 ± 53.28 ± 56.56 ± 70 ± 69.38 ± 55.16 ± 61.09 ± 57.81 ± 62.97 ± 53.13 ± 54.69 ± 60.31 ± 65.47 ±
WOA
1.3 4.3 4.64 5.64 11.23 2.41 5.37 2.17 2.79 11.43 3.24 3.21 1.87 4.53 3.7 1.64
Biosensors 2022, 12, 821 15 of 21

Table 5. Classification rates (%) of the best model of the KUHAR dataset using RF.

Activity Pre Rec F1

Standing 76.29 73.32 74.77
Sitting 98.3 95.7 96.98
Talking-Sit 94.8 96.23 95.51
Talking-Stand 92.86 95.79 94.3
Standing-Sit 96 92.31 94.12
Laying 93.85 94.38 94.12
Laying-Stand 95.2 91.21 93.16
Picking 95.71 95.3 95.5
Jumping 53.36 91.81 67.5
Pushing-up 89.82 81.82 85.63
Sitting-up 97.11 95.89 96.5
Walking 98.48 99.23 98.86
Walking-backward 80.62 33.64 47.47
Walking-circle 97.19 98.11 97.65
Running 96.52 97 96.76
Stairs up 96.94 95 95.96
Stairs down 97.84 94.44 96.11

Table 6. Classification rates (%) of the best model of the OPPO dataset using either RF or SVM.

Activity Pre Rec F1

NULL 95.37 98.49 96.9
Open Door 1 94.03 80.77 86.9
Open Door 2 97.5 90.7 93.98
Close Door 1 91.23 81.25 85.95
Close Door 2 96.67 93.55 95.08
Open Fridge 89.04 83.87 86.38
Close Fridge 88.99 71.85 79.51
Open Dishwasher 79.63 69.35 74.14
Close Dishwasher 68.52 66.07 67.27
Open Drawer 1 46.88 62.5 53.57
Close Drawer 1 46.67 25.93 33.33
Open Drawer 2 57.14 64 60.38
Close Drawer 2 27.27 45 33.96
Open Drawer 3 68.85 79.25 73.68
Close Drawer 3 97.14 70.83 81.93
Clean Table 81.4 57.38 67.31
Drink from Cup 91.94 58.76 71.7
Toggle Switch 87.5 60.87 71.79

Table 7. Classification rates (%) of the best model of the PAMAP2 dataset using SVM.

Activity Pre Rec F1

Lying 95.77 93.61 94.68
Sitting 97 95.67 96.33
Standing 84.06 91.34 87.55
Normal Walking 86.19 96.98 91.27
Running 99.78 98.7 99.24
Cycling 96.43 99.61 97.99
Nordic Walking 95.71 95.71 95.71
Ascending Stairs 91.21 95.82 93.46
Descending Stairs 95.2 93.83 94.51
Vacuuming 98.58 61.5 75.75
Ironing 83.99 90.79 87.26
Rope Jumping 97.15 99.09 98.11
Biosensors 2022, 12, 821 16 of 21

Table 8. Classification rates (%) of the best model of the SisFallB dataset using RF.

Activity Pre Rec F1

Fall 100 100 100
No Fall 100 100 100

Additionally, to test the validity of the combination of MH optimization algorithms

with the feature selection based on deep learning, Figure 5 presents a comparison of
each applied technique (i.e., deep learning (DL) model, DL model using RF and SVM
(best model (initialize results) and the optimized DL model (best model (Opt.)), which
is a combination of MH and DL). The figure shows that the combination of MH and DL
achieved the best results. In summary, we noticed that the applications of MH optimization
algorithms showed superior performance and can be further investigated for human
activity recognition and fall detection applications using body-attached sensors.

Table 9. Classification rates (%) of the best model of the SisFallM dataset using RF.

Activity Pre Rec F1

Walking Slowly 98.71 98.71 98.71
Walking Quickly 100 100 100
Jogging Slowly 98 98.39 98.2
Jogging Quickly 98.64 97.75 98.19
Walking upstairs and downstairs slowly 92.25 92.58 92.42
Walking upstairs and downstairs quickly 77.78 81.29 79.5
Sitting in a half-height chair, waiting a moment, and getting up slowly 79.86 77.08 78.45
Sitting in a half-height chair, waiting a moment, and getting up quickly 76.28 81.51 78.81
Sitting in a low-height chair, waiting a moment, and getting up slowly 70 74.34 72.1
Sitting in a low-height chair, waiting a moment, and getting up quickly 80 67.8 73.39
Sitting a moment, trying to get up, and collapsing into a chair 86.55 80.47 83.4
Sitting a moment, lying slowly, waiting a moment, and sitting again 93.91 93.91 93.91
Sitting a moment, lying quickly, waiting a moment, and sitting again 88.46 84.15 86.25
Being on one’s back, changing to lateral position, waiting a moment, and changing to one’s back 95.97 95.97 95.97
Standing, slowly bending at knees, and getting up 86.15 78.87 82.35
Standing, slowly bending without bending knees, and getting up 87.07 88.28 87.67
Standing, getting into a car, remaining seated, and getting out of the car 82.49 91.76 86.88
Stumbling while walking 97.62 95.35 96.47
Gently jumping without falling (trying to reach a high object) 88.51 82.8 85.56
Falling 99.44 99.44 99.44

Table 10. Classification rates (%) of the best model of the UCI-HAR dataset using RF.

Activity Pre Rec F1

Walking 99.59 97.98 98.78
Walking Upstairs 98.31 98.51 98.41
Walking Downstairs 96.74 99.05 97.88
Sitting 92.05 84.93 88.35
Standing 88.43 93.42 90.86
Laying Down 99.08 100 99.54
Biosensors 2022, 12, 821 17 of 21

Table 11. Classification rates (%) of the best model of the UniMiB-SHAR dataset using RF.

Activity Pre Rec F1

Standing Up From Sitting 64.33 68.24 66.23
Standing Up From Lying 87.5 77.21 82.03
Walking 76.07 75.61 75.84
Running 64.33 65.58 64.95
Going UpS 77.7 70.59 73.97
Jumping 70.06 72.37 71.2
Going DownS 94.41 96.43 95.41
Lying DownFS 92.06 88.55 90.27
Sitting Down 92.75 88.61 90.63
Falling Forward 98.14 98.6 98.37
Falling Rightward 69.79 73.63 71.66
Falling Backward 99.14 99.31 99.22
Hitting Obstacle 73.68 73.68 73.68
Falling With Protection 72.37 77.46 74.83
Falling Backward-Sitting Chair 81.82 60 69.23
Syncope Fall 52.17 61.15 56.3
Falling Leftward 98.28 97.72 98

Table 12. Classification rates (%) of the best model of the WISDM dataset using RF.

Activity Pre Rec F1

Walking 97.09 96.01 96.54
Walking Upstairs 100 98.81 99.4
Walking Downstairs 98.46 100 99.22
Sitting 95.83 95.67 95.75
Standing 100 99.95 99.97
Jogging 99.38 99.75 99.57

Figure 5. Comparison of the average accuracy of the applied DL model, DL model using RF and
SVM (best model (Init.)), and the optimized DL model (best model (Opt.).

5. Comparison with Previous Related Studies

The accuracy reported in this simulation for the different considered datasets are
compared in Table 13 to some relevant results obtained in previous studies.
From the table, we can see that the proposed approach recorded the best results in all
datasets, except for PAMAP2 in which Gao et al. [45] achieved the best results, as reported
in their study. The KUHAR dataset is so recent and there is not as much rich literature as
the above-mentioned datasets. Binary analysis for the SisFall dataset (accuracy of SisFallB
reaches 99.98%, see Table 2) revealed the excellent performance of the proposed model
Biosensors 2022, 12, 821 18 of 21

regarding the fall detection task, similar to the results reported by Mrozek et al. [79]. To the
best of our knowledge, our results are the first multiclass analysis of the SisFall dataset.

Table 13. Results of related models’ accuracies (%) for common datasets.

OPPO PAMAP2 UCI-HAR UniMiB WISDM

Multi-ResAtt [80] 86.85 Multi-ResAtt [80] 90.08 Daho et al. [11] 95.23 Multi-ResAtt [80] 74.94 LSTM-CNN [35] 95.01
Gao et al. [45] 82.75 Gao et al. [45] 93.16 LSTM-CNN [35] 95.31 DanHAR [45] 79.03 U-Net [81] 96.40
Teng et al. [73] 81 Teng et al. [73] 92.97 DSmT [82] 95.31 Teng et al. [73] 78.07 MHCA [83] 96.40
iSPLInception [41] 88.14 DanHAR [45] 93.16 Net-att3-pc-tanh [34] 93.83 Predsim ResNet [84] 80.33 DanHAR [45] 98.85
Proposed 93.90 Proposed 92.63 Proposed 95.51 Proposed 86.25 Proposed 98.95

6. Conclusions
This paper focused on applying metaheuristic (MH) optimization algorithms for daily
human activity recognition and fall detection using wearable sensors such as smartphones
and body-attached sensors. In general, we developed an efficient HAR and fall detection
model using an integration of deep learning and MH algorithms. A new feature extraction
technique based on an efficient deep learning model called ResRNN was developed to
expose the related feature from the collected sensor data. Afterward, the feature selec-
tion process was implemented using nine MH algorithms, namely, the Aquila optimizer
(AO), arithmetic optimization algorithm (AOA), slime mold algorithm (SMA), marine
predators algorithm (MPA), genetic algorithm (GA), grey wolf optimizer (GWO), whale
optimization algorithm (WOA), artificial bee colony (ABC) algorithm, and particle swarm
optimization (PSO) algorithm. These MH methods were implemented to obtain opti-
mal features, reduce computation time, and boost classification accuracy. The proposed
approach was implemented in two main areas. The first was daily activity recognition
(multi-classification) and the second was fall detection (binary classification). For compre-
hensive analysis and study, we used seven datasets that contained different and complex
activities. The used datasets were UCI-HAR, Sis-Fall, WISDOM, UNIMIB-SHAR, OPPO,
KU-HAR, and PAMMP2. Extensive experiments were conducted using two classifiers, sup-
port vector machine (SVM) and random forest (RF). The findings of this study showed
that the MH optimization algorithms showed significant performance in HAR and fall
detection applications. This study opens the possibilities of the MH to be applied for
sensing applications. We suggest using more advanced methods, such as modified MH
optimization methods integrated with intelligent search mechanisms, that may provide
more robust results, such as levy flight, opposition-based learning, and a hybridization of
two MH algorithms.

Author Contributions: Conceptualization, M.A.E.; Data curation, A.M.H.; Formal analysis, M.A.A.A.-
q.; Funding acquisition, M.A.A.A.-q.; Investigation, M.A.A.A.-q.; Methodology, M.A.A.A.-q., A.M.H.
and A.D.; Software, M.A.A.A.-q., A.M.H. and A.D.; Validation, M.A.E.; Visualization, A.M.H. and
A.D.; Writing—original draft, M.A.A.A.-q., A.M.H., A.D. and M.A.E.; Writing—review and edit-
ing, M.A.E. All authors contributed equally to this paper. All authors have read and agreed to the
published version of the manuscript.
Funding: This work was supported by the National Natural Science Foundation of China (Grant No.
62150410434), by LIESMARS Special Research Funding, and in part by the Scientific Research Center
at Buraydah Private Colleges under the research project # BPC-SRC/2022-010.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data are publicly available as described in the main text.
Conflicts of Interest: The authors declare no conflict of interest.
Biosensors 2022, 12, 821 19 of 21

References
1. Hasegawa, T. Smartphone sensor-based human activity recognition robust to different sampling rates. IEEE Sens. J. 2020,
21, 6930–6941. [CrossRef]
2. Beddiar, D.R.; Nini, B.; Sabokrou, M.; Hadid, A. Vision-based human activity recognition: A survey. Multimed. Tools Appl. 2020,
79, 30509–30555. [CrossRef]
3. Gu, F.; Chung, M.H.; Chignell, M.; Valaee, S.; Zhou, B.; Liu, X. A survey on deep learning for human activity recognition. ACM
Comput. Surv. (CSUR) 2021, 54, 1–34. [CrossRef]
4. Bouchabou, D.; Nguyen, S.M.; Lohr, C.; LeDuc, B.; Kanellos, I. A survey of human activity recognition in smart homes based on
IoT sensors algorithms: Taxonomies, challenges, and opportunities with deep learning. Sensors 2021, 21, 6037. [CrossRef]
5. Li, X.; He, Y.; Jing, X. A survey of deep learning-based human activity recognition in radar. Remote Sens. 2019, 11, 1068. [CrossRef]
6. Al-Qaness, M.A.; Abd Elaziz, M.; Kim, S.; Ewees, A.A.; Abbasi, A.A.; Alhaj, Y.A.; Hawbani, A. Channel state information from
pure communication to sense and track human motion: A survey. Sensors 2019, 19, 3329. [CrossRef]
7. Fatani, A.; Dahou, A.; Al-Qaness, M.A.; Lu, S.; Elaziz, M.A. Advanced feature extraction and selection approach using deep
learning and Aquila optimizer for IoT intrusion detection system. Sensors 2021, 22, 140. [CrossRef]
8. Dahou, A.; Abd Elaziz, M.; Chelloug, S.A.; Awadallah, M.A.; Al-Betar, M.A.; Al-qaness, M.A.; Forestiero, A. Intrusion Detection
System for IoT Based on Deep Learning and Modified Reptile Search Algorithm. Comput. Intell. Neurosci. 2022, 2022, 1–15.
[CrossRef]
9. Li, L.; Pan, J.S.; Zhuang, Z.; Chu, S.C. A Novel Feature Selection Algorithm Based on Aquila Optimizer for COVID-19
Classification. In Proceedings of the International Conference on Intelligent Information Processing, Bucharest, Romania,
29–30 September 2022; pp. 30–41.
10. Bansal, P.; Gehlot, K.; Singhal, A.; Gupta, A. Automatic detection of osteosarcoma based on integrated features and feature
selection using binary arithmetic optimization algorithm. Multimed. Tools Appl. 2022, 81, 8807–8834. [CrossRef]
11. Dahou, A.; Al-qaness, M.A.; Abd Elaziz, M.; Helmi, A. Human activity recognition in IoHT applications using arithmetic
optimization algorithm and deep learning. Measurement 2022, 199, 111445. [CrossRef]
12. Sahlol, A.T.; Yousri, D.; Ewees, A.A.; Al-Qaness, M.A.; Damasevicius, R.; Elaziz, M.A. COVID-19 image classification using deep
features and fractional-order marine predators algorithm. Sci. Rep. 2020, 10, 15364. [CrossRef] [PubMed]
13. Rajinikanth, V.; Kadry, S.; Taniar, D.; Damaševičius, R.; Rauf, H.T. Breast-cancer detection using thermal images with marine-
predators-algorithm selected features. In Proceedings of the 2021 Seventh International Conference on Bio Signals, Images, and
Instrumentation (ICBSII), Chennai, India, 25–27 March 2021; pp. 1–6.
14. Al-qaness, M.A.; Ewees, A.A.; Fan, H.; Abualigah, L.; Abd Elaziz, M. Boosted ANFIS model using augmented marine predator
algorithm with mutation operators for wind power forecasting. Appl. Energy 2022, 314, 118851. [CrossRef]
15. Wazery, Y.M.; Saber, E.; Houssein, E.H.; Ali, A.A.; Amer, E. An efficient slime mould algorithm combined with k-nearest neighbor
for medical classification tasks. IEEE Access 2021, 9, 113666–113682. [CrossRef]
16. Liu, Y.; Heidari, A.A.; Ye, X.; Liang, G.; Chen, H.; He, C. Boosting slime mould algorithm for parameter identification of
photovoltaic models. Energy 2021, 234, 121164. [CrossRef]
17. Maleki, N.; Zeinali, Y.; Niaki, S.T.A. A k-NN method for lung cancer prognosis with the use of a genetic algorithm for feature
selection. Expert Syst. Appl. 2021, 164, 113981. [CrossRef]
18. Abualigah, L.; Dulaimi, A.J. A novel feature selection method for data mining tasks using hybrid sine cosine algorithm and
genetic algorithm. Clust. Comput. 2021, 24, 2161–2176. [CrossRef]
19. Lappas, P.Z.; Yannacopoulos, A.N. A machine learning approach combining expert knowledge with genetic algorithms in feature
selection for credit risk assessment. Appl. Soft Comput. 2021, 107, 107391. [CrossRef]
20. Sathiyabhama, B.; Kumar, S.U.; Jayanthi, J.; Sathiya, T.; Ilavarasi, A.; Yuvarajan, V.; Gopikrishna, K. A novel feature selection
framework based on grey wolf optimizer for mammogram image analysis. Neural Comput. Appl. 2021, 33, 14583–14602. [CrossRef]
21. Rajammal, R.R.; Mirjalili, S.; Ekambaram, G.; Palanisamy, N. Binary Grey Wolf Optimizer with Mutation and Adaptive K-nearest
Neighbour for Feature Selection in Parkinson’s Disease Diagnosis. Knowl. Based Syst. 2022, 246, 108701. [CrossRef]
22. Preeti, P.; Deep, K. A Random Walk Grey Wolf Optimizer based on dispersion factor for feature selection on Chronic Disease
Prediction. Expert Syst. Appl. 2022, 206, 117864. [CrossRef]
23. Mohammadzadeh, H.; Gharehchopogh, F.S. A novel hybrid whale optimization algorithm with flower pollination algorithm for
feature selection: Case study Email spam detection. Comput. Intell. 2021, 37, 176–209. [CrossRef]
24. Hassouneh, Y.; Turabieh, H.; Thaher, T.; Tumar, I.; Chantar, H.; Too, J. Boosted whale optimization algorithm with natural
selection operators for software fault prediction. IEEE Access 2021, 9, 14239–14258. [CrossRef]
25. Moorthy, U.; Gandhi, U.D. A novel optimal feature selection technique for medical data classification using ANOVA based whale
optimization. J. Ambient Intell. Humaniz. Comput. 2021, 12, 3527–3538. [CrossRef]
26. Shafi, A.; Molla, M.; Jui, J.J.; Rahman, M.M. Detection of colon cancer based on microarray dataset using machine learning as a
feature selection and classification techniques. SN Appl. Sci. 2020, 2, 1243. [CrossRef]
27. Agrawal, V.; Chandra, S. Feature selection using Artificial Bee Colony algorithm for medical image classification. In Proceedings
of the 2015 Eighth International Conference on Contemporary Computing (IC3), Noida, India, 20–22 August 2015; pp. 171–176.
Biosensors 2022, 12, 821 20 of 21

28. Rani, M.; Gagandeep. Employing Artificial Bee Colony Algorithm for Feature Selection in Intrusion Detection System. In
Proceedings of the 2021 8th International Conference on Computing for Sustainable Global Development (INDIACom), New
Delhi, India, 17–19 March 2021; pp. 496–500.
29. Sharkawy, R.; Ibrahim, K.; Salama, M.; Bartnikas, R. Particle swarm optimization feature selection for the classification of
conducting particles in transformer oil. IEEE Trans. Dielectr. Electr. Insul. 2011, 18, 1897–1907. [CrossRef]
30. Sakri, S.B.; Rashid, N.B.A.; Zain, Z.M. Particle swarm optimization feature selection for breast cancer recurrence prediction. IEEE
Access 2018, 6, 29637–29647. [CrossRef]
31. Kunhare, N.; Tiwari, R.; Dhar, J. Particle swarm optimization and feature selection for intrusion detection system. Sādhanā 2020,
45, 109. [CrossRef]
32. Joshi, G.; Walambe, R.; Kotecha, K. A review on explainability in multimodal deep neural nets. IEEE Access 2021, 9, 59800–59821.
[CrossRef]
33. Deotale, D.; Verma, M.; Suresh, P.; Kotecha, K. Optimized hybrid RNN model for human activity recognition in untrimmed
video. J. Electron. Imaging 2022, 31, 051409. [CrossRef]
34. Wang, K.; He, J.; Zhang, L. Attention-based convolutional neural network for weakly labeled human activities’ recognition with
wearable sensors. IEEE Sens. J. 2019, 19, 7598–7604. [CrossRef]
35. Xia, K.; Huang, J.; Wang, H. LSTM-CNN architecture for human activity recognition. IEEE Access 2020, 8, 56855–56866. [CrossRef]
36. Sikder, N.; Ahad, M.A.R.; Nahid, A.A. Human Action Recognition Based on a Sequential Deep Learning Model. In Proceedings
of the 2021 Joint 10th International Conference on Informatics, Electronics & Vision (ICIEV) and 2021 5th International Conference
on Imaging, Vision & Pattern Recognition (icIVPR), Krakow, Poland, 16–18 June 2021; pp. 1–7.
37. Kumar, P.; Suresh, S. DeepTransHHAR: Inter-subjects Heterogeneous Activity Recognition Approach in the Non-identical
Environment Using Wearable Sensors. Natl. Acad. Sci. Lett. 2022, 45, 317–323. [CrossRef]
38. Dua, N.; Singh, S.N.; Semwal, V.B. Multi-input CNN-GRU based human activity recognition using wearable sensors. Computing
2021, 103, 1461–1478. [CrossRef]
39. Khatun, M.A.; Yousuf, M.A.; Ahmed, S.; Uddin, M.Z.; Alyami, S.A.; Al-Ashhab, S.; Akhdar, H.F.; Khan, A.; Azad, A.; Moni, M.A.
Deep CNN-LSTM with Self-Attention Model for Human Activity Recognition using Wearable Sensor. IEEE J. Transl. Eng. Health
Med. 2022, 10, 2700316. [CrossRef] [PubMed]
40. Ghate, V.; Hemalatha, S. Hybrid deep learning approaches for smartphone sensor-based human activity recognition. Multimed.
Tools Appl. 2021, 80, 35585–35604. [CrossRef]
41. Ronald, M.; Poulose, A.; Han, D.S. iSPLInception: An inception-ResNet deep learning architecture for human activity recognition.
IEEE Access 2021, 9, 68985–69001. [CrossRef]
42. Tufek, N.; Yalcin, M.; Altintas, M.; Kalaoglu, F.; Li, Y.; Bahadir, S.K. Human action recognition using deep learning methods on
limited sensory data. IEEE Sens. J. 2019, 20, 3101–3112. [CrossRef]
43. Gao, W.; Zhang, L.; Huang, W.; Min, F.; He, J.; Song, A. Deep neural networks for sensor-based human activity recognition using
selective kernel convolution. IEEE Trans. Instrum. Meas. 2021, 70, 2512313. [CrossRef]
44. Huang, W.; Zhang, L.; Teng, Q.; Song, C.; He, J. The convolutional neural networks training with channel-selectivity for human
activity recognition based on sensors. IEEE J. Biomed. Health Inform. 2021, 25, 3834–3843. [CrossRef] [PubMed]
45. Gao, W.; Zhang, L.; Teng, Q.; He, J.; Wu, H. DanHAR: Dual attention network for multimodal human activity recognition using
wearable sensors. Appl. Soft Comput. 2021, 111, 107728. [CrossRef]
46. Tang, Y.; Zhang, L.; Min, F.; He, J. Multi-scale deep feature learning for human activity recognition using wearable sensors. IEEE
Trans. Ind. Electron. 2022. [CrossRef]
47. Berlin, S.J.; John, M. Particle swarm optimization with deep learning for human action recognition. Multimed. Tools Appl. 2020,
79, 17349–17371. [CrossRef]
48. Zhang, R. Sports action recognition based on particle swarm optimization neural networks. Wirel. Commun. Mob. Comput. 2022,
2022, 6912315. [CrossRef]
49. Guha, R.; Khan, A.H.; Singh, P.K.; Sarkar, R.; Bhattacharjee, D. CGA: A new feature selection model for visual human action
recognition. Neural Comput. Appl. 2021, 33, 5267–5286. [CrossRef]
50. Helmi, A.M.; Al-Qaness, M.A.; Dahou, A.; Damaševičius, R.; Krilavičius, T.; Elaziz, M.A. A novel hybrid gradient-based optimizer
and grey wolf optimizer feature selection method for human activity recognition using smartphone sensors. Entropy 2021,
23, 1065. [CrossRef]
51. Sikder, N.; Nahid, A.A. KU-HAR: An open dataset for heterogeneous human activity recognition. Pattern Recognit. Lett. 2021,
146, 46–54. [CrossRef]
52. Chavarriaga, R.; Sagha, H.; Calatroni, A.; Digumarti, S.T.; Tröster, G.; Millán, J.D.R.; Roggen, D. The Opportunity challenge: A
benchmark database for on-body sensor-based activity recognition. Pattern Recognit. Lett. 2013, 34, 2033–2042. [CrossRef]
53. Reiss, A.; Stricker, D. Introducing a new benchmarked dataset for activity monitoring. In Proceedings of the 2012 16th
International Symposium on Wearable Computers, Newcastle, UK, 18–22 June 2012, pp. 108–109.
54. Sucerquia, A.; López, J.D.; Vargas-Bonilla, J.F. SisFall: A fall and movement dataset. Sensors 2017, 17, 198. [CrossRef]
55. Anguita, D.; Ghio, A.; Oneto, L.; Parra, X.; Reyes-Ortiz, J.L. A public domain dataset for human activity recognition using
smartphones. In Proceedings of the European Symposium on Artificial Neural Networks, Computational Intelligence and
Machine Learning, Bruges, Belgium, 24–26 April 2013.
Biosensors 2022, 12, 821 21 of 21

56. Micucci, D.; Mobilio, M.; Napoletano, P. Unimib shar: A dataset for human activity recognition using acceleration data from
smartphones. Appl. Sci. 2017, 7, 1101. [CrossRef]
57. Weiss, G.M.; Lockhart, J. The impact of personalization on smartphone-based activity recognition. In Proceedings of the
Workshops at the Twenty-Sixth AAAI Conference on Artificial Intelligence, Toronto, ON, Canada, 22–26 July 2012.
58. Abualigah, L.; Yousri, D.; Abd Elaziz, M.; Ewees, A.A.; Al-Qaness, M.A.; Gandomi, A.H. Aquila optimizer: A novel meta-heuristic
optimization algorithm. Comput. Ind. Eng. 2021, 157, 107250. [CrossRef]
59. Abualigah, L.; Diabat, A.; Mirjalili, S.; Abd Elaziz, M.; Gandomi, A.H. The arithmetic optimization algorithm. Comput. Methods
Appl. Mech. Eng. 2021, 376, 113609. [CrossRef]
60. Faramarzi, A.; Heidarinejad, M.; Mirjalili, S.; Gandomi, A.H. Marine Predators Algorithm: A nature-inspired metaheuristic.
Expert Syst. Appl. 2020, 152, 113377. [CrossRef]
61. Li, S.; Chen, H.; Wang, M.; Heidari, A.A.; Mirjalili, S. Slime mould algorithm: A new method for stochastic optimization. Future
Gener. Comput. Syst. 2020, 111, 300–323. [CrossRef]
62. Mirjalili, S.; Lewis, A. The whale optimization algorithm. Adv. Eng. Softw. 2016, 95, 51–67. [CrossRef]
63. Karaboga, D. An Idea Based on Honey Bee Swarm for Numerical Optimization; Technical Report-tr06; Erciyes University: Kayseri,
Turkey, 2005.
64. Tereshko, V.; Loengarov, A. Collective decision making in honey-bee foraging dynamics. Comput. Inf. Syst. 2005, 9, 1.
65. Karaboga, D.; Basturk, B. A powerful and efficient algorithm for numerical function optimization: Artificial bee colony (ABC)
algorithm. J. Glob. Optim. 2007, 39, 459–471. [CrossRef]
66. Mirjalili, S.; Mirjalili, S.M.; Lewis, A. Grey wolf optimizer. Adv. Eng. Softw. 2014, 69, 46–61. [CrossRef]
67. Goldberg, D.E.; Richardson, J. Genetic algorithms with sharing for multimodal function optimization. In Proceedings of the
Second International Conference on Genetic Algorithms, Cambridge, MA, USA, 28–31 July 1987; Volume 4149.
68. Goldberg, D.E. Genetic Algorithms; Pearson Education India: Bengaluru, India , 2013.
69. Kennedy, J.; Eberhart, R. Particle swarm optimization. In Proceedings of the ICNN’95-International Conference on Neural
Networks, Perth, WA, Australia, 27 November–1 December 1995; Volume 4, pp. 1942–1948.
70. Preece, S.J.; Goulermas, J.Y.; Kenney, L.P.; Howard, D. A comparison of feature extraction methods for the classification of
dynamic activities from accelerometer data. IEEE Trans. Biomed. Eng. 2008, 56, 871–879. [CrossRef] [PubMed]
71. Shoaib, M.; Bosch, S.; Incel, O.D.; Scholten, H.; Havinga, P.J. Complex human activity recognition using smartphone and
wrist-worn motion sensors. Sensors 2016, 16, 426. [CrossRef]
72. Voicu, R.A.; Dobre, C.; Bajenaru, L.; Ciobanu, R.I. Human physical activity recognition using smartphone sensors. Sensors 2019,
19, 458. [CrossRef]
73. Teng, Q.; Wang, K.; Zhang, L.; He, J. The layer-wise training convolutional neural networks using local loss for sensor-based
human activity recognition. IEEE Sens. J. 2020, 20, 7265–7274. [CrossRef]
74. Ma, H.; Li, W.; Zhang, X.; Gao, S.; Lu, S. AttnSense: Multi-level Attention Mechanism For Multimodal Human Activity
Recognition. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, Macao, China,
10–16 August 2019; pp. 3109–3115.
75. Sousa Lima, W.; Souto, E.; El-Khatib, K.; Jalali, R.; Gama, J. Human activity recognition using inertial sensors in a smartphone:
An overview. Sensors 2019, 19, 3213. [CrossRef]
76. Anguita, D.; Ghio, A.; Oneto, L.; Parra, X.; Reyes-Ortiz, J.L. Human activity recognition on smartphones using a multiclass
hardware-friendly support vector machine. In Proceedings of the International Workshop on Ambient Assisted Living,
Vitoria-Gasteiz, Spain, 3–5 December 2012; pp. 216–223.
77. Cristianini, N.; Shawe-Taylor, J. An Introduction to Support Vector Machines and Other Kernel-Based Learning Methods; Cambridge
University Press: Cambridge, UK, 2000.
78. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
79. Mrozek, D.; Koczur, A.; Małysiak-Mrozek, B. Fall detection in older adults with mobile IoT devices and machine learning in the
cloud and on the edge. Inf. Sci. 2020, 537, 132–147. [CrossRef]
80. Al-qaness, M.A.; Dahou, A.; Abd Elaziz, M.; Helmi, A. Multi-ResAtt: Multilevel Residual Network with Attention for Human
Activity Recognition Using Wearable Sensors. IEEE Trans. Ind. Inform. 2022. [CrossRef]
81. Zhang, Y.; Zhang, Z.; Zhang, Y.; Bao, J.; Zhang, Y.; Deng, H. Human activity recognition based on motion sensor using u-net.
IEEE Access 2019, 7, 75213–75226. [CrossRef]
82. Li, X.; Dezert, J.; Khyam, M.O.; Noor-A-Rahim, M.; Ge, S.S.; Dezert, J.; Dong, Y. DSmT-Based Fusion Strategy for Human Activity
Recognition in Body Sensor Networks. IEEE Trans. Ind. Inform. 2020, 16, 7138–7149.
83. Zhang, H.; Xiao, Z.; Wang, J.; Li, F.; Szczerbicki, E. A novel IoT-perceptive human activity recognition (HAR) approach using
multihead convolutional attention. IEEE Internet Things J. 2019, 7, 1072–1080. [CrossRef]
84. Teng, Q.; Zhang, L.; Tang, Y.; Song, S.; Wang, X.; He, J. Block-wise training residual networks on multi-channel time series for
human activity recognition. IEEE Sens. J. 2021, 21, 18063–18074. [CrossRef]