Content uploaded by Ibrahim Karatas
Author content
All content in this area was uploaded by Ibrahim Karatas on Jan 22, 2025
Content may be subject to copyright.
Research article
Deep learning-based system for prediction of work at height in
construction site
Ibrahim Karatas
Osmaniye Korkut Ata University, Faculty of Engineering and Natural Sciences, Department of Civil Engineering, Osmaniye, Turkey
ARTICLE INFO
Keywords:
Construction safety
Fall from height
Deep learning
Construction management
ABSTRACT
Falling from height (FFH) is a major cause of injuries and fatalities on construction sites. Research
has emphasized the role of technological advances in managing FFH safety risks. In this inves-
tigation, the objective is to forecast if a laborer is operating at an elevated position by utilizing an
accelerometer, gyroscope, and pressure information through the application of deep-learning
techniques. The study involved analyzing worker data to quickly implement safety measures
for working at heights. A total of 45 analyses were conducted using DNN, CNN, and LSTM deep-
learning models, with 5 different window sizes and 3 different overlap rates. The analysis
revealed that the DNN model, utilizing a 1-s window size and a 75 % overlap rate, attained an
accuracy of 94.6 % with a loss of 0.1445. Conversely, the CNN model, employing a 5-s window
size and a 75 % overlap rate, demonstrated an accuracy of 94.9 % with a loss of 0.1696. The
results of this study address information gaps by efciently predicting workers’ working condi-
tions at heights without the need for complex calculations. By implementing this method at
construction sites, it is expected to reduce the risk of FFH and align occupational health and safety
practices with technological advancements.
1. Introduction
Construction sites have always been known as one of the most hazardous industries due to the unique nature of outdoor operations,
working at heights, complex facilities and equipment on site, and employee attitudes and behaviors towards safety [1,2]. Construction
projects not only have hazardous working environments but also subject workers to intense physical exertion and serious safety and
health risks, all leading to an increase in fatal and non-fatal accidents [3]. Globally, the construction sector employs approximately 29
% of all industrial workers, yet it has the highest rate of occupational accidents among all sectors at 40 % [4]. In the last 20 years, over
28,000 workers have died in the construction industry, with approximately 200,000 non-fatal injuries reported annually [5]. The most
common accidents in the construction industry include falls from heights (FFH), collisions with objects, electrocution, and being
trapped [6]. Moreover, there has been an increase in industrial accidents in South Korea, with accidents at construction sites making up
over one-third of all industrial accidents. Falls account for 47.7 %–52.1 % of the total number of fatalities in the construction industry
[6]. In addition to the United States and South Korea, other countries such as Australia, China, and Turkey have experienced signicant
economic and human losses due to FFH [2,7,8]. Given the presence of elevated work areas such as ladders, scaffolding, and roofs,
construction workers in such environments are at a heightened risk of injury and fatality. Previous studies have emphasized the
signicance of proactive and preventive measures, including hazard identication, control, safety training, and the prompt recognition
E-mail address: ibrahimkaratas@osmaniye.edu.tr.
Contents lists available at ScienceDirect
Heliyon
journal homepage: www.cell.com/heliyon
https://doi.org/10.1016/j.heliyon.2025.e41779
Received 9 September 2024; Received in revised form 6 January 2025; Accepted 7 January 2025
Heliyon 11 (2025) e41779
Available online 17 January 2025
2405-8440/© 2025 The Author. Published by Elsevier Ltd. This is an open access article under the CC BY license
( http://creativecommons.org/licenses/by/4.0/ ).
of fall hazards on construction sites, as crucial in mitigating these risks [2,6,8,9]. In the context of FFH accidents, it is essential to
identify and address the fall hazards present at construction sites and communicate these situations to experts in order to minimize
such incidents. While safety training programs have been implemented to mitigate risks, they have only been partially effective in
reducing the number of falls at construction sites [8,10]. Furthermore, safety equipment like seat belts has been employed to mitigate
the risk of FFHs. Nonetheless, although conventional measures such as seat belts are effective in minimizing the impact of risks,
documented incidents of FFH reveal that traditional methods are inadequate in addressing the diverse fall hazards present in today’s
dynamic work settings [11]. Hence, it is imperative to employ cutting-edge technologies in order to mitigate these incidents and
enhance the well-being and safety of laborers [8,12–15].
Numerous new technologies have been developed and researched in the literature to identify and mitigate ongoing fall accidents at
construction sites. These include integrating building information modeling (BIM) systems with safety performance to improve hazard
detection and communication [1,3,16,17], using a camera or closed-circuit television system to monitor construction activities
[18–20] and employing wearable technologies to recognize unsafe behavioral activities and improve safety management [2,3,21–23].
Effective occupational health and safety measures rely on three fundamental pillars: prediction, prevention, and mitigation. These
pillars form the basis for enhancing safety risk management procedures, judgments, and results through the utilization of suitable
technologies [24]. Thus, the primary objective of this research is to propose a methodology for forecasting the occurrence of elevated
work among workers using data collected through sensors. This approach involves implementing preventive measures based on the
generated forecasts, ultimately leading to a reduction in incidents of FFH.
Contemporary technological advancements possess the capability to adequately mitigate the hazard of falling from elevated sur-
faces. Appropriate utilization of technological tools in preempting fall hazards within the realm of occupational health and safety in
construction environments has the potential to enhance the accuracy of risk anticipation [24]. Conversely, the signicant expenses
associated with the hardware and software components of contemporary technologies, the necessity of providing training for staff
members on the utilization of such technologies, and the impact of wearable devices on the limited movement of employees are
identied as obstacles in this context [25,26]. However, in spite of these obstacles, the utilization of emerging technologies presents an
opportunity for a proactive strategy in the anticipation and management of fall hazards [26].
In the realm of occupational health and safety, extensive research has been carried out regarding the utilization of novel tech-
nologies for the anticipation and mitigation of fall incidents from elevated positions. Fang et al. (2019) introduced an automated
computer vision methodology employing Mask Region Based Convolutional Neural Network (R-CNN) for the identication of hazards
among construction site personnel. The study revealed recall and precision rates of 90 % and 75 % respectively for hazard identi-
cation using this particular approach. The authors suggested that the computer vision system could be implemented by construction
site supervisors to autonomously identify unsafe actions and offer guidance to individuals regarding their risk of experiencing an FFH
[27]. Anjum et al. (2022) introduced a deep learning-based method for assessing height. This method uses a single known value in an
image to measure the working height of workers, monitor compliance with safety regulations, and ensure worker safety. The study
analyzed over 300 images for binary classication (safe and unsafe) and achieved an overall accuracy of 85.33 % [6]. In a study
concerning the monitoring of workers, Hong et al. (2023) introduced a framework aimed at categorizing the predetermined safety
actions of scaffold workers and assessing the fulllment of each safety action, with the purpose of enhancing safety management in
response to the challenges associated with manual observations. To achieve this objective, they developed a Convolutional Neural
Network (CNN) model utilizing Gramian angular elds (GAFs) to categorize adherence to safety protocols among 35 scaffold workers
through the utilization of ve Inertial Measurement Unit (IMU) sensors. The outcomes of the model indicate that the proposed
approach has the potential to automatically discern whether workers are adhering to safety regulations or not [28]. Automatic
detection of fall hazards using IMU sensors, including accelerometer and gyroscope data, has been identied as an effective approach
to monitoring fall hazards from height on construction sites [29]. In this way, it is possible to take timely preventive measures against
dangerous situations. Additionally, these wearable sensors can help prevent FFH by enabling advanced precautions.
Choo et al. (2023) have proposed a system that utilizes wearable sensors to identify workers operating at heights and determine the
anchoring status of safety hooks in order to prevent falls. The IMU sensor collects data, which is then processed using machine learning
algorithms to assess the status of safety hooks for workers at heights. Despite previous efforts to identify unsafe working conditions and
behaviors at heights, researchers have encountered challenges due to the complexity of tasks and dynamic working conditions, which
have hindered the establishment of precise methodologies for effective and timely detection. The evaluation of the constructed model
was conducted through single-subject cross-validation (LOSOCV) in order to accommodate various new workers and working envi-
ronments. As per the ndings, the system for identifying work at height demonstrated an accuracy of 96 %, whereas the system for
detecting safety hook attachments exhibited 86 % accuracy [2]. In addition, in the work on the height detection system by Choo et al.
(2023), a rule-based height estimation was made by applying the Kalman Filter and Complementary Filter algorithms to the collected
data. In addition, it was determined whether the worker was only working at height or not.
In the evaluation of construction studies, there is a notable interest in leveraging technology to enhance occupational health and
safety practices. It is evident that the adoption of technology at construction sites substantially reduces accidents and injuries, thereby
cultivating a safer work environment. Specically, there is a heightened focus on addressing FFH incidents for the overall well-being of
workers. Extensive research has delved into identifying fall hazards in the construction industry. Specically, computer vision-based
studies have been employed for this purpose. However, the analysis process can be time-consuming and demanding due to the high
processing requirements of image processing stages. Furthermore, existing studies have primarily focused on identifying and pre-
dicting two scenarios: whether workers are operating at elevated heights or not. The dataset employed in this research comprises the
information gathered by Choo et al. (2023) in their study. In contrast to their work, the objective is to develop a deep learning-based
prediction model specically for detecting work at heights. Furthermore, the aim is to predict not only work at heights in general but
I. Karatas
Heliyon 11 (2025) e41779
2
also work at heights on the ground, on ladders, on mobile scaffolds, and horse scaffolds separately. In the context of creating deep
learning models, it is imperative that the data obtained from sensors adheres to specic window sizes and overlap ratios. To achieve
this, the sliding window technique is commonly employed. This methodology is frequently employed in the realm of human activity
recognition challenges. It is specically utilized in the initial phases of data segmentation to process unrened data for subsequent
modeling purposes. The popularity of this method stems from its simplicity of implementation and its exemption from preprocessing
requirements. Particularly advantageous for applications requiring instantaneous responses, this method presents an optimal solution.
The process involves segmenting the incoming sensor data into windows of a consistent size, with no intervals between them [30]. In
addition to the dimensions of the window, the percentage of overlap plays a signicant role in the context of activity recognition tasks
when windows traverse one another. Prior research has employed window durations varying from 0.1 to 8 s along with diverse overlap
percentages during the process of segmenting data [30–38]. One of the aims of this study was to determine the optimal window size
and overlap ratio for predicting the analyzed data. The window sizes tested were 1s, 2s, 3s, 4s, and 5s, with overlap ratios of 25 %, 50
%, and 75 %. This was necessary to divide the data collected at 25Hz fully and achieve the highest prediction success.
The main contributions of our study are as follows.
•By utilizing data from accelerometers, gyroscopes, and pressure sensors gathered from individuals in a workplace setting, one can
potentially infer the worker’s occupational conditions and identify the necessary safety measures for tasks performed at elevated
locations.
•In exploring this study, it has been aimed to ascertain the most precise model through a comparative analysis of various deep-
learning approaches for predicting working at height situations.
•Analyzing various window sizes and overlap ratios for analysis, along with different deep learning methods, to determine the
optimal window size and overlap ratios for the most accurate model.
Fig. 1. Research Methodology owchart.
I. Karatas
Heliyon 11 (2025) e41779
3
1.1. Knowledge gaps
In academic literature, researchers have utilized accelerometers and gyroscopes to assess workers’ movements and their impact on
efciency at construction sites. However, there is limited research on the use of these sensors in studying occupational health and
safety, particularly in relation to falls from height. Furthermore, image-processing techniques have also been employed in studies
pertaining to occupational health and safety. This study emphasizes the effectiveness of utilizing kinematic-based sensors to navigate
the complexities of construction site environments. These sensors are both lightweight and wireless, which enhances their practicality
for real-world applications. Also, in contrast to most similar research that relies on traditional machine learning techniques, our
approach employs advanced deep learning methods to improve accuracy rates signicantly. The proposed deep learning framework
allows for rapid predictions without necessitating complex calculations for operations at heights, and it also facilitates swift preventive
measures against falls from heights (FFH). A range of deep learning models was evaluated to identify the one with the highest accuracy.
Furthermore, this research explores various options to establish the optimal window size and overlap ratio during the data segmen-
tation phase, which is crucial for determining the best data preprocessing techniques before model development. These investigations
set this study apart from previous research in height level estimation.
2. Research methodology
This research aims to predict the height levels of construction workers by analyzing data collected from accelerometers, gyro-
scopes, and barometers using advanced deep-learning techniques. Additionally, it seeks to identify the approach that provides the
highest accuracy in these predictions by evaluating various window sizes and overlap ratios within the deep learning methods
employed for height level estimation. The objective is to accurately forecast the working height levels of workers, while also exploring
the feasibility of a framework designed to monitor their safety. This framework will investigate whether the working heights of
construction workers can be effectively estimated through the combination of motion sensor data and barometric data. This study is
grounded in two primary hypotheses: (1) to explore whether distinctive features derived from motion data—such as accelerometer,
gyroscope, and barometric readings—can be effectively used to estimate a worker’s height level, and (2) to determine the optimal
window size and overlap ratio for analysis using a deep learning method that achieves the highest accuracy in estimating the height
level of construction workers. A schematic representation of the research methodology, aligned with this framework, is illustrated in
Fig. 1.
The initial step involves preparing the data obtained from the literature for the model. The data was labeled according to the
method described in Choo et al. (2023) [2]. Following data preprocessing and segmentation, the dataset has been split into training
and test sets employing 5 distinct window sizes and 3 different overlap rates. The performance of the training data has been compared
with that of the test data to assess prediction accuracy. Subsequently, 5 different deep learning models have been applied to analyze the
data, and model evaluation metrics have been utilized to identify the most suitable model.
A. Data Preparation and Data Segmentation
The study utilized data from Choo et al. (2023) [2] and involved 20 workers in a construction site setting. The data was gathered
from sensors located at three different positions: base, belt, and hook. However, only the data from the belt sensors was used in this
study to predict working at heights. The area labeled as the "base" refers to the sensor mounted on the oor, while the "hook" designates
the sensor attached to the end of the worker’s safety hook. The term "belt" indicates the sensor positioned on the safety belt worn by the
worker. Given that the worker consistently wears a seat belt, the data gathered from the sensor afxed to the belt is utilized in this
research. Each sensor captured accelerometer, gyroscope, and barometer data. In this study, data consisting of atmospheric pressure
signals from the barometer, as well as acceleration and gyroscopic signals from the Inertial Measurement Unit (IMU) sensor, were
recorded at a frequency of 8 Hz. The descriptive statistics for this dataset, comprising a total of 368647 entries, are presented in
Table 1. This study aims to assess the utilization of scaffolds by workers during their work activities. Subsequently, based on this
assessment, an evaluation of their work performance at elevated heights has been conducted. The method employed in the research
article facilitated the identication of the specic type of scaffold (such as mobile scaffold, horse scaffold, ladder, or ground) on which
the monitored workers were operating, as illustrated in Fig. 2. The collected data was subsequently labeled. Following an initial
analysis, outlier values were eliminated using the median absolute deviation (MAD) method.
Following the collection of raw data, data segmentation was conducted to prepare it for the model. This process entailed dividing
Table 1
Descriptive statistics for the dataset.
Pressure Acc_x Acc_y Acc_z Gyr_x Gyr_y Gyr_z
Mean 1017,10 1186 2726 9142 0,038 −0,056 −0,017
Std 5,33 2260 1486 0,927 0,203 0,269 0,506
Min 1008,88 −19,321 −13,195 −17,339 −7341 −12,067 −4889
25 % 1010,92 0,464 1853 8867 −0,043 −0,161 −0,205
50 % 1017,88 1714 2833 9251 0,038 −0,056 −0,018
75 % 1020,69 2623 3651 9596 0,117 0,051 0,167
Max 1026,39 36,982 14,161 24,309 1847 5942 5118
I. Karatas
Heliyon 11 (2025) e41779
4
the data into windows of specic lengths and shifting them over each other with a predetermined overlap rate until the nal data was
ready for the model. Window sizes ranging from 0.1 to 10 s, as depicted in Fig. 3, have been identied as appropriate for activity
recognition based on earlier investigations. The overlap rate among pennies in past research typically ranged from 25 % to 75 % [30,
39–41]. In this research, data was gathered at a frequency of 8 Hz, with overlap rates of 25 %, 50 %, and 75 % being chosen.
Consequently, window sizes of 1 s, 2 s, 3 s, 4 s, and 5 s were determined for complete separation of the data. Following this process, the
resulting segmented data was prepared for model input. After segmentation, the data was split into 80 % for training and 20 % for
testing. Three deep learning techniques (DNN, CNN, LSTM) were employed for activity classication in this study.
B. Deep Learning Models
In the domain of machine learning models, the process of statistical feature extraction plays a pivotal role in the examination phase.
Nevertheless, this procedure may consume considerable time and resources. Conversely, deep learning techniques present a more
effective strategy for feature extraction. These techniques scrutinize motion data instead of statistical characteristics, leading to a
substantial decrease in the time and resources necessary for examination. Through the utilization of deep learning techniques, or-
ganizations and educational entities can optimize their examination procedures and attain more precise outcomes in a shorter time
frame.
2.1. Deep neural network (DNN)
Neural networks usually consist of a restricted number of hidden layers. In contrast, a Deep Neural Network (DNN) is essentially a
standard neural network with added "depth". The depth of a neural network is dened by the number of hidden layers positioned
between the input and output layers. DNNs are more adept at processing large datasets owing to their greater number of layers.
Moreover, DNNs are frequently employed as the dense layer in other deep-learning models [42]. DNN architectures are typically
structured with a minimum of three fully connected layers, comprising an initial input layer, multiple intermediate hidden layers, and
a nal output layer. Within a fully connected network, individual neurons within a given layer are interconnected with all neurons
from the preceding layer, facilitating comprehensive information exchange. The outputs generated by the hidden neurons are inter-
preted as a collection of distinct features derived from the input data specic to their respective layer. Successive layering within the
network can be viewed as a process wherein features are abstracted and extracted at increasingly higher levels. Neurons situated in the
nth layer are responsible for deriving features based on computations involving those originating from the (n-1)th layer [43]. The DNN
architecture utilized in this study is depicted in Fig. 4. This architecture comprises 6 hidden layers. To enhance the training and
generalization of the data in each network layer, batch normalization and dropout methods were implemented. Batch normalization
standardizes the inputs of each layer using a xed scale across the layer, ensuring that the outputs of the layers remain stable during the
later stages of the learning process. Conversely, dropout deactivates a random set of neurons during each training step, aiding in
preventing overtting by reducing the model’s reliance on each neuron. In this study, the dropout value is set to 0.2.
2.2. Convolutional neural network (CNN)
CNN is a form of deep learning model extensively employed in tasks such as pattern recognition and image analysis. Its application
has been prevalent within the domain of occupational health and safety, particularly within the context of construction sites [6]. It
Fig. 2. Types of scaffolding for which data was collected [2].
I. Karatas
Heliyon 11 (2025) e41779
5
consists of several layers that can extract signicant features from input data. CNNs are highly efcient in automatically extracting
spatial information from data, especially two-dimensional image data. They are not as effective in processing one-dimensional data,
such as physical or business data [44]. A common CNN architecture usually comprises a Convolutional layer, a Pooling layer, and a
Fully connected layer. The Convolutional layer within a CNN employs lters to derive characteristics from the input data, with each
feature necessitating an individual lter. The action of moving these lters across the input data is referred to as convolution. Within
the CNN design, there are segments, each encompassing a convolutional layer with ReLU activation alongside a maximum max-pooling
layer [44]. Furthermore, these lters ensure the preservation of distinct features extracted from individual sensor channels prior to
their propagation to the subsequent network layers. The connections of the convolutional layers are established with the successive
layers, in contrast to the fully interconnected structure seen in traditional neural network architectures. Also, for the purpose of
classication, it is common practice to append a fully connected layer after the nal block in order to consolidate the information
gathered from all sensor channels. Subsequently, class probabilities are derived by incorporating a softmax layer at the conclusion of
Fig. 3. Data segmentation process.
Fig. 4. DNN model architecture.
I. Karatas
Heliyon 11 (2025) e41779
6
the network [43]. In certain cases, CNN may be more efcient than DNN due to its use of weight sharing and subsampling to decrease
the number of weights and connections. The CNN architecture employed in this investigation is depicted in Fig. 5.
2.3. Long short-term memory (LSTM)
The LSTM model is a form of deep learning that excels in solving complex problems. Specically, it is a widely successful variant of
Recurrent Neural Networks (RNN), featuring layers with LSTM cells capable of storing information over time in internal memory.
LSTM networks are adept at capturing temporal dependencies across various applications, including automatic translation, image
captioning, and sensor or video-based event recognition [43,44]. Several studies have employed the LSTM model in the context of
activity recognition [33,42,44]. LSTM cells are specically engineered to retain and recall information over extended periods by
retaining it within an internal memory. These cells are capable of updating, outputting, or deleting this internal state based on their
inputs and the state from the preceding time step. LSTM cells are organized in hierarchical layers, akin to biological neurons. The
information produced by each cell is conveyed to the succeeding cell within the corresponding stratum and subsequently to the
subsequent stratum in the neural network. Upon reaching the nal stratum, the information is forwarded to the dense and softmax
strata to address the task of classication. The schematic in Fig. 6 depicts the fundamental composition of an LSTM cell, which
comprises the forgetting gate (red rectangular part), input gate (green rectangular part), and output gate (yellow rectangular part)
[45]. The architecture of the LSTM model used in this study is depicted in Fig. 7. It comprises 6 LSTM layers, with each layer containing
a specic number of LSTM cells. Following the LSTM layers, a dense layer utilizing the Relu activation function and an output layer
employing the softmax activation function were constructed.
In the eld of literature studies, various deep learning techniques have been employed for the purpose of recognizing human
activities and enhancing occupational health and safety protocols within construction sites as well as other industrial settings [6,31,
33–35,46–49]. In this study, the primary aim is to determine the most suitable model through the analysis of literature-derived data
using various deep learning methods and diverse window and overlap ratios. To achieve this objective, the analysis results have been
assessed using a range of evaluation metrics. Various metrics such as Accuracy, F1-score, and Loss values are considered in the
evaluation process. The metrics Accuracy and F1-score are determined through the application of specic formulae denoted by
Equations (1) and (2). Within the realm of deep learning classication tasks, Loss values are computed utilizing Equation (5). This
metric is crucial for assessing the model’s predictions against the ground truth. It quanties the disparity between the model’s pre-
dictions and the actual values, with a lower Loss value signifying a higher level of proximity between the predicted and actual values.
Fig. 5. CNN model architecture.
I. Karatas
Heliyon 11 (2025) e41779
7
Accuracy =TP +TN
TP +FP +TN +FN (1)
Precision =TP
TP +FP (2)
Recall =TP
TP +FN (3)
F1Score =2*Precision*Recall
Precision +Recall (4)
Loss = − ∑(y*log(p)) (5)
where y is the true value and p is the predicted value of the model.
3. Results and discussion
This study utilized deep learning methods to analyze acceleration, gyroscope, and pressure data, aiming to predict the working
conditions of workers on three different types of scaffolds as well as when they are on the ground. The analysis results allow us to
predict whether a worker is on a scaffold or on the ground, and if on a scaffold, which type it is (mobile scaffold, horse scaffold, or
ladder). This analysis has also facilitated an evaluation of working at heights. Additionally, it has been sought to identify the most
Fig. 6. The basic structure of the LSTM cell [45].
Fig. 7. LSTM model architecture.
I. Karatas
Heliyon 11 (2025) e41779
8
suitable model by creating data sets from different window sizes and overlap rates for each deep learning method. 15 data sets were
generated for each model, and a total of 45 analyses were conducted for the purpose of identifying the most appropriate model for
working at heights. The analysis results revealed the most suitable model, with Fig. 8 displaying the accuracy values obtained. The
models with a 75 % overlap rate generally exhibit higher accuracy compared to models with different overlap rates. The window
lengths vary for each model. Regarding the accuracy of the deep learning models, the DNN model demonstrates a higher accuracy rate
with a 1-s window size, while the CNN model shows a higher accuracy rate with other window sizes. To effectively evaluate the
prediction success of these models, it is important to consider the loss and f1 score values in addition to accuracy.
When analyzing the F1 score and Loss values, it is evident that the highest prediction success is achieved with 75 % overlap rates
(Figs. 9 and 10). Upon reviewing the prediction results of LSTM models, it is apparent that they exhibit the highest level of loss.
Additionally, CNN and DNN models demonstrate better prediction results compared to the LSTM model.
In the above gures, it has been observed that deep learning analysis yields higher accuracy rates at 75 % overlap rates. Therefore,
Table 2presents the analysis results at a 75 % overlap rate. It is evident that the CNN model generally outperforms the DNN model. The
most successful models were the DNN model with a 1s window size, achieving a prediction success of 94.6 %, and the CNN model with
a 5s window size, achieving a prediction success of 94.9 %.
Upon evaluating the models with the highest accuracy rates individually, it was found that the DNN model, analyzed with a 1-s
window size and a 75 % overlap rate, achieved a prediction accuracy of approximately 95 % for scaffold type. Fig. 11(a) presents
the confusion matrix of this analysis, while Fig. 11(b) displays the learning curve graph. The confusion matrix indicates that the model
more accurately predicts the worker’s working situations, including mobile scaffold, horse scaffold, ladder, and ground. Evaluating a
model’s performance can be effectively done by plotting its learning curve. These curves illustrate the relationship between accuracy
and the quantity of training examples. By graphing the training and cross-validation scores for a particular model across different
training dimensions, we can observe how the scores change as the number of training examples increases. Additionally, these curves
can help identify whether the model suffers from bias or variance. A signicant margin between training and validation accuracy could
indicate overtting of the model.
In the evaluation of the confusion matrix, the DNN model correctly predicted 10550 out of 10973 mobile scaffold data. However, it
had some confusion with the ladder class. Out of 9253 ladder data tested, 8814 were accurately predicted, but there were some in-
stances of confusion with the mobile scaffold and horse scaffold categories. Lastly, out of 9245 horse scaffold data, 8742 were correctly
predicted. The data set was also mixed up with ladder and ground information. Amongst the 7393 ground data points that were
analyzed, 6766 were accurately forecasted, often mistaken for mobile and horse scaffold data. However, based on the learning curve
graph, we can infer that the DNN model is not prone to overtting, as there is only a minimal difference between the training and
validation lines.
The CNN model was assessed using a 5-s window size and a 75 % overlap rate, achieving an accuracy of approximately 95 % in
predicting scaffold types. Fig. 12(a) illustrates the confusion matrix for this analysis, while Fig. 12(b) depicts the learning curve graph.
According to the confusion matrix, 92 % of the 2194 mobile scaffold data tested with the CNN model were correctly predicted, with 4
% misclassied as ground. For the 1851 ladder data tested, 93 % were accurately predicted. The analysis indicates that there is some
confusion between the ladder and horse scaffold categories. Among the 1849 instances of horse scaffold data, 92 % were accurately
classied, but 4 % of these instances were mistakenly identied as ladder or ground data. Similarly, 93 % of the 1479 ground data
Fig. 8. Accuracy values of the analysis results.
I. Karatas
Heliyon 11 (2025) e41779
9
Fig. 9. Loss values of the analysis results.
Fig. 10. F1 score values of the analysis results.
Table 2
Results of the 75 % overlap rate deep learning models.
Model Win Length Accuracy Loss F1-score
DNN 1s 0,9460 0,1445 0,9441
2s 0,9401 0,1660 0,9380
3s 0,9389 0,1770 0,9371
4s 0,9239 0,2503 0,9219
5s 0,9250 0,2498 0,9232
CNN 1s 0,9360 0,1627 0,9336
2s 0,9457 0,1510 0,9441
3s 0,9465 0,1618 0,9449
4s 0,9503 0,1718 0,9488
5s 0,9490 0,1696 0,9475
LSTM 1s 0,8987 0,2724 0,8955
2s 0,9259 0,2030 0,9238
3s 0,9345 0,2130 0,9321
4s 0,9334 0,2397 0,9309
5s 0,9377 0,2585 0,9357
I. Karatas
Heliyon 11 (2025) e41779
10
samples were correctly predicted, with 3 % being misclassied as mobile or horse scaffold data. Despite the greater separation between
the training and validation lines on the learning curve graph compared to the DNN model, it is apparent that the CNN model is not
overtting.
In this study, a proposed method aims to determine whether construction site workers are working on scaffolding or on the ground.
If the workers are on scaffolding, the method seeks to identify the specic scaffolding they are using. The study evaluates various
window sizes and overlap ratios during model analysis to establish the optimal settings for the model. Based on the obtained results,
the study aims to assess whether the workers are working at height and if they have taken the necessary precautions for working at that
height. The dataset utilized in this study was sourced from Choo et al. (2023) [2]. The study involved determining height by employing
pressure data and Kalman and complementary lters, resulting in signicant computational overhead. Notably, the analysis focused
solely on ascertaining whether the workers were operating at height. Furthermore, deep learning techniques were applied to forecast
the specic scaffold on which they worked. The process reduced the crowd and employed a quicker method to determine whether the
worker was working at height or not, with an accuracy rate of 95 %. In addition to pressure data, this study utilized accelerometer and
gyroscope data for height estimation, considering the unique movements of workers on each scaffold. Furthermore, by gathering this
data, future analyses such as activity recognition and productivity calculation can be performed. Anjum et al. (2022) [6], sought to
measure the working height of workers using a ladder in their study. They employed a deep learning-based computer vision method to
achieve a height estimation accuracy of around 85 %. The study effectively estimated the working heights of the workers using just one
sensor. In their review research, Khan et al. (2023) [8], highlighted falls from height as a critical occupational health and safety issue in
construction sites and stressed the need to integrate advanced construction technologies for prevention. However, they found the
Fig. 11. Confusion Matrix and Learning Curve for 1s (75 %) DNN model.
Fig. 12. Confusion Matrix and Learning Curve for 5s (75 %) CNN model.
I. Karatas
Heliyon 11 (2025) e41779
11
current technologies to be inadequate. Our study proposes a deep learning model that can accurately predict whether a worker is
working at height using data from a single sensor. This innovation is expected to facilitate easier control for occupational health and
safety experts at construction sites and help reduce FFH accidents. In their review article, Newaz et al. (2022) [24], provided an
extensive evaluation delving into the impact of FFH technologies and their suitability on construction sites. Drawing from their
research ndings, a total of 7 FFH technologies were pinpointed to delineate their role in forecasting, averting, and alleviating FFH
hazards. These encompass (1) safety risk assessment and propagation, (2) real-time sensing and monitoring, (3) automated prevention
through design, (4) ontology and knowledge modeling, (5) virtual reality for FFH training, (6) personal fall arrest systems, and (7)
collective fall protection systems. The present study aims to enhance the anticipation and prevention of FFH risks among construction
workers by integrating real-time sensing and monitoring technology into the realm of automated prediction of workers operating at
elevated levels.
This proposed method can be employed to establish a worker safety monitoring system on construction sites. For example, a system
that automatically classies workers based on their working environment could send alerts to employers or safety teams in hazardous
situations. This capability is especially critical for workers on ladders or scaffolding, where the risk of falls is prevalent. Consequently,
occupational safety teams can intervene swiftly to prevent potential accidents. Furthermore, the data collected can be routinely
analyzed to pinpoint potential hazards on construction sites, enabling the implementation of proactive safety measures. Additionally,
these insights can be integrated into occupational safety training programs. By analyzing worker behavior in relation to the ndings, it
becomes possible to identify which working environments carry greater risks, thereby allowing for the development of targeted
training programs that focus on these specic areas. Real-time monitoring plays a crucial role in identifying issues such as improper
equipment usage and deviations from established procedures. For example, if a worker is detected operating at excessive heights, this
indicates a potential safety concern. Consequently, an automatic warning can be issued, prompting the worker to use the appropriate
occupational health and safety equipment. By creating a digital twin of a construction site that models the working environment and
integrates deep learning analytics, we can effectively identify potential vulnerabilities. This method enables us to simulate the heights
at which workers operate, providing essential insights into safety risks.
4. Conclusion
The construction industry is currently undergoing a technological revolution, largely due to the integration of IoT and sensor
technologies. In addition to these, articial intelligence methods are commonly employed in the literature to monitor and analyze
workers in the realm of occupational health and safety. These advanced tools serve various purposes, including identifying worker
activities through sensors, automatically calculating worker productivity, assessing occupational safety risks, determining workers’
activities at heights, and detecting worker fatigue. The primary objective of this study is to utilize accelerometer, gyroscope, and
pressure data to assess the scaffold on which a worker is operating and consequently establish the precariousness associated with
working at elevated heights.
In the construction industry, assessing the safety risks faced by workers traditionally relies on labor-intensive manual observation
and the expertise of safety professionals. However, this approach is time-consuming. To improve efciency, a proactive framework for
assessing labor safety risks is essential. This framework should also address the issue of falls from height (FFH). As part of our study, we
aim to analyze the scaffolds where workers are stationed using data gathered from the workers themselves to gain insights into FFH
risks. According to FFH evaluations in many developed countries such as Korea, the USA, and the UK, mobile scaffold and ladder work
is considered as working at height. In contrast, ground and horse scaffold work is not. In this study, based on prediction results ob-
tained from sensors, it can be inferred that working at height conditions should be applied when it is determined that work is being
conducted automatically on a scaffold that is categorized as working at height.
In this research, it has been employed deep learning methods to analyze acceleration, gyroscope, and pressure data gathered from
construction workers. The aim of the study has predicted the working conditions of these workers across three different types of
scaffolding and on the ground. The analysis enabled us to determine whether a worker was on scaffolding or on the ground, and if on
scaffolding, to identify the specic type—namely, mobile scaffolding, horse scaffolding, or ladder. It has been evaluated each deep
learning model based on various window sizes and overlap ratios, resulting in a total of 45 distinct analyses. The ndings demonstrated
that the CNN model, congured with a 75 % overlap rate and a 5-s window size, achieved accuracy, loss, and F1 score values of 0.9490,
0.1696, and 0.9475, respectively. Similarly, the DNN model with the same overlap rate and a 1-s window size explained accuracy, loss,
and F1 score values of 0.9460, 0.1445, and 0.9441, respectively. Among the deep learning models assessed, both the CNN (5-s window
size with a 75 % overlap rate) and the DNN (1-s window size with a 75 % overlap rate) exhibited exceptional prediction accuracy of
approximately 95 % when identifying the locations of workers. This remarkable level of accuracy underscores the potential for
integrating these systems into safety monitoring installations on construction sites.
The ndings of this study illustrate the potential for enhancing construction management, particularly in complex and large-scale
projects, through the integration of sensors in the work environment. This innovative technology is poised to make a signicant impact
on occupational health and safety, especially within construction sites. The model developed in this research is capable of real-time
analysis of workers’ operational statuses. In hazardous areas, this functionality enables prompt interventions, thereby improving
occupational safety and reducing the risk of accidents. Additionally, this study enhances prediction accuracy by leveraging data from a
variety of sensors, including accelerometers, gyroscopes, and barometers, rather than relying solely on a single sensor. Achieving high-
accuracy predictions for preventing falls from heights is expected to provide considerable benets to the construction industry,
particularly regarding worker safety and minimizing project downtime. Nevertheless, it is important to acknowledge that the research
has its limitations. While our model exhibits high accuracy rates, there may still be instances of false positive and false negative
I. Karatas
Heliyon 11 (2025) e41779
12
predictions. For example, a worker could be mistakenly assessed as being ‘in danger’ based on movement data, resulting in unnec-
essary alarms. Additionally, environmental factors on construction sites—such as dust, humidity, temperature uctuations, and
impacts—can interfere with the sensors’ functionality. Moreover, ensuring that workers consistently wear the sensors presents
practical challenges. The sensitivity and calibration of the employed sensors signicantly inuence the model’s performance. Subpar
or improperly positioned sensors can diminish data accuracy and adversely affect prediction outcomes. Continuous monitoring of
workers’ movements may also lead to privacy concerns, raising both ethical and legal implications. To mitigate these issues, strategies
like anonymization and the analysis of only the necessary data can be implemented. Furthermore, expanding the dataset with data
collected under varying conditions can enhance the diversity and scope of the training information. To tackle these challenges, it is
essential to develop smaller, more robust sensors and conduct regular maintenance. These advancements can support the seamless
integration of technology within construction sites. Future research can expand on the development of sensors by implementing them
in real construction site environments and involving numerous workers. Furthermore, the scope of occupational health and safety
procedures can be broadened by gathering additional data and conducting analyses to assess whether appropriate protective measures
are being taken based on height conditions. The accuracy of the proposed model can be assessed and rened by gathering data across
various environmental conditions and from more complex workplaces. Future research priorities could include enhancing the dura-
bility of sensors, developing devices with lower energy consumption, and exploring innovative technologies capable of collecting more
sensitive data. Furthermore, integrating motion data with other types of information, such as GPS and RFID, could lead to more
comprehensive analyses in the future. This data integration allows for more precise monitoring of workers’ locations, heights, and
behaviors. Conducting long-term analyses of the collected data can help identify recurring risk factors in the work environment and
support the development of proactive measures. Future research can leverage workers’ movement data not only to enhance safety but
also to conduct ergonomic assessments and identify potential risks to the musculoskeletal system. With ongoing advancements in
technology, the efcacy of predictive models can be further improved by incorporating various articial intelligence methodologies.
To facilitate real-time analysis for employers and technical staff, it is essential to enhance the integration of software and hardware.
Specically, the proposed model could benet from being integrated with mobile applications or cloud-based solutions. Moreover, the
methods developed in this study can be adapted for use in sectors beyond construction and tested across different working environ-
ments. Such adaptations could increase the system’s exibility and expand its overall applicability.
Data availability statement
All data generated or analyzed during this study are included in this published article.
Code availability
Not applicable.
Funding
The authors declare no funding for this research.
Declaration of competing interest
The authors declare that they have no known competing nancial interests or personal relationships that could have appeared to
inuence the work reported in this paper.
Acknowledgments
Not applicable.
References
[1] S. Dong, H. Li, Q. Yin, Building information modeling in combination with real time location systems and sensors for safety performance enhancement, Saf. Sci.
102 (2018) 226–237, https://doi.org/10.1016/j.ssci.2017.10.011.
[2] H. Choo, B. Lee, H. Kim, B. Choi, Automated detection of construction work at heights and deployment of safety hooks using IMU with a barometer, Autom.
ConStruct. 147 (2023) 104714, https://doi.org/10.1016/j.autcon.2022.104714.
[3] C.R. Ahn, S. Lee, C. Sun, H. Jebelli, K. Yang, B. Choi, Wearable sensing technology applications in construction safety and health, Journal of Construction
Engineering Mangement 145 (2019) 3119007, https://doi.org/10.1061/(ASCE)CO.1943-7862.0001708.
[4] K. Bhattacharjee, N. Bugalia, A. Mahalingam, An analysis of safety practices for small, medium, and large construction projects: a resilience engineering
perspective, Saf. Sci. 169 (2024) 106330, https://doi.org/10.1016/j.ssci.2023.106330.
[5] G. Albeaino, P. Brophy, I. Jeelani, M. Gheisari, R.R.A. Issa, Impact of drone presence on construction individuals working at heights, J. Constr. Eng. Manage. 149
(2023) 04023119, https://doi.org/10.1061/JCEMD4.COENG-13861.
[6] S. Anjum, N. Khan, R. Khalid, M. Khan, D. Lee, C. Park, Fall prevention from ladders utilizing a deep learning-based height assessment method, IEEE Access 10
(2022) 36725–36742, https://doi.org/10.1109/ACCESS.2022.3164676.
[7] K. Koç, ¨
O. Ekmekcioglu, A.P. Gurgun, Developing a national data-driven construction safety management framework with interpretable fatal accident
prediction, Journal of Construction Engineering Mangement 149 (2023).
I. Karatas
Heliyon 11 (2025) e41779
13
[8] M. Khan, C. Nnaji, M.S. Khan, A. Ibrahim, D. Lee, C. Park, Risk factors and emerging technologies for preventing falls from heights at construction sites, Autom.
ConStruct. 153 (2023) 104955, https://doi.org/10.1016/j.autcon.2023.104955.
[9] A. Asadzadeh, M. Arashpour, H. Li, T. Ngo, A. Bab-Hadiashar, A. Rashidi, Sensor-based safety management, Autom. ConStruct. 113 (2020) 103128, https://doi.
org/10.1016/j.autcon.2020.103128.
[10] S. Demirkesen, D. Arditi, Construction safety personnel’s perceptions of safety training practices, Int. J. Proj. Manag. 33 (2015) 1160–1169.
[11] C.M. Zuluaga, A. Albert, M.A. Winkel, Improving safety, efciency, and productivity: evaluation of fall protection systems for bridge work using wearable
technology and utility analysis, J. Constr. Eng. Manage. 146 (2020) 04019107, https://doi.org/10.1061/(ASCE)CO.1943-7862.0001764.
[12] O. Maali, C.-H. Ko, P.H. Nguyen, Applications of existing and emerging construction safety technologies, Autom. ConStruct. 158 (2024) 105231, https://doi.
org/10.1016/j.autcon.2023.105231.
[13] F. Mosto, V. To˘
gan, Predicting construction accident outcomes using graph convolutional and dual-edge safety networks, Arab J Sci Eng (2023), https://doi.
org/10.1007/s13369-023-08609-8.
[14] P. Liu, C. Zhang, D. Arditi, H. Li, S. Demirkesen, Using eye-tracking technology to assess the effect of daily safety training on hazard recognition skills, IEEE
Trans. Eng. Manage. 71 (2024) 8548–8561, https://doi.org/10.1109/TEM.2024.3367230.
[15] W. Fang, P.E. Love, L. Ding, S. Xu, T. Kong, H. Li, Computer vision and deep learning to manage safety in construction: matching images of unsafe behavior and
semantic rules, IEEE Trans. Eng. Manage. 70 (2023) 4120–4132, https://doi.org/10.1109/TEM.2021.3093166.
[16] I. Karatas, A. Budak, Inuence on construction waste of the lean–building information modelling interaction, Proceedings of the Institution of Civil Engineers -
Engineering Sustainability Ahead of Print (2023), https://doi.org/10.1680/jensu.22.10018.
[17] I. Karatas, A. Budak, Investigating the impact of lean-BIM synergy on labor productivity in the construction execution phase, Journal of Engineering Research 11
(2023) 322–333, https://doi.org/10.1016/j.jer.2023.10.021.
[18] D. Roberts, M. Golparvar-Fard, End-to-end vision-based detection, tracking and activity analysis of earthmoving equipment lmed at ground level, Autom.
ConStruct. 105 (2019) 102811, https://doi.org/10.1016/j.autcon.2019.04.006.
[19] X. Hou, C. Li, Q. Fang, Computer vision-based safety risk computing and visualization on construction sites, Autom. ConStruct. 156 (2023) 1–18, https://doi.
org/10.1016/j.autcon.2023.105129.
[20] J. Lee, S. Lee, Construction site safety management: a computer vision and deep learning approach, Sensors 23 (2023), https://doi.org/10.3390/s23020944.
[21] G. Aksüt, T. Eren, Selection of wearable sensors for health and safety use in the construction industry, J. Civ. Eng. Manag. 29 (2023) 577–586, https://doi.org/
10.3846/jcem.2023.19175.
[22] R. Gao, B. Mu, S. Lyu, H. Wang, C. Yi, Review of the application of wearable devices in construction safety: a bibliometric analysis from 2005 to 2021, Buildings
12 (2022) 344, https://doi.org/10.3390/buildings12030344.
[23] I. Karatas, A. Budak, Prediction of labor activity recognition in construction with machine learning algorithms, ICONTECH INTERNATIONAL JOURNAL OF
SURVEYS, ENGINEERING, TECHNOLOGY 5 (2021) 38–47, https://doi.org/10.46291/ICONTECHvol5iss3pp38-47.
[24] M.T. Newaz, M. Ershadi, L. Carothers, M. Jefferies, P. Davis, A review and assessment of technologies for addressing the risk of falling from height on
construction sites, Saf. Sci. 147 (2022) 105618, https://doi.org/10.1016/j.ssci.2021.105618.
[25] Q. Fang, H. Li, X. Luo, L. Ding, H. Luo, C. Li, Computer vision aided inspection on falling prevention measures for steeplejacks in an aerial environment, Autom.
ConStruct. 93 (2018) 148–164, https://doi.org/10.1016/j.autcon.2018.05.022.
[26] P. Liu, M. Xie, J. Bian, H. Li, L. Song, A hybrid PSO–svm model based on safety risk prediction for the design process in metro station construction, Int. J.
Environ. Res. Publ. Health 17 (2020) 1–24, https://doi.org/10.3390/ijerph17051714.
[27] W. Fang, B. Zhong, N. Zhao, P.E. Love, H. Luo, J. Xue, S. Xu, A deep learning-based approach for mitigating falls from height with computer vision:
convolutional neural network, Adv. Eng. Inf. 39 (2019) 170–177, https://doi.org/10.1016/j.aei.2018.12.005.
[28] S. Hong, J. Yoon, Y. Ham, B. Lee, H. Kim, Monitoring safety behaviors of scaffolding workers using Gramian angular eld convolution neural network based on
IMU sensing data, Autom. ConStruct. 148 (2023) 104748, https://doi.org/10.1016/j.autcon.2023.104748.
[29] Y.-C. Fang, R.-J. Dzeng, Accelerometer-based fall-portent detection algorithm for construction tiling operation, Autom. ConStruct. 84 (2017) 214–230, https://
doi.org/10.1016/j.autcon.2017.09.015.
[30] O. Banos, J.-M. Galvez, M. Damas, H. Pomares, I. Rojas, Window size impact in human activity recognition, Sensors 14 (2014) 6474–6499, https://doi.org/
10.3390/s140406474.
[31] E.L. Jacobsen, J. Teizer, S. Wandahl, Work estimation of construction workers for productivity monitoring using kinematic data and deep learning, Autom.
ConStruct. 152 (2023) 104932, https://doi.org/10.1016/j.autcon.2023.104932.
[32] M. Zhang, S. Chen, X. Zhao, Z. Yang, Research on construction workers’ activity recognition based on smartphone, Sensors 18 (2018), https://doi.org/10.3390/
s18082667.
[33] J. Zhao, E. Obonyo, Convolutional long short-term memory model for recognizing construction workers’ postures from wearable inertial measurement units,
Adv. Eng. Inf. 46 (2020) 101177, https://doi.org/10.1016/j.aei.2020.101177.
[34] S. Mekruksavanich, A. Jitpattanakul, Automatic recognition of construction worker activities using deep learning approaches and wearable inertial sensors,
Intelligent Automation & Soft Computing 36 (2023) 2111–2128, https://doi.org/10.32604/iasc.2023.033542.
[35] K. Kim, Y.K. Cho, Effective inertial sensor quantity and locations on a body for deep learning-based worker’s motion recognition, Autom. ConStruct. 113 (2020)
103126, https://doi.org/10.1016/j.autcon.2020.103126.
[36] M.F. Antwi-Afari, H. Li, J. Seo, A.Y.L. Wong, Automated detection and classication of construction workers’ loss of balance events using wearable insole
pressure sensors, Autom. ConStruct. 96 (2018) 189–199, https://doi.org/10.1016/j.autcon.2018.09.010.
[37] S.-M. M¨
akela, A. L¨
ams¨
a, J.S. Ker¨
anen, J. Liikka, J. Ronkainen, J. Peltola, J. H¨
aiki¨
o, S. J¨
arvinen, M. Bordallo L´
opez, Introducing VTT-ConIot: a realistic dataset for
activity recognition of construction workers using IMU devices, Sustainability 14 (2022) 220, https://doi.org/10.3390/su14010220.
[38] I. Karatas, A. Budak, Deep learning-based recognition of construction activities in real construction site environment, Eng. Construct. Architect. Manag. (2024),
https://doi.org/10.1108/ECAM-08-2024-1036 ahead-of-print.
[39] R. Akhavian, A.H. Behzadan, Smartphone-based construction workers’ activity recognition and classication, Autom. ConStruct. 71 (2016) 198–209, https://
doi.org/10.1016/j.autcon.2016.08.015.
[40] J. Ryu, J. Seo, H. Jebelli, S. Lee, Automated action recognition using an accelerometer-embedded wristband-type activity tracker, J. Constr. Eng. Manage. 145
(2019) 4018114, https://doi.org/10.1061/(ASCE)CO.1943-7862.0001579.
[41] L. Joshua, K. Varghese, Accelerometer-based activity recognition in construction, J. Comput. Civ. Eng. 25 (2011) 370–379, https://doi.org/10.1061/(ASCE)
CP.1943-5487.0000097.
[42] J. Wang, Y. Chen, S. Hao, X. Peng, L. Hu, Deep learning for sensor-based activity recognition: a survey, Pattern Recogn. Lett. 119 (2019) 3–11, https://doi.org/
10.1016/j.patrec.2018.02.010.
[43] F. Li, K. Shirahama, M.A. Nisar, L. K¨
oping, M. Grzegorzek, Comparison of feature learning methods for human activity recognition using wearable sensors,
Sensors 18 (2018), https://doi.org/10.3390/s18020679.
[44] S. Mekruksavanich, A. Jitpattanakul, Biometric user identication based on human activity recognition using wearable sensors: an experiment using deep
learning models, Electronics 10 (2021) 308, https://doi.org/10.3390/electronics10030308.
[45] Y. Zhao, X. Wang, Y. Luo, M.S. Aslam, Research on Human Activity Recognition Algorithm Based on LSTM-1DCNN, CMC (2023) 1–10, https://doi.org/
10.32604/cmc.2023.040528.
[46] T.T. Alemayoh, J.H. Lee, S. Okamoto, New sensor data structuring for deeper feature extraction in human activity recognition, Sensors 21 (2021) 2814, https://
doi.org/10.3390/s21082814.
I. Karatas
Heliyon 11 (2025) e41779
14
[47] S.S. Bangaru, C. Wang, S.A. Busam, F. Aghazadeh, ANN-based automated scaffold builder activity recognition through wearable EMG and IMU sensors, Autom.
ConStruct. 126 (2021) 103653, https://doi.org/10.1016/j.autcon.2021.103653.
[48] T. Slaton, C. Hernandez, R. Akhavian, Construction activity recognition with convolutional recurrent networks, Autom. ConStruct. 113 (2020) 103138, https://
doi.org/10.1016/j.autcon.2020.103138.
[49] K.-S. Song, J.G. Lim, D.-G. Lee, D.-S. Kwon, Motion similarity-based safety hook fastening state recognition via deep siamese neural networks, IEEE Sens. Lett. 7
(2023) 1–4, https://doi.org/10.1109/LSENS.2023.3317759.
I. Karatas
Heliyon 11 (2025) e41779
15
