A systematic review of non-intrusive human activity recognition in smart homes using deep learning

IAES International Journal of Artificial Intelligence (IJ-AI)
Vol. 13, No. 3, September 2024, pp. 3188~3202
ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i3.pp3188-3202  3188
Journal homepage: http://ijai.iaescore.com
A systematic review of non-intrusive human activity recognition
in smart homes using deep learning
Mariam El Ghazi, Noura Aknin
Information Technology and Modeling, Systems Research Unit (TIMS), Abdelmalek Essaadi University, Tetouan, Morocco
Article Info ABSTRACT
Article history:
Received Jun 21, 2023
Revised Nov 16, 2023
Accepted Jan 16, 2024
Smart homes are a viable solution for improving the independence and
privacy of elderly and dependent people thanks to IoT sensors. Reliable
human activity recognition (HAR) devices are required to enable precise
monitoring inside smart homes. Despite various reviews on HAR, there is a
lack of comprehensive studies that include a diverse range of approaches,
including sensor-based, wearable, ambient, and device-free methods.
Considering this research gap, this study aims to systematically review the
HAR studies that apply deep learning as their main solution and utilize a
non-intrusive approach for activity monitoring. Out of the 2,171 studies in
the IEEE Explore database, we carefully selected and thoroughly analyzed
37 studies for our research, following the guidelines provided by the
preferred reporting items for systematic reviews and meta-analyses
(PRISMA) methodology. In this paper, we explore various modalities, deep
learning approaches, and datasets employed in the context of non-intrusive
HAR. This study presents essential data for researchers to employ deep
learning techniques for HAR in smart home environments. Additionally, it
identifies and highlights the main trends, challenges, and future directions.
Keywords:
Datasets
Deep learning
Human activity recognition
Internet of things sensors
Non-intrusive
Preferred reporting items for
systematic reviews and meta-
analyses
This is an open access article under the CC BY-SA license.
Corresponding Author:
Mariam El Ghazi
Information Technology and Modeling, Systems Research Unit, Abdelmalek Essaadi University
Tetouan, Morocco
Email: mariam.elghazi@etu.uae.ac.ma
1. INTRODUCTION
The rising demand for human activity recognition (HAR) systems in healthcare institutions is
reshaping patient and elderly care. These systems, integrated into smart homes, help relieve the strain on
hospitals and nursing homes by providing real-time healthcare support to individuals, allowing them to
maintain their independence. This review excludes vision-based approaches because of privacy and user
acceptability issues [1]. We will focus on non-intrusive modalities like sensor-based wearables, ambient
sensors, or device-free (wireless fidelity (WiFi) and radio-frequency identification (RFID)) for HAR.
Besides, sensor-based approaches offer better recognition performance and low computational costs.
Non-intrusive HAR in smart health systems faces numerous challenges in sensor choice, activity
type, and deep model selection. Deep learning is effective for HAR in smart homes, as it extracts features
using heuristics and human expertise, overcoming the limitations of traditional methods [2]. Deep learning
(DL) can work with various network types and overcomes limitations in machine learning. Deep learning
models execute feature extraction and model-building simultaneously, learning relevant features from raw
data. They excel in complex activity recognition tasks due to their adaptability and generalization capabilities
[3]. Deep learning models like convolutional neural networks (CNNs), recurrent neural networks (RNNs),
and long short-term memory networks (LSTMs) are essential for HAR tasks like image and video

Int J Artif Intell ISSN: 2252-8938 
A systematic review of non-intrusive human activity recognition in smart homes using … (Mariam El Ghazi)
3189
recognition. Hybrid models combine various algorithms, capturing spatial and temporal features, making
them suitable for sensor data processing. CNNs perform automatic feature learning and extract higher
features from deep layers, while LSTMs excel in time-series data, resulting in better performance and high
accuracy [2]. Researchers have developed various HAR models for different applications: elderly, children,
and babies monitoring, safety, sleep monitoring, development, crowd surveillance, healthcare, lifestyle
patterns, exercise, gait analysis, abnormal activity recognition, and human activity prediction in smart homes
and other fields [4].
While significant advancement has been achieved in the field of non-intrusive HAR using deep
learning techniques in smart homes, there is a necessity for a systematic review that consolidates and
synthesizes the existing literature. Previous reviews have explored aspects of HAR or deep learning in
isolation. Thus, there is a lack of holistic research that mainly focuses on the trends, challenges, and open
issues in non-intrusive HAR using deep learning in intelligent homes. This review paper responds to this gap
by systematically examining the current literature and identifying research gaps; it also attempts to provide a
valuable resource for researchers and developers working in the field of HAR, enabling them to discover
current HAR modalities, deep learning approaches, benchmark datasets, the latest trends, and identify the
gaps and discover future directions to advance the research.
2. METHOD
2.1. The review protocol-PRISMA
The preferred reporting items for systematic reviews and meta-analyses (PRISMA) methodology led
the current study as a framework for conducting systematic reviews and meta-analyses [5]. To guarantee that
our work satisfied the requirement of a high-quality systematic review, we followed the 27-item PRISMA
review process. The systematic searching strategy consisted of three steps: identification, screening, and
eligibility.
2.2. Formulation of research questions
To determine the specific research questions and extract the search string, we applied the PICO
model [6]:
− Population (P): elderly and dependent individuals living in smart homes.
− Intervention (I): non-intrusive HAR approaches and methodologies.
− Comparison (Co): comparison of different modalities, deep learning approaches, and datasets used in
non-intrusive HAR.
− Outcomes (O): understanding the effectiveness, trends, challenges, and future directions in non-intrusive
HAR for smart homes.
To analyze each study, five research questions were formulated:
RQ1: How is the distribution of studies based on publication time?
RQ2: What modalities are used for non-intrusive HAR?
RQ3: What deep learning methods are employed for HAR, and how do they perform?
RQ4: Which datasets are commonly used for non-intrusive HAR?
RQ5: What are the trends, challenges, and future directions for non-intrusive HAR using deep learning?
2.3. Systematic searching strategies
The search process for the systematic review comprised three fundamental steps: identification,
screening, and eligibility. Figure 1 provides a comprehensive overview of the entire process through the flow
diagram. The subsequent sections will delve into a detailed explanation of the steps mentioned above,
shedding light on the intricacies of each stage in the systematic review process.
2.3.1. Identification
In the identification process, our primary goal was to enhance keyword coverage in databases by
aligning our keyword selection with the research questions. We meticulously examined each keyword to
achieve this, identifying its variations based on synonyms and related terms. This comprehensive approach
ensured that our search strategy encompassed a wide range of relevant terms and phrases, ultimately
enhancing the depth and breadth of our database search.
However, the main keywords used in this study are HAR, deep learning, sensor, wearable, and
device-free. We searched in the IEEE Explore database which provides comprehensive and advanced
searching functions. We constructed a full search string using the Boolean operator “AND” and “OR”, phrase
searching: (((“human activity recognition” OR “HAR” OR “human action recognition” OR “motion
recognition”) AND (“deep learning” OR “autoencoder” OR “auto-encoder” OR ”deep belief network” OR

 ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 3, September 2024: 3188-3202
3190
”convolutional neural network” OR ”convolution neural network” OR ”recurrent neural network” OR
”RNN” OR ”LSTM” OR ”long short term memory” OR ”generative adversarial network” OR ”GAN” OR
”reinforcement learning” OR ”attention” OR ”deep semi-supervised learning” OR ”graph neural network”)
AND (“sensor” OR “sensor-based” OR ”ambient” ”wearables” OR “smart home” OR ”smart homes” OR
”assisted living” OR “ambient assisted living” OR ”multi-resident” OR ”multiple residents” OR
“device-free” OR “WIFI” OR “RFID”))).
The review focused on original research papers and conference papers published between 2019 and
the present. In addition to this, we conducted a meticulous manual search to identify articles relevant to
non-intrusive HAR using deep learning. Through this rigorous process, we successfully retrieved a total of
2,171 articles.
Figure 1. PRISMA flowchart for study selection
2.3.2. Screening
All selected articles identified in the previous stage went through the screening process. We
screened papers for our systematic review based on these inclusion and exclusion criteria,
− Inclusion criteria : i) all the studies included the important keywords (deep learning, HAR, non-intrusive,
sensor, WIFI, RFID, wearable, ambient, smart home, healthcare, multi-resident); ii) studies in artificial
intelligence (AI), convolutional neural nets, recurrent neural nets, deep learning (artificial intelligence),
body sensor networks, sensor fusion, health care, and patient monitoring; iii) studies published between
January 1st
2019, to June 8th
2023; iv) conference papers and journal articles; and v) open access.
− Exclusion criteria: i) review paper, book chapter; ii) studies are not accessible in full text; iii) studies are
not written in English; iv) studies on video-based HAR, image HAR; v) studies before 2018; and
vi) methods other than deep learning.
Among the 2,171 papers in the IEEE Explore database, 161 articles were retained after applying the
inclusion and exclusion criteria. These articles proceeded to the screening phase, where they were evaluated
based on their title, keywords, and abstracts. During the eligibility phase, 49 articles were reviewed in full
text. The search results were exported in BibTeX format to be used as input for reference management tools.
We utilized JabRef as our reference manager tool, primarily for managing the downloaded references and
removing duplicate papers obtained from different search engines. Additionally, JabRef offers the option to
automatically download the full text of all added references, saving a significant amount of time.

3191
2.3.3. Eligibility
In the final stage, a comprehensive manual review of the articles was conducted, involving a
thorough reading of the full text. This meticulous eligibility process was applied to the articles retrieved from
IEEE Explore, ensuring strict adherence to the predetermined criteria. Ultimately, 49 articles met the
inclusion criteria and were included in this stage.
2.4. Quality appraisal
The aim of creating a quality assessment (QA) is to assess the overall quality of the selected studies.
Therefore, we utilize specific quality criteria to evaluate the strength and the relevance of the studies'
findings:
QA1. Does the study align with the research objectives?
QA2. Is the method or approach used in HAR mentioned in the study?
QA3. Is the research methodology clearly articulated and described in the study?
QA4. Is the dataset used in the study described in detail?
QA5. Has the performance of the deep learning model used in the study been explained comprehensively?
In the study, the researchers examined 37 selected studies to assess their credibility by using five
QA questions. All authors of the study reviewed and extracted data from each of the 37 studies as reviewers.
The authors appraised the QA to decide the articles’ content quality. Each reviewer rated the articles into
three levels of priority: high, medium, or low [7]. The articles ranked as high were eligible for review in the
following process. The articles marked as “low” are excluded. The articles marked as “medium” are
discussed by the authors for eligibility. A total of 12 articles that did not respond to the inclusion/exclusion
criteria or QA were removed. Finally, 37 articles were reviewed in full text for data extraction concerning
these data items: paper, year, aim, modality, DL model, dataset, accuracy, application, limitation, and future
work.
3. RESULTS AND DISCUSSION
This section presents the results of our systematic review, addressing the research questions
formulated earlier. We selected thirty-seven studies through our systematic review to identify sensor
modalities for HAR using deep learning methods. In Table 1 (see in Appendix), we provide a summary of
these selected studies, including their identity (ID), publication aim, publication year, the deep learning
methods employed, a discussion of the accuracy achieved by each method, the dataset utilized, the
application area, study limitations, and future research directions.
3.1. RQ1: how is the distribution of studies based on publication time?
The chosen studies were published within the last 5 years. Figure 2 displays the number of studies
published between the years 2019-2022. However, the graph does not include data for the year 2023, as the
research for that specific year is still in progress. Overall, four articles were published in 2019. Nineteen
articles were published in 2020, five in 2021, and eight in 2022. Based on the results, we observe that
numerous studies have been published in the last five years. Consequently, the deep learning method is a
popular approach to improving the HAR in a smart environment.
3.2. RQ2: what modalities are used for non-intrusive HAR?
In addressing our research question and drawing insights from the gathered data in selected articles,
several noteworthy observations. Figure 3 illustrates a prevalent trend wherein a majority of the papers adopt
the sensor-based approach or the wearable approach, few studies are conducted using the device-free
approach such as radar, WIFI, channel state information (CSI), or radio frequency identification (RFID), the
researchers need to investigate this area of the research area. Yu et al. [8] adopts a hybrid modality approach
by combining various types of sensors to recognize activities. By integrating various sensor types, they aim
to enhance the accuracy and overall performance of HAR systems.
Based on the data collected from Table 1 which can be found in the Appendix, we have made
several observations regarding the modalities used for HAR,
− Sensors can be embedded into devices like smartphones and smartwatches, which contain different types
of sensors (accelerometer (Acc), gyroscope (Gyr), magnetometer (Mag), and global positioning system
(GPS)) and are used widely by HAR researchers [9], [10]. In a study conducted by Ribeiro and Santos
[11], they have used the inertial measurement unit (IMU), which measures the values of Acc, Gyr, Mag,
and other sensors. Its advantage is the low cost and high accuracy, and it can be mounted to different
body parts [12].
− The use of wearables is inconvenient to the subjects, and thus, smartphones are more desirable. But it
only captures simple activities.

 ISSN: 2252-8938
3192
− The current advances in biosensors and tattoo sensors can be used for human activity recognition. These
types of wearable sensors are non-invasive and can be used for real-time monitoring. Thus, a need to
create miniaturized sensor-based devices for remote healthcare monitoring.
− Lately, another good option is a device-free HAR. The recent advancements in RFID technology are
witnessed in the internet of things (IoT) solutions, and a few applications have been proposed for activity
recognition using device-free RFID technology. In addition, Further attention has to be given to the
device-free approach because it provides more privacy and is less intrusive. However, the device-free
approach has many challenges, such as the RFID tags capturing only indoor activities and the WiFi can
only capture activities within the covered area.
− One of the primary hurdles in the HAR field is the recognition of activities involving multiple residents.
Several studies have employed ambient sensors to identify multiple occupants’ activities and determine
their indoor positions. Another challenge lies in detecting concurrent or complex activities. To address this,
Plötz et al. [13] employed object sensors capable of detecting composite activities humans perform during
their interactions with the environment. Incorporating objects in the recognition process proves crucial in
identifying more complex activities. Additionally, Yu et al. [8] proposed an alternative approach by
adopting a hybrid sensor setup, thereby enhancing the comprehensiveness of the acquired data.
− The sensor placement is crucial to get the correct reading, whether attached or worn by the subject. It
depends more on the nature of the activity. If it is a simple or more complex activity, then many sensors
are attached to the body to collect comprehensive data. Kulchyk and Etemad [14] investigated the effect
of sensor placement on activity recognition. They concluded that the optimal sensor location was on the
ankle, especially for activities of daily living (ADL), rather than on the sternum, shoulder, or thigh. Good
accuracy can be achieved with at least two sensors located in body parts to get a wide range of motion.
Figure 2. Trend in the number of selected studies
per year (2019-2022)
Figure 3. Number of studies by sensor modality
3.3. RQ3: what deep learning methods are employed for HAR, and how do they perform?
3.3.1. Deep learning methods
This section delves into the performance evaluation of deep learning models for HAR. We analyzed
the findings from various research papers to gain insights. Figure 4 shows hybrid models emerged as the
most popular choice, with 12 papers focusing on them. Following closely behind were CNN and LSTM, with
10 and 6 papers, respectively. Researchers explored a range of deep learning approaches to tackle HAR
challenges, including bidirectional long short-term memory (Bi-LSTM), generative adversarial network
(GAN), gated recurrent unit (GRU), and more recent advancements like residual network (ResNet) and
transfer learning, which showed promising results. Most deep learning approaches have demonstrated
exceptional performance, particularly those based on hybrid models. For instance, in a study conducted by
Thakur et al. [15], a ConvLSTM model was employed for HAR. CNN was utilized to capture spatial
information, while LSTM was employed to capture temporal data, resulting in an impressive accuracy of
98%. Similarly, Lu et al. [16] adopted a hybrid model combining CNN and GRU, achieving an accuracy of
95% on the PAMAP2 dataset. It should be noted that the performance of these models can vary depending on
factors such as the specific dataset used, the preprocessing stage, hardware configurations, and the size of the
dataset.

3193
Figure 4. Deep learning models used for HAR
Moreover, RNN leverage the temporal sequencing of sensor readings, considering their time-order
relationship. On the other hand, CNN excel in capturing complex features embedded within recurrent
patterns [17]. A recent and modern approach using LSTM algorithms stands out (Bi-LSTM, Casc-LSTM,
ENs2-LSTM) [18]. Thus, in the HAR problem, the hybrid model can achieve better performance as CNN is
capable of extracting features, autoencoders are used for dimensionality reduction and LSTM learns temporal
relationships [15].
A trending technique such as transfer learning has shown good results with vision-based HAR models
but there is a lack of research on using transfer learning in sensor-based HAR [4] and reinforcement learning
needs to be investigated to improve HAR performance and result [18]. Another issue that needs to be
investigated is the HAR model for synchronized activities and prediction of future action. There is a lack of
research on this topic. Researchers should focus on selecting the suitable optimizer and tuning the right
hyperparameters for a better performance of the model. Many challenges are associated with the deep learning
model. On one hand overfitting/underfitting, in the case of the limited amount of data, affects the generalization
ability. On the other hand, the specialized hardware and memory integration in HAR devices is required to
achieve high-performance HAR models, which leads to high costs and limits its use in the real environment [4].
3.3.1. Performance evaluation of the deep learning methods
Before understanding the approach used, it is crucial to assess the performance of the proposed
classification model. The effectiveness of a deep learning model can be measured by it is accuracy,
sensitivity, specificity, recall, precision, and F-measure. After reviewing 37 research papers, it was noted that
most researchers employed accuracy and F-measure as the metrics to assess the proficiency of their models.
These performance metrics are essential for assessing machine learning classification models’ performance in
a given context. Selecting the appropriate metric is a vital step in assessing a dataset. While accuracy proves
valuable in a balanced dataset class, it may not be the best fit for imbalanced datasets. Metrics such as
precision, recall, f-measure, and specificity may better suit such scenarios. The confusion matrix is frequently
used in these cases, presenting a tabular representation of class labels that depict predicted and actual classes
along two axes [19]. Table 2 defines all these metrics and their respective formula [12], [19].
Table 2 summarizes the performance metrics, including accuracy, precision, recall, specificity, and
F-measure. These metrics are essential for measuring the performance of classification models. To
understand these metrics, it is important to know the four key terms used in measuring the performance
metrics, namely true positive (tp), true negative (tn), false positive (fp), and false negative (fn) [12], [19].
3.4. RQ4: which datasets are commonly used for non-intrusive HAR?
Based on the data extracted from the selected papers, it is evident that datasets such as warehouse
instance segmentation dataset for object manipulation (WISDM), UCI, PAMAP2, and opportunity have
gained significant popularity in the field of non-intrusive HAR. The description of these datasets is presented
in Table 3. This table provides information regarding the number of subjects, activity types, and sensor types
associated with each dataset. Additionally, noteworthy datasets have emerged such as KU-HAR [20], precis
HAR [21], fall-up dataset [22], DOMUS [23], CASAS aruba [24], MIT PlaceLab [25], smart environment-
Ulster University [26], CASAS-daily life Kyoto [27], and UJAmI SmatLab [28].
Nevertheless, researchers must broaden their investigations by incorporating additional datasets to
evaluate their models’ performance comprehensively. Specifically, there is a distinct need to explore
device-free approaches like WiFI, RFiD, RADAR, and other emerging technologies within non-intrusive HAR
datasets. This area of research presents ample opportunities for further exploration and advancement. This is
primarily due to the scarcity of publicly available datasets in this particular domain. Notably, Noori et al. [29]
employed a RADAR approach, and the data used was collected directly by the authors themselves.
0
5
10
15
Number
of
studies
Deep learning model
Deep learning models used for HAR

 ISSN: 2252-8938
3194
Table 2. Performance metrics [19]
Metric Formula Definition
Accuracy 𝑡𝑝 + 𝑡𝑛
𝑡𝑝 + 𝑡𝑛 + 𝑓𝑝 + 𝑓𝑛
The ratio of correct predictions and overall predictions
Precision 𝑡𝑝
𝑡𝑝 + 𝑓𝑝
The proportion ratio of accurately predicted positive instances to the total
predicted positive instances.
Recall of sensitivity 𝑡𝑝
𝑡𝑝 + 𝑓𝑛
The ratio of correct predictions to the samples in the actual class
Specificity 𝑡𝑛
𝑡𝑛 + 𝑓𝑝
The ratio of actual class 0 to the correctly predicted 0
F1 score/F-measure 2(𝑟𝑒𝑐𝑎𝑙𝑙 ∗ 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛)
𝑟𝑒𝑐𝑎𝑙𝑙 + 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛
The weighted average of precision and Recall if the class distribution is uneven
Table 3. Public dataset for sensor-based HAR
Dataset Type of Activity subjects Type of Sensor REF
OPPORTUNITY
UCI Smartphone
PAMAP2
USC-HAD
WISDM
DSADS
Ambient kitchen
Darmstadt Daily
Actitracker
SHO
MHEALTH
Daphnet Gait
ActiveMiles
HASC
ActRecTut
ADL
ADL
ADL
ADL
ADL
ADL
Food preparation
Routines ADL
ADL
ADL
ADL
Gait
ADL
ADL
Gesture
4
30
9
14
29
8
20
1
36
10
10
10
10
1
2
ALL TYPES
Acc, Gyr
Acc, Gyr, Mg
Acc, Gyr
Acc
Acc, Gyr, Mg
Ob
Acc
Acc
Acc, Gyr, Mg
Acc, Gyr, C
Acc
Acc
Acc
Acc, Gyr
[30]‒[32]
[17]
[33]
[34]
[35]
[13]
[13]
[36]
[33]
[37]
[38]
[39]
[40]
[41]
Berkeley MHAD Daily Living 12 W, Am [42]
VanKasteren benchmark Real-world Home 9 Ob [43]
SaMO-UJA dataset Multi-residents-complex-fine-grained Activities 2 Ob, binary, W [44]
Casas multi-resident Washington State University-smart apartment 2 AM [45]
(Acc=accelerometer, Gyr=gyroscope, Mg=magnetometer, Ob=object sensor, AM=ambient, W=wearable).
The data extracted from our systematic review can be found in Table 1 in the Appendix. We have
made several observations about HAR datasets:
− The choice of the dataset is crucial to obtain the highest accuracy and achieve the best performance. The
most essential dataset features that the researcher needs to investigate, such as (the activities recorded, the
choice of sensor modalities, the value of sample rates, the placement of the sensor, the resident
information, and the application domain) are essential and must be considered when selecting a potential
dataset for HAR problem.
− Despite numerous public datasets, there remains a growing need to create additional public datasets in the
field of HAR. Researchers must address these challenges in future studies, including class imbalance,
multimodal data, composite activities, heterogeneity, and multi-resident scenarios. Acknowledging and
tackling these challenges is imperative to advance the understanding and effectiveness of HAR
methodologies.
− A gap in research that needs further investigation is the lack of a dataset for multi-residents in a smart
home environment. In addition, the activities in the existing datasets are recorded in a controlled
environment, which is not the case in natural human activity; the processing of multi-occupancy datasets
needs high computational resources.
3.5. RQ5: what are the trends, challenges, and future directions for non-intrusive HAR using deep
learning?
3.5.1. Challenges and open issues
The human activity recognition sensor-based approach faces numerous challenging issues:
− The heterogeneity of sensory data can be with users performing different motion patterns over time. Data
distributions of activities are changing over time, and emerging activities may occur. Sensor variation
may cause data disturbance with sensors.
− The sensor placement affects the data collected by the sensors. Therefore, it affects the accuracy of the
activity. Then, identifying the optimum number and placement of sensors is crucial in HAR systems.

3195
− Feature extraction: the identification of relevant characteristics that differentiate various activities is made
possible by feature extraction, which is crucial in recognizing human activity. It is an essential step since
it enables the collection of important data required for correctly distinguishing separate activities. The
quality and applicability of the features extracted from the sensor data significantly impact how well
activity recognition approaches perform.
− Multi-resident: in a smart home environment, multi-resident activity can occur. Hence, designing
solutions for handling multi-residents is necessary. The activities performed by the people at home may
be parallel when each person performs individually or collaboratively and when they collaborate to
achieve the activity, also multi-occupancy datasets require more computational resources and it is
difficult to obtain accurate results and performance in the analysis.
− Real-time HAR: a difference between empirical and real-time data can be problematic and needs to be
implemented in real-life scenarios.
− Concurrent activities: in reality, the person may not sequentially perform activities. A person may carry
out more than one activity simultaneously, that is to say, multi-tasking; the concurrent activity is executed
by a single subject, making recognition difficult.
− Device-free dataset: a lack of available public datasets based on wireless RFID technology.
− High computational cost: many researchers rely on high-performance computers (GPU). They should
concentrate on creating reliable, portable models that do not require specialist hardware to work in the
real world.
− Dataset imbalance: one of the common challenges in HAR is the imbalanced dataset, which can lead to
bias in the result and affect the model’s ability to predict. Therefore, using evaluation metrics can play a
vital role in addressing this problem.
3.5.2. Future directions
Researchers may further investigate the following interesting aspects:
− The importance is to ensure the efficacy and safety of the model by improving the precision, sensitivity,
and specificity of wearable sensors.
− The use of the advances of IoT and the miniaturization of the sensors to facilitate their integration with
wearable sensors.
− More investigation is needed into Hybrid sensor fusion methods that help improve accuracy and achieve
better performance.
− More research should be conducted on action prediction for synchronized activities.
− Other issues, such as data ownership, data sharing, data security, and data interoperability, are among the
main concerns of the HAR problem.
− Further improvements are expected in the problems related to imbalanced labels and data volume, feature
extraction, and data processing.
− A future area of research that needs to be considered is the creation of public domain datasets for
multi-residents that record complex activities and for device-free modality.
− Facing the complications of processing multi-resident datasets, algorithms based on reinforcement learning,
and transfer learning have the potential to resolve the problem and are useful in real-time analysis.
− Researchers should explore other techniques to improve experiment results and quality metrics.
− For the problem of high computational cost, researchers should focus on developing a lightweight model
that performs well on experimentation and real-life data.
− The need for explainable AI: DL models is called “black boxes”. Thus, searching for explainability
techniques is crucial to understanding how the model makes decisions.
4. CONCLUSION
This paper presents the findings of a systematic review following the PRISMA methodology, which
aimed to provide valuable insights into non-intrusive HAR using deep learning. The analysis involved a
meticulous comparison of 37 selected studies, focusing on utilizing deep learning technologies, modalities
employed, and datasets. The review encompassed the thorough examination of each study, including its
preparation, implementation, and evaluation, offering a comprehensive overview of research conducted
between January 2019 and June 2023. It is worth noting that previous research predominantly focused on
sensor-based HAR, and this review extends beyond that scope. One of the significant contributions of this
study is the presentation of multiple approaches for HAR, including the utilization of sensors, wearables, and
device-free methods, all of which prove to be suitable for monitoring the activities of elderly individuals
while preserving their privacy. Furthermore, this review explores various deep learning models and their
applications in HAR, providing valuable insights into their efficacy and potential. Throughout the review

 ISSN: 2252-8938
3196
process, the different HAR modalities, datasets, and deep learning models are extensively discussed,
shedding light on several gaps in the existing literature. Moreover, emerging trends and challenges are
identified, offering researchers valuable focus areas when tackling this topic. In summary, this systematic
review is a valuable resource for researchers in non-intrusive HAR using deep learning. By addressing the
identified gaps and considering the highlighted trends and challenges, researchers can significantly contribute
to understanding and applying deep learning in HAR.
APPENDIX
Table 1. Summary of selected studies (Continue…)
Id Y Aim M Dl Dataset Accuracy Application Limitation/future work Ref
S1 2023 Physical sports
activities
W RNN IMSB,
WISDM,
ERICA
85.01%
88.46%
93.18%
Gaming
Sport
Healthcare
- Improving the accuracy
- Real-time application
- Integration with other
technologies
[46]
S2 2023 Multi-sensing
data fusion with
neuromorphic
computing for
HAR
IMU,
S, R
Hopfield
network
neurons
Self 98,98% HAR Explore the stronger potential of
neuromorphic computing of
multi-sensing data in HAR.
[8]
S3 2022 HAR
Activity
prediction
Anomaly
detection
S DNN,
OCD-AE
LSTM
Aruba,
Cairo
93%, 76%
90%
54.6%
Elderly
monitoring in
smart home
- Limited length of sequence
used for training the models
- Correct the imbalance dataset
to improve the DL model
[47]
S4 2022 Multi-view CNN-
LSTM
architecture for
radar-based HAR
R CNN
and
LSTM
four datasets F1score:
74.7%
smartphone
security
smart-offices
- Generalization capability of the
learning methods, with a single
radar-sensor
- Online continual learning
- Framework for the absolute
context of the targets and
generalize to an unseen
environment
[48]
S5 2022 Deep CNN-
LSTM with a
self-attention
model for human
activity
recognition
S CNN
and
LSTM
H-Activity
(Self)
MHEALTH
UCI-HAR
99.93%
98.76%
93.11%
Healthcare
sports, and
fitness
tracking
- Address the problem of class
imbalances in the dataset,
strengthen our dataset by
adding more participants, and
adjust the network structure.
- A real-time classification of
security and health risks
affecting the elderly.
[49]
S6 2022 A multichannel
CNN-GRU model
for HAR
S CNN
and
GRU
WISDM
UCI-HAR
PAMAP2
96.41%,
96.67%,
96.25
HAR - Process the data more
effectively.
- Train the model on larger
benchmark datasets or Self-
collected activity data to verify
its generality for sensor-based
HAR.
[16]
S7 2022 Channel
attention-based
deep resnet for
complex HAR
W ResNet WISDM-
HARB, UT-
Smoke, UT-
Complex
94.91%,
98.75%,
97.73%
HAR - Optimizing the
hyperparameters can decrease
model size and computation
time, which can improve
efficiency.
- Incorporating spatial and
channel attention mechanisms
can enhance the accuracy of
convolutional neural networks.
[50]
S8 2022 Multitemporal
sampling module
for real-time
human activity
recognition
S multi-
temporal
sampling
module +
CNN
UCI-HAR
WISDM
OPPORTUN
ITY
PAMAP2
F1 score:
0.94
0.86
0.84
0.75
HAR - Optimize the trade-off between
accuracy and efficiency in
sensor devices
- Measure the latency and
memory of the devices and
reflect them in the search
process.to improve the
accuracy
[51]
S9 2022 HAR based on
multichannel
convolutional
neural network
with data
augmentation
W AMC-
CNN WISDM
MHEALTH
F1 score:
95.18%
99.86%
HAR - Reducing the computational
complexity and improving the
real-time interaction are still
- the contents of follow-up
research.
[52]
S10 2022 Convolutional
autoencoder
LSTM for
smartphone-based
HAR
S
SP
ConvAE-
LSTM
WISDM,
UCI,
PAMAP2,
OPPORTUN
ITY
97.76%
98.14%
94.33%
95.69%
HAR - Investigate the applicability of
the proposed method in real
life should be analyzed.
- Compare the model
performance to other recent
DL-based methods and
Experiment with more
available datasets
[15]

3197
Table 1. Summary of selected studies
S11 2021 Design of
optimal DL
model for
HAR
W BiLSTM UCI-HAR
USC-HAD
Over
93.100%.
in the two
datasets
HAR - Security and privacy
issues
- Time series Data
preprocessing
challenges
- Improve the accuracy of
the model
[53]
S12 2021 Hierarchical
deep
learning-
based HAR
model
(HiHAR)
W CNN +
BiLSTM
UCI HAPT
MobiAct
97.98%
96.16%,
Health,
abnormal
activity
security, and
fall detection
for elderly
people
- The proposed model
requires a large amount
of labeled data for
training, and the
computational
complexity is high
- Developing more
efficient and lightweight
models that can be
deployed on resource-
constrained devices, and
exploring the possibility
of using transfer
learning to improve the
performance of the
model on new domains.
[54]
S13 2021 Ultra-
wideband
radar-based
activity
recognition
using deep
learning
S
R
LSTM Self 99.6%. HAR
Elderly
Position
- Different sensor fusion
strategies might be
explored
- Perform more complex
experiments in real-time
environments. Include
the heart rate into the
system for detecting
emergencies.
[29]
S14 2021 Unsupervised
domain
adaptation in
activity
recognition: a
gan-based
approach
S GAN HA, HB, and HC
PAMAP2
UCI-sport
Accuracy
between
40% and
60%.
HAR - Test shift-GAN on
image and text data.
- Investigate a more
unified network to
eliminate the need for
the SVM classifier.
[55]
S15 2021 Address the
problems of:
insufficient
training data
and biased
training data
S DIM
BLS
WISDM
HAPT
Overall
accuracy
93.6 %
HAR Online HAR baseline
model training
[56]
S16 2020 wearable
wireless
multi-sensor
system for
HAR
W CNN
LSTM
ConvLSTM
Self, not public
dataset
90.8%
90.5%
94%
HAR - The dataset used is not
publicly available
- Present public
benchmark dataset
concerning activity
types and sample size.
[57]
S17 2020 passive
device-free
WiFi CSI
based human
identity
identification
approach
using RNN
(Wihi) for
HAR.
WC RNN Self 96% human
identity
identification
HAR
- Multi-target
identification, walking
path
- The testing range
- Collect enough, datasets
to improve the
performance of human
identity identification
also overcomes these
limitations and
challenging
[58]
S18 2020 attention-based
encoder-
decoder
framework for
multi-sensory
time-series
analytic of
wearable
sensor
W LSTM Self 99,27% Squat
activities
Deploy a lightweight
model on resource
limited hardware to
improve results and
reduce cost.
[59]
S19 2020 investigate the
effectiveness
of
personalization
methods for
human activity
recognition
(HAR) using
accelerometer
signals
S AdaBoost-
HC
AdaBoost-
CNN
UniMiB-SHAR
MobiAct
Motion Sense
accuracy
increased
by about
11% after
model
personali-
zation
HAR Including the gyroscope
in the analysis proving
personalization methods
on other publicly
available datasets
[60]

 ISSN: 2252-8938
3198
S20 2020 DNN for HAR
with wearable
sensors:
leave-one-
subject-out
cross-validation
for model
selection
W CNN+
LOSOCV
10 fold cross
validation+CNN
MHEALTH 85.1%
99.85%
HAR - Finding the best
window size
- LOSOCV is time
consuming
- Evaluate the approach
with
- Different data sets and
improve the accuracy
of HAR through
personalization.
[61]
S21 2020 Improved loss
function for
sensor-HAR
based on LSTM
RNNs
S LSTM UCI dataset
Opportunity
dataset
92.98%
90.36%
HAR To improve the
classification
performance of both
CNNs and RNNs, it
may be beneficial to
propose novel loss
functions, such as
harmonic loss
functions. These loss
functions could take
into account the relative
importance of different
sequence errors to
improved classification
accuracy.
[62]
S22 2020 Deep learning
multi-channel
architecture
using a CNN
and BLSTM
W CNN and
BLSTM
Phone
WISDM
Phone+watch
97.91%
96.60%
99.13%
HAR - Extend the tuning
method
- Explore better ways to
automatically tune
and adjust parameters
rather than the grid
search method
[63]
S23 2020 Using ResNet
transfer deep
learning
methods in
person
identification
according to
physical actions
W RestNet
+Transfer deep
learning
UCI database 94.21% Person
Identification
Not mentioned [64]
S24 2020 Deep human
activity
recognition with
localisation of
wearable sensors
W CNN RWHAR F1-score
of 0.90
HAR
Localization
Combining
complementary
information from both
waist and shin data
helped in further
improving the activity
recognition accuracy
[65]
S25 2020 Sensor-based
open-set human
activity
recognition
using
representation
learning with
mixup triplets
S Mixup Triplet
with deep metric
learning
UCI HAR
USC HAD
PAMAP2
F1 score:
0.66
0.58
0.67
HAR Apply the proposed
method in more
realistic scenario by
adding the concept of
incremental learning.
[66]
S26 2020 A hybrid
network based on
dense connection
and weighted
feature
aggregation for
HAR
S ConvLSTM OPPO
UniMiB-SHAR
F1 score:
92.3%
97.3%,
HAR Verify the robustness
and practicality of the
model, and test it on
other datasets
[67]
S27 2020 Sensor-based
HAR using
deep stacked
multilayered
perceptron
model
S Deep Stacked
Multilayered
Perceptron
+ANN
UCI-HHAR
HAR-SP
97.3%
99.4%
HAR Use stacked ensemble
learning to achieve
higher performance for
HAR classification.
[68]
S28 2020 Continuous
human activity
classification
from FMCW
radar With Bi-
LSTM networks
R Bi-lstm Carnegie
Mellon MOCAP
90% HAR - Apply HAR for multi-
residents in the radar
field of view
Explore techniques for
decomposing
signatures for
classification.
[69]
S29 2020 LSTM-CNN
architecture for
human activity
recognition
SP LSTM-CNN UCI,
WISDM,
OPPORTUNITY
95.78
95.85%
92.63%.
HAR Not mentioned [70]

3199
S30 2020 Human activity
recognition based
on gramian
angular field and
deep
convolutional
neural network
S Gramian
angular Field
(GAF) and
deep CNN
WISDM,
UCI HAR,
OPPO
96.83%
89.48%
97.27%
HAR To make machine learning
models more practical for
wearable devices, it is
recommended to compress and
miniaturize them. This will
reduce computation, save
hardware resources, and
increase equipment standby
time.
[71]
S31 2020 Temporal-
frequency
attention-based
human activity
recognition using
commercial WiFi
devices
WC LSTM line-of-
sight
non-line-of
sight
96.6%
93%.
S32 2020 Human daily
activity
recognition
performed using
wearable inertial
sensors combined
with deep
learning
algorithms
W CNN University
of
California
(UCI),
93.77% HAR
rehabilitation
exercise
Evaluate the amount of
rehabilitation exercise for
individuals with reduced
mobility, such as patients
undergoing dialysis.
[73]
S33 2020 Improved
Bayesian
convolution
network for HAR
to analyze health
care data using
wearable IoT
device
W Improved
Bayesian
Convolution
Network
(IBCN)
Self Over 96% HAR The system's architecture
comprises Wi-Fi and Cloud-
based applications, which
allows for the seamless
addition of new users while
enabling updates with the
latest training sets.
[74]
S34 2019 WiFi CSI-based
HAR using RNN
WC RNN Self More than
95%
S35 2019 Deep SRUs-
GRUs based
activity
recognition
system based on
wearable body
multi-sensors data
W SRU+GRU MHEALT
H
99.80% HAR - Analyze the proposed model
on complex and bigger
datasets with more complex
activities to get a real-time
human behavior monitoring
system.
- investigate the consumption
of computational of the
model.
[76]
S36 2019 InnoHAR: a deep
neural network
for complex
human activity7
recognition6
W inception
neural
network and
RNN
OPPO
PAMAP2
SP
F1 score:
94.6%
93.5%
94.5%
HAR - Investigate the kernel size
and the connection method
- Address the problem of data
imbalance in real-life HAR.
[77]
S37 2019 Design and
implementation
of a convolutional
neural network on
an edge
computing
smartphone for
HAR
S CNN Self 96.4% HAR The utilization of the
proposed model requires
significant computational and
energy resources. However, it
is useful for creating smart
wearables and devices that do
not rely on cloud
connectivity. This could lead
to enhanced user privacy and
security.
[78]
S=sensor
W=wearable
R=radar
SP=smartphone
AdaBoost-HC=AdaBoost combined with hand-
crafted
SRU=simple recurrent unit
RWHAR=real-world human activity
recognition
DIM=deep InfoMax
BLS=based incremental learning
CSI=channel state information
OPPO=opportunity dataset
Self=dataset self-collected
REFERENCES
[1] K. V. Kumar and J. Harikiran, “Privacy preserving human activity recognition framework using an optimized prediction algorithm,”
IAES International Journal of Artificial Intelligence, vol. 11, no. 1, pp. 254–264, Mar. 2022, doi: 10.11591/ijai.v11.i1.pp254-264.
[2] D. Bouchabou, S. M. Nguyen, C. Lohr, B. Leduc, and I. Kanellos, “A survey of human activity recognition in smart homes based
on iot sensors algorithms: Taxonomies, challenges, and opportunities with deep learning,” Sensors, vol. 21, no. 18, 2021, doi:
10.3390/s21186037.
[3] T. T. Alemayoh, J. H. Lee, and S. Okamoto, “New sensor data structuring for deeper feature extraction in human activity
recognition,” Sensors, vol. 21, no. 8, 2021, doi: 10.3390/s21082814.
[4] N. Gupta, S. K. Gupta, R. K. Pathak, V. Jain, P. Rashidi, and J. S. Suri, “Human activity recognition in artificial intelligence framework:
a narrative review,” Artificial Intelligence Review, vol. 55, no. 6, pp. 4755–4808, Jan. 2022, doi: 10.1007/s10462-021-10116-x.
[5] M. Vrabel, “Preferred reporting items for systematic reviews and meta-analyses,” Oncology Nursing Forum, vol. 42, no. 5, pp.
552–554, 2015, doi: 10.1188/15.ONF.552-554.
[6] S. A. Miller and J. L. Forrest, “Enhancing your practice through evidence-based decision making: PICO, learning how to ask good
questions,” Journal of Evidence-Based Dental Practice, vol. 1, no. 2, pp. 136–141, 2001, doi: 10.1016/s1532-3382(01)70024-3.

 ISSN: 2252-8938
3200
[7] R. L. S. Silva and F. W. Neiva, “Systematic literature review in computer science-a practical guide,” Relatorios Tecnicos Do
DCC/UFJF, vol. 1, no. 8, 2016, doi: 10.13140/RG.2.2.35453.87524.
[8] Z. Yu et al., “An intelligent implementation of multi-sensing data fusion with neuromorphic computing for human activity
recognition,” IEEE Internet of Things Journal, vol. 10, no. 2, pp. 1124–1133, Jan. 2023, doi: 10.1109/JIOT.2022.3204581.
[9] A. Bulling, U. Blanke, and B. Schiele, “A tutorial on human activity recognition using body-worn inertial sensors,” ACM
Computing Surveys, vol. 46, no. 3, Jan. 2014, doi: 10.1145/2499621.
[10] W. Qi, H. Su, C. Yang, G. Ferrigno, E. D. Momi, and A. Aliverti, “A fast and robust deep convolutional neural networks for
complex human activity recognition using smartphone,” Sensors, vol. 19, no. 17, Aug. 2019, doi: 10.3390/s19173731.
[11] N. F. Ribeiro and C. P. Santos, “Inertial measurement units: a brief state of the art on gait analysis,” ENBENG 2017-5th
Portuguese Meeting on Bioengineering, Proceedings, Mar. 2017, doi: 10.1109/ENBENG.2017.7889458.
[12] N. S. Khan and M. S. Ghani, “A survey of deep learning based models for human activity recognition,” Wireless Personal
Communications, vol. 120, no. 2, pp. 1593-1635, 2021, doi: 10.1007/s11277-021-08525-w.
[13] T. Plötz, N. Y. Hammerla, and P. Olivier, “Feature learning for activity recognition in ubiquitous computing,” IJCAI International
Joint Conference on Artificial Intelligence, pp. 1729–1734, 2011, doi: 10.5591/978-1-57735-516-8/IJCAI11-290.
[14] J. Kulchyk and A. Etemad, “Activity recognition with wearable accelerometers using deep convolutional neural network and the
effect of sensor placement,” Proceedings of IEEE Sensors, pp. 1–4, 2019, doi: 10.1109/SENSORS43011.2019.8956668.
[15] D. Thakur, S. Biswas, E. S. L. Ho, and S. Chattopadhyay, “ConvAE-LSTM: convolutional autoencoder long short-term memory
network for smartphone-based human activity recognition,” IEEE Access, vol. 10, pp. 4137–4156, Jun. 2022, doi:
10.1109/ACCESS.2022.3140373.
[16] L. Lu, C. Zhang, K. Cao, T. Deng, and Q. Yang, “A multichannel CNN-GRU model for human activity recognition,” IEEE
Access, vol. 10, pp. 66797–66810, Jun. 2022, doi: 10.1109/ACCESS.2022.3185112.
[17] Y. Zheng, Q. Liu, E. Chen, Y. Ge, and J. L. Zhao, “Time series classification using multi-channels deep convolutional neural
networks,” International conference on web-age information management, pp. 298–310, 2014, doi: 10.1007/978-3-319-08010-9_33.
[18] P. P. A. -Colpas et al., “Human activity recognition data analysis: history, evolutions, and new trends,” Sensors, vol. 22, no. 9,
Apr. 2022, doi: 10.3390/s22093401.
[19] M. E. Ghazi and N. Aknin, “A comparison of sampling methods for dealing with imbalanced wearable sensor data in human
activity recognition using deep learning,” International Journal of Advanced Computer Science and Applications, vol. 14, no. 10,
pp. 290–305, 2023, doi: 10.14569/IJACSA.2023.0141032.
[20] N. Sikder and A. Al Nahid, “KU-HAR: an open dataset for heterogeneous human activity recognition,” Pattern Recognition
Letters, vol. 146, pp. 46–54, Jun. 2021, doi: 10.1016/j.patrec.2021.02.024.
[21] A. -C. Popescu, I. Mocanu, and B. Cramariuc, “Precis HAR,” IEEE Dataport, 2019, doi: 10.21227/mene-ck48.
[22] L. M. -Villaseñor, H. Ponce, J. Brieva, E. M. -Albor, J. N. -Martínez, and C. P. -Asturiano, “Up-fall detection dataset: a
multimodal approach,” Sensors, vol. 19, no. 9, Apr. 2019, doi: 10.3390/s19091988.
[23] M. Gallissot, J. Caelen, N. Bonnefond, B. Meillon, and S. Pons, “Using the multicom domus dataset,” Laboratoire d'Informatique
de Grenoble, pp. 1-18, 2011.
[24] D. J. Cook, “Learning setting-generalized activity models for smart spaces,” IEEE Intelligent Systems, vol. 27, no. 1, pp. 32–38,
Jan. 2012, doi: 10.1109/MIS.2010.112.
[25] B. Logan, J. Healey, M. Philipose, E. M. Tapia, and S. Intille, “A long-term evaluation of sensing modalities for activity
recognition,” International conference on Ubiquitous computing, pp. 483–500, 2007, doi: 10.1007/978-3-540-74853-3_28.
[26] C. D. Nugent, M. D. Mulvenna, X. Hong, and S. Devlin, “Experiences in the development of a Smart Lab,” International Journal
of Biomedical Engineering and Technology, vol. 2, no. 4, pp. 319–331, 2009, doi: 10.1504/IJBET.2009.027796.
[27] D. J. Cook and M. S. -Edgecombe, “Assessing the quality of activities in a smart environment,” Methods of Information in
Medicine, vol. 48, no. 5, pp. 480–485, 2009, doi: 10.3414/ME0592.
[28] M. Espinilla, L. Martinez, J. Medina, and C. Nugent, “The experience of developing the UJAmI Smart Lab,” IEEE Access, vol. 6,
pp. 34631–34642, Jun. 2018, doi: 10.1109/ACCESS.2018.2849226.
[29] F. M. Noori, M. Z. Uddin, and J. Torresen, “Ultra-wideband radar-based activity recognition using deep learning,” IEEE Access,
vol. 9, pp. 138132–138143, Jun. 2021, doi: 10.1109/ACCESS.2021.3117667.
[30] R. Chavarriaga et al., “The opportunity challenge: a benchmark database for on-body sensor-based activity recognition,” Pattern
Recognition Letters, vol. 34, no. 15, pp. 2033–2042, Nov. 2013, doi: 10.1016/j.patrec.2012.12.014.
[31] F. J. Ordóñez and D. Roggen, “Deep convolutional and LSTM recurrent neural networks for multimodal wearable activity
recognition,” Sensors, vol. 16, no. 1, Jan. 2016, doi: 10.3390/s16010115.
[32] B. Almaslukh, B. Almaslukh, J. Almuhtadi, and A. Artoli, “An effective deep autoencoder approach for online smartphone-based
human activity recognition online smartphone-based human activity recognition view project an effective deep autoencoder
approach for online smartphone-based human activity recognition,” IJCSNS International Journal of Computer Science and
Network Security, vol. 17, no. 4, 2017.
[33] W. Jiang and Z. Yin, “Human activity recognition using wearable sensors by deep convolutional neural networks,” MM 2015-
Proceedings of the 2015 ACM Multimedia Conference, pp. 1307–1310, Oct. 2015, doi: 10.1145/2733373.2806333.
[34] M. A. Alsheikh, A. Selim, D. Niyato, L. Doyle, S. Lin, and H. P. Tan, “Deep activity recognition models with triaxial
accelerometers,” AAAI Workshop-Technical Report, pp. 8–13, 2016, doi: 10.48550/arXiv.1511.04664.
[35] L. Zhang, X. Wu, and D. Luo, “Recognizing human activities from raw accelerometer data using deep neural networks,”
Proceedings-2015 IEEE 14th
International Conference on Machine Learning and Applications, ICMLA 2015, pp. 865–870, 2016,
doi: 10.1109/ICMLA.2015.48.
[36] M. Zeng et al., “Convolutional neural networks for human activity recognition using mobile sensors,” in Proceedings of the 2014
6th
International Conference on Mobile Computing, Applications and Services, MobiCASE 2014, 2015, pp. 197–205, doi:
10.4108/icst.mobicase.2014.257786.
[37] F. M. Rueda, R. Grzeszick, G. A. Fink, S. Feldhorst, and M. T. Hompel, “Convolutional neural networks for human activity
recognition using body-worn sensors,” Informatics, vol. 5, no. 2, May 2018, doi: 10.3390/informatics5020026.
[38] N. Y. Hammerla, S. Halloran, and T. Plötz, “Deep, convolutional, and recurrent models for human activity recognition using
wearables,” IJCAI International Joint Conference on Artificial Intelligence, pp. 1533–1540, 2016.
[39] D. Ravi, C. Wong, B. Lo, and G. Z. Yang, “A deep learning approach to on-node sensor data analytics for mobile or wearable
devices,” IEEE Journal of Biomedical and Health Informatics, vol. 21, no. 1, pp. 56–64, 2017, doi: 10.1109/JBHI.2016.2633287.
[40] T. Hayashi, M. Nishida, N. Kitaoka, and K. Takeda, “Daily activity recognition based on DNN using environmental sound and
acceleration signals,” 2015 23rd
European Signal Processing Conference, EUSIPCO 2015, pp. 2306–2310, 2015, doi:
10.1109/EUSIPCO.2015.7362796.

3201
[41] S. W. Heads and L. Leuzinger, “On the placement of the cretaceous orthopteran brauckmannia groeningae from Brazil, withnotes
on the relationships of schizodactylidae (Orthoptera, Ensifera),” ZooKeys, vol. 77, pp. 17–30, 2011, doi: 10.3897/zookeys.77.769.
[42] F. Ofli, R. Chaudhry, G. Kurillo, R. Vidal, and R. Bajcsy, “Berkeley MHAD: a comprehensive multimodal human action
database,” in IEEE Workshop on Applications of Computer Vision, 2013, pp. 53–60, doi: 10.1109/WACV.2013.6474999.
[43] T. L. M. V. Kasteren, G. Englebienne, and B. J. A. Kröse, “Activity recognition using semi-Markov models on real world smart home
datasets,” Journal of Ambient Intelligence and Smart Environments, vol. 2, no. 3, pp. 311–325, 2010, doi: 10.3233/AIS-2010-0070.
[44] E. D. -L. -H. -Franco, E. R. B. Monroy, P. A. -Colpas, F. M. -Palechor, and M. E. Estévez, “UJA human activity recognition
multi-occupancy dataset,” The Annual Hawaii International Conference on System Sciences, pp. 1938–1947, 2021, doi:
10.24251/hicss.2021.236.
[45] G. Singla, D. J. Cook, and M. S. -Edgecombe, “Recognizing independent and joint activities among multiple residents in smart
environments,” Journal of Ambient Intelligence and Humanized Computing, vol. 1, no. 1, pp. 57–63, 2010, doi: 10.1007/s12652-
009-0007-1.
[46] M. M. Afsar et al., “Body-worn sensors for recognizing physical sports activities in exergaming via deep learning model,” IEEE
[47] K. A. Alaghbari, M. H. Mohamad, A. Hussain, and M. R. Alam, “Activities recognition, anomaly detection and next activity prediction
based on neural networks in smart homes,” IEEE Access, vol. 10, pp. 28219–28232, 2022, doi: 10.1109/ACCESS.2022.3157726.
[48] H. U. R. Khalid, A. Gorji, A. Bourdoux, S. Pollin, and H. Sahli, “Multi-view CNN-LSTM architecture for radar-based human
activity recognition,” IEEE Access, vol. 10, pp. 24509–24519, 2022, doi: 10.1109/ACCESS.2022.3150838.
[49] M. A. Khatun et al., “Deep CNN-LSTM with self-attention model for human activity recognition using wearable sensor,” IEEE
Journal of Translational Engineering in Health and Medicine, vol. 10, pp. 1–16, Jun. 2022, doi: 10.1109/JTEHM.2022.3177710.
[50] S. Mekruksavanich, A. Jitpattanakul, K. Sitthithakerngkiet, P. Youplao, and P. Yupapin, “ResNet-SE: channel attention-based
deep residual network for complex activity recognition using wrist-worn wearable sensors,” IEEE Access, vol. 10, pp. 51142–
51154, Jun. 2022, doi: 10.1109/ACCESS.2022.3174124.
[51] J. Park, W. S. Lim, D. W. Kim, and J. Lee, “Multitemporal sampling module for real-time human activity recognition,” IEEE
[52] W. Shi, X. Fang, G. Yang, and J. Huang, “Human activity recognition based on multichannel convolutional neural network with
data augmentation,” IEEE Access, vol. 10, pp. 76596–76606, Jun. 2022, doi: 10.1109/ACCESS.2022.3192452.
[53] F. N. Al-Wesabi et al., “Design of optimal deep learning based human activity recognition on sensor enabled internet of things
environment,” IEEE Access, vol. 9, pp. 143988–143996, Jun. 2021, doi: 10.1109/ACCESS.2021.3112973.
[54] N. T. H. Thu and D. S. Han, “HiHAR: A hierarchical hybrid deep learning architecture for wearable sensor-based human activity
recognition,” IEEE Access, vol. 9, pp. 145271–145281, Jun. 2021, doi: 10.1109/ACCESS.2021.3122298.
[55] A. R. Sanabria, F. Zambonelli, and J. Ye, “Unsupervised domain adaptation in activity recognition: A gan-based approach,” IEEE
[56] J. Xiao, L. Chen, H. Chen, and X. Hong, “Baseline model training in sensor-based human activity recognition: an incremental
learning approach,” IEEE Access, vol. 9, pp. 70261–70272, Jun. 2021, doi: 10.1109/ACCESS.2021.3077764.
[57] G. Ascioglu and Y. Senol, “Design of a wearable wireless multi-sensor monitoring system and application for activity recognition
using deep learning,” IEEE Access, vol. 8, pp. 169183–169195, 2020, doi: 10.1109/ACCESS.2020.3024003.
[58] J. Ding, Y. Wang, and X. Fu, “Wihi: WiFi based human identity identification using deep learning,” IEEE Access, vol. 8, pp.
129246–129262, 2020, doi: 10.1109/ACCESS.2020.3009123.
[59] C. Fan and F. Gao, “A new approach for smoking event detection using a variational autoencoder and neural decision forest,”
IEEE Access, vol. 8, pp. 120835–120849, 2020, doi: 10.1109/ACCESS.2020.3006163.
[60] A. Ferrari, D. Micucci, M. Mobilio, and P. Napoletano, “On the personalization of classification models for human activity
recognition,” IEEE Access, vol. 8, pp. 32066–32079, 2020, doi: 10.1109/ACCESS.2020.2973425.
[61] D. Gholamiangonabadi, N. Kiselov, and K. Grolinger, “Deep neural networks for human activity recognition with wearable
sensors: leave-one-subject-out cross-validation for model selection,” IEEE Access, vol. 8, pp. 133982–133994, 2020, doi:
10.1109/ACCESS.2020.3010715.
[62] Y. Hu et al., “Harmonic loss function for sensor-based human activity recognition based on LSTM recurrent neural networks,”
IEEE Access, vol. 8, pp. 135617–135627, 2020, doi: 10.1109/ACCESS.2020.3003162.
[63] I. K. Ihianle, A. O. Nwajana, S. H. Ebenuwa, R. I. Otuka, K. Owa, and M. O. Orisatoki, “A deep learning approach for human
activities recognition from multimodal sensing devices,” IEEE Access, vol. 8, pp. 179028–179038, Jun. 2020, doi:
10.1109/ACCESS.2020.3027979.
[64] S. Kilic, I. Askerzade, and Y. Kaya, “Using ResNet transfer deep learning methods in person identification according to physical
actions,” IEEE Access, vol. 8, pp. 220364–220373, Jun. 2020, doi: 10.1109/ACCESS.2020.3040649.
[65] I. A. Lawal and S. Bano, “Deep human activity recognition with localisation of wearable sensors,” IEEE Access, vol. 8, pp.
155060–155070, 2020, doi: 10.1109/ACCESS.2020.3017681.
[66] M. Lee and J. Bae, “Deep learning based real-time recognition of dynamic finger gestures using a data glove,” IEEE Access, vol.
8, pp. 219923–219933, 2020, doi: 10.1109/ACCESS.2020.3039401.
[67] T. Lv, X. Wang, L. Jin, Y. Xiao, and M. Song, “A hybrid network based on dense connection and weighted feature aggregation
for human activity recognition,” IEEE Access, vol. 8, pp. 68320–68332, 2020, doi: 10.1109/ACCESS.2020.2986246.
[68] F. Rustam et al., “Sensor-based human activity recognition using deep stacked multilayered perceptron model,” IEEE Access, vol.
8, pp. 218898–218910, 2020, doi: 10.1109/ACCESS.2020.3041822.
[69] A. Shrestha, H. Li, J. L. Kernec, and F. Fioranelli, “Continuous human activity classification from FMCW radar with Bi-LSTM
networks,” IEEE Sensors Journal, vol. 20, no. 22, pp. 13607–13619, 2020, doi: 10.1109/JSEN.2020.3006386.
[70] K. Xia, J. Huang, and H. Wang, “LSTM-CNN architecture for human activity recognition,” IEEE Access, vol. 8, pp. 56855–
56866, 2020, doi: 10.1109/ACCESS.2020.2982225.
[71] H. Xu et al., “Human activity recognition based on gramian angular field and deep convolutional neural network,” IEEE Access,
vol. 8, pp. 199393–199405, Jun. 2020, doi: 10.1109/ACCESS.2020.3032699.
[72] X. Yang, R. Cao, M. Zhou, and L. Xie, “Temporal-frequency attention-based human activity recognition using commercial WiFi
devices,” IEEE Access, vol. 8, pp. 137758–137769, 2020, doi: 10.1109/ACCESS.2020.3012021.
[73] C. T. Yen, J. X. Liao, and Y. K. Huang, “Human daily activity recognition performed using wearable inertial sensors combined
with deep learning algorithms,” IEEE Access, vol. 8, pp. 174105–174114, 2020, doi: 10.1109/ACCESS.2020.3025938.
[74] Z. Zhou, H. Yu, and H. Shi, “Human activity recognition based on improved bayesian convolution network to analyze health care
data using wearable IoT device,” IEEE Access, vol. 8, pp. 86411–86418, 2020, doi: 10.1109/ACCESS.2020.2992584.
[75] J. Ding and Y. Wang, “WiFi CSI-based human activity recognition using deep recurrent neural network,” IEEE Access, vol. 7, pp.
174257–174269, 2019, doi: 10.1109/ACCESS.2019.2956952.

 ISSN: 2252-8938
3202
[76] A. Gumaei, M. M. Hassan, A. Alelaiwi, and H. Alsalman, “A hybrid deep learning model for human activity recognition using
multimodal body sensing data,” IEEE Access, vol. 7, pp. 99152–99160, 2019, doi: 10.1109/ACCESS.2019.2927134.
[77] C. Xu, D. Chai, J. He, X. Zhang, and S. Duan, “InnoHAR: A deep neural network for complex human activity recognition,” IEEE
Access, vol. 7, no. c, pp. 9893–9902, 2019, doi: 10.1109/ACCESS.2018.2890675.
[78] T. Zebin, P. J. Scully, N. Peek, A. J. Casson, and K. B. Ozanyan, “Design and implementation of a convolutional neural network
on an edge computing smartphone for human activity recognition,” IEEE Access, vol. 7, pp. 133509–133520, 2019, doi:
10.1109/ACCESS.2019.2941836.
BIOGRAPHIES OF AUTHORS
Mariam El Ghazi received a Bachelor’s degree in Information Systems and
Software from Abdelmalek Essaadi in Tetouan, Morocco, in 2010. In 2012, she completed her
Master’s degree in Software Quality at the same institution. Currently, she is a Ph.D. student in
artificial intelligence at the information technology and modeling, systems research unit at her
university. Her primary research interest centers around deep learning, where she actively
engages in cutting-edge research to make meaningful contributions to the field. She is
dedicated to advancing knowledge in human activity recognition using deep learning and is
known for her collaborative spirit, frequently engaging with fellow researchers. She eagerly
shares her research findings with the broader academic community at several national and
international conferences. For inquiries, you can contact her at mariam.elghazi@etu.uae.ac.ma.
Noura Aknin is a Professor of Electrical and Computer Engineering at
Abdelmalek Essaadi University since 2000. She received Ph.D. degree in Electrical
Engineering in 1998. She currently serves as the Vice-Dean of Scientific Research and
Cooperation at the Faculty of Science and leads the Research Laboratory specializing in
Information Technology and Modeling Systems. She coordinates a network administration and
security of information systems bachelor, electronic and optic for embedded systems bachelor
and cybersecurity and administration of networks and information systems bachelor. She is an
IEEE senior member and is a co-founder of the IEEE Morocco Section in November 2004 and
she was the women in engineering coordinator. She actively participates in the IEEE
Communications Society, Computer Society, and Women in Engineering Society. Her
research interests focus mainly on mobile and wireless networks, social web and e-learning,
big data, IoT, and AI. She has authored and co-authored more than 200 papers presented at
both national and international conferences and published in journals. She has also served in
the organizing committees and TPC and presented keynote talks at many international
conferences worldwide. Moreover, she has supervised several Ph.D. and Masters Theses. She
can be contacted at email: noura.aknin@uae.ac.ma.

A systematic review of non-intrusive human activity recognition in smart homes using deep learning

More Related Content

Similar to A systematic review of non-intrusive human activity recognition in smart homes using deep learning (20)

More from IAESIJAI (20)

Recently uploaded (20)

A systematic review of non-intrusive human activity recognition in smart homes using deep learning