EdgeFall: a promising cloud-edge-end architecture for elderly fall care

I

Elder citizens face sudden fall, which can lead to injuries of both destructive and non-virulent. These sudden falls are later more precarious than diseases like heart attack, blood sugar, blood pressure because these can be untreated for a lengthy time which can lead to death. Elder citizen who experiences a precipitous fall, carry out their communal life narrowed. Therefore, a shrewd and adequate anti-fallen system is required for aiding elderly health care, specifically to those who live individually. So, it can identify and anticipate a precipitous fall through appropriate human activity recognition. In this study, we have suggested an end-edge-cloud based wearable EdgeFall architecture for elderly care. We have performed simulation setups to clarify the query of why we need such a strategy, and its validity. We have achieved maximum 91.87% accuracy with 1.6% false alarm rate (FAR). These empirical results indicate the superiority of using tightly couple multiple information for recognizing human activity. We can accomplish a low FAR with an enhanced accuracy. We can observe that our proposed end-edge-cloud based architecture can reduce the execution time to millisecond range (ms) of 14.16 to 15.74. This work serves as the starting mark for future related research activities.

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 13, No. 4, August 2023, pp. 4721∼4733
ISSN: 2088-8708, DOI: 10.11591/ijece.v13i4.pp4721-4733 ❒ 4721
EdgeFall: a promising cloud-edge-end architecture for
elderly fall care
Kazi Md Shahiduzzaman, Md. Salah Uddin Yusuf
Department of Electrical and Electrical Engineering, Khulna University of Engineering and Technology, Khulna, Bangladesh
Article Info
Article history:
Received Jul 4, 2022
Revised Oct 26, 2022
Accepted Nov 6, 2022
Keywords:
Accelerometer
Elderly care
End-edge-cloud architecture
Gyroscope
Human activity
Long term short memory
Wearable camera
ABSTRACT
Elder citizens face sudden fall, which can lead to injuries of both destructive
and non-virulent. These sudden falls are later more precarious than diseases like
heart attack, blood sugar, blood pressure because these can be untreated for a
lengthy time which can lead to death. Elder citizen who experiences a precipi-
tous fall, carry out their communal life narrowed. Therefore, a shrewd and ad-
equate anti-fallen system is required for aiding elderly health care, specifically
to those who live individually. So, it can identify and anticipate a precipitous
fall through appropriate human activity recognition. In this study, we have sug-
gested an end-edge-cloud based wearable EdgeFall architecture for elderly care.
We have performed simulation setups to clarify the query of why we need such
a strategy, and its validity. We have achieved maximum 91.87% accuracy with
1.6% false alarm rate (FAR). These empirical results indicate the superiority of
using tightly couple multiple information for recognizing human activity. We
can accomplish a low FAR with an enhanced accuracy. We can observe that our
proposed end-edge-cloud based architecture can reduce the execution time to
millisecond range (ms) of 14.16 to 15.74. This work serves as the starting mark
for future related research activities.
This is an open access article under the CC BY-SA license.
Corresponding Author:
Md. Salah Uddin Yusuf
Department of Electrical and Electronic Engineering, Khulna University of Engineering and Technology
Khulna, Bangladesh
Email: suyusuf@eee.kuet.ac.bd
1. INTRODUCTION
Any abnormality in everyday human activity brings an unusual fall. An adequate routine human
activity recognition can identify, anticipate and inhibit many diseases. Walking, running (jogging), jumping,
walking stairs, sitting, standing, and laying are closely related to elderly fall-related service [1]–[3]. Besides,
human activity recognition is an elemental part of most explorations concerned to fall detection and prediction
[3]. That is why we choose to lead our investigation with human activity recognition from the context of fall
detection and prediction.
We can express the interpretation of fall as the sudden change of state from a higher post ion like
standing to a lower position like laying or sitting [4]. People over 65 years of age are at significant risk
of falling, corresponding to many published articles. A survey from the World Health Organization (WHO)
shows that there are around 650 thousand injuries occur that are linked to falls every year [1], [5]–[7]. These
falls induce curable or incurable injuries from flesh cramping, fractured extremities, posterior damage, and
brain injury to life expiry. Every inmate has to put in an abundance of payment in medicine, therapy, and
treatment [4], [8]–[10]. The after-effects of falls are likewise severe as they lead to falls, depression, restraint
Journal homepage: http://ijece.iaescore.com
4722 ❒ ISSN: 2088-8708
of regular living movements, fewer personal tasks, and poorer conditions of life, specifically for indepen-
dent elderly citizens [8]. A surprisingly striking factor is the aged populations will explode in the upcoming
20 years [11]. A survey was done on people equal to and over 65 years to answer the requirement of a fall-
related service in [12]. The number of participants was 125. The analysis shows that falling on the floor for
more than an hour leads to death within a six months period [12]. Over eight hundred thousand old individu-
als, 65 years or above, are using medical alert systems like MobileHelp, Bay alarm medical, LifeFone, Philips
Lifeline, and QMedic, with the lowest cost of USD 19.99 [13]. These services offer wearable or mobile devices
to bring mobility to the users. Besides the subscription fees, these devices still require user aid for pressing the
SOS button during an accident. This is the communal demand to promote a dynamic solution that can predict
and detect a precipitous fall. We crave to build up such a structure that can be self-employed in a necessary and
affordable.
Now, we recommend the interpretation of a wearable smart anti-fall from the point of its performance
and the opportunity of fitting into the smart environment. We can suggest the definition as a wearable device
capable of predicting and detecting an abrupt fall. For example, the system designed and developed in [14]
which can detect and predict a sudden fall while Mauldin et al. [15], Naihan et al. [16], [17] illustrate the
opportunity of enforcing end-edge-cloud approach for fall detection. So, we should develop the architecture
with various sensors like wearable cameras, accelerometers, gyroscopes, global positioning system (GPS), and
alarms. The functions of such architecture should not restrict to predicting and detecting falls but continu-
ing the auxiliary functions like position detection and sending alarms and messages to the involved parties.
Comfortable to carry, small battery utilization and significant reliability are the elementary properties of such
a system. Any wearable platform like spectacles on the eyes, helmet on the head, and gear around the waist
can build up an anti-fall system. This system should be transparent to fuse with a smart environment like smart
health care and smart home system [18], [19]. This intelligent system can be an integral part of information
and communications technology (ICT)-based healthcare solutions for elderly people in the smart home concept
similar to [20]. It is significant to define two more terms which are:
− Fall prediction: A process that forecasts a sudden fall and tries to restrain it. It provides a clue to the user
about an abrupt fall in progress so that the user can become alert and escape the fall for further emotional
and health consequences.
− Fall detection: A process that identifies a fall and gives alarms to the corresponding third parties. The third
parties include the relative of the sufferers, caretakers and near medical amenities so that the effect of falls
can be diminished and the medication can be set up as promptly as possible.
So, a robust prediction process is required before falling down, and a detection process is after a
fall occurs. In [7], the authors reviewed sensors, algorithms, and performance on the 20 most authoritative
and often cited papers among 6,830 papers with keywords linked to falling detection. The report reveals that
an accelerometer is usually employed for data collection, machine learning-based algorithms have become
mainstream in this area, and accuracy is the key performance evaluation factor. There is a positive sugges-
tion of applying camera and accelerometer for data collection i,e, data fusion. This study shows nothing
about the impact of false alarm rate (FAR) on fall detection or prediction. While the review on fall detection
in [21], we see that wearable-based development is more suitable for real-world implementation. Researches
like Saadeh et al. [14], Mualdin et al. [15], Rachakonda et al. [17], Ozcan et al. [22], Micucci et al. [23],
George et al. [24], Davide et al. [25], Reyes-Ortiz et al. [26], Li et al. [27], Howcroft [28], Engel and Ding [29],
Wang et al. [30], and others have developed fall detection systems with a wearable camera, kinit camera sen-
sors, accelerometer, gyroscope, internal measurement unit (IMU), Wi-Fi system with either machine-learning
interfaced or simply threshold-based techniques. Almost all instances apply experimental setup to classify a
binary type grouping, i,e, defining fall or not fall. Classical machine learning techniques like support vector
machine (SVM), k- neighbor nearest (k-NN), and artificial neural network (ANN), are not up to the level for
defining different human activities of daily living (ADL) [23]. We have noticed that threshold-based algorithms
can lead to a false alarm in recognizing multiclass human activity through our own empirical results [31]. In
the case of fall prediction, a few researches like Li et al. [27], Engel and Ding [29], Otanasap [32], Tong et al.
[33] have developed methods using wearable and ambient devices with both dynamic threshold and machine
learning techniques. These algorithms can maintain adequate lead time to the users before falling down so they
can be aware. Therefore the algorithms developed strategies for such kind of employment should recognize
human activities to separate them from falls.
Int J Elec & Comp Eng, Vol. 13, No. 4, August 2023: 4721-4733
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 4723
One of the dominant concerns regarding these developments is the on-chip or on-board system, which
have defined power sources and computational capacity. It is worthwhile to investigate end-edge-cloud-based
architecture for prediction. Because this architecture can hold larger and perpetual power sources and compu-
tational capacity. Therefore, our technical interest in this work is to reinforce the accuracy, lower the FAR, and
raise the performance of the classification network so it can realize different human activities, lower latency in
classifying tasks, less power utilization, and adequate lead time prediction. In this paper, we are mainly focus-
ing on the classification process. We crave to estimate the usefulness of end-edge-cloud architecture using data
fusion and modern machine learning approaches in ADLs recognition. Because this recognition is the initial
stride toward fall detection and prediction. Thus, we propose EdgeFall, an end-edge-cloud-based architecture
where different data sources are fused, and a recurrent neural network (RNN) approach will analyze human
activity. We have set up our simulation through four datasets which are UniMIB SHAR [23], MobiAct [24],
UCI HAR [25] and HAPT [26] for demonstrating the potency of data fusion in human activity recognition
(summary is given in Table 1). We can reproduce various daily activities of a group of volunteers with these
datasets. We have further studied the classic machine learning-based activity recognition model with recurrent
neural network (RNN) based model on these datasets. The performance evaluating metrics will help us design
human activity relevant to the real world.
Table 1. Characteristics of considered public datasets
Dataset UniMIB SHAR [23] MobiAct [24] UCI HAR [25] HAPT [26]
Size of original data 255 MB 1.24 GB 241 MB 7.54 KB
Representation Both ADL and fall Both ADL and fall Daily Human Daily Human
activities activities
No. of sensors 1 3 2 2
Sensors used Accelerometer Accelerometer, Accelerometer and Accelerometer and
Gyroscope and Gyroscope Gyroscope
Orientation sensor
Device used to Samsung Galaxy Nexus Samsung Galaxy S3 Samsung Galaxy S II Samsung Galaxy S
collect data I9250 with Bosh with LSM330DLC II with HARApp
BMA220 module
Position of the Trouser front pocket (half Not Specified Waist-mounted Different body parts
device of the time in the left one (trunk, upper and
and the remaining time lower extremities)
in the right one)
No of Subjects 30 (24 Females, 6 males) 57 (15 Females, 30 (the number of 17 (the number of
42 males) female and male is) female and male is)
not specified) not specified)
Physical Age: 18-60, Age: 20-47, Age: 19-48 Not specified
information about Height: 160-190, Height: 160-189,
the subjects Weight: 50-82 Weight: 50-120
No. of activities 9 types of ADL and 9 types of ADL and 6 type of daily 12 type of daily
8 type of fall 4 type of fall activities activities
Nature of data Raw accelerometer Raw accelerometer, Extracted feature Feature from
data gyroscope and accelerometer and accelerometer and
orientation data gyroscope data gyroscope data
Sampling frequency 50 Hz 20 Hz 50 Hz 50 Hz
in original dataset
Samples in modified 4,632 67,348 10,299 10,411
dataset
Depicted activities walk, walk up-stair, walk, walk down-stair, walk, walk up-stair, walk, walk up-stair,
walk down-stair, sitting, sitting, standing, walk down-stair, sitting, walk down-stair, sitting,
standing, laying down jogging standing, laying down standing,laying down
Signal window 3 second 1 second 1.5 second 1.5 second
These results likewise indicate how adequately EdgeFall architecture can analyze different human
activities. We observe from the assessment that the model developed by ActDec-SysOpt outperforms classical-
based recognition models. So, our significant contributions are:
− Designing and resembling the end-edge-cloud-based EdgeFall architecture to figure out the usefulness in
human activity detection.
EdgeFall: a promising cloud-edge-end architecture for elderly fall care (Kazi Md Shahiduzzaman)
4724 ❒ ISSN: 2088-8708
− Spot the challenges like FAR, accuracy, precision, and recall in loosely coupled data fusion for recognizing
human activities.
− Performance improvement in classifying and realizing human daily activity through tightly coupled multiple
information sources and ActDec-SysOpt algorithm.
The remaining paper dwells on four more sections. The next section will present the proposed Edge-
Fall paradigm for activity recognition. Algorithms used for this application are depicted in section 4 with
respective flowcharts. The EdgeFall is simulated using the ActDec-SysOpt algorithm for human activity detec-
tion and described there. Then, the empirical setups for evaluating EdgeFall architecture with data fusion are
analyzed in section 5. The performance evaluating metrics are also enlisted in this section. Research findings
and highlights on prospective research directions are mentioned in the conclusion.
2. EdgeFall: THE PROPOSED ARCHITECTURE
The proposed EdgeFall system is a three-tiered system as depicted in Figure 1 which are the artificial
intelligence (AI)-aware medical cloud for model development, the mobile edge for human activity recognition,
user identification, fall detection, prediction, and the body area network (BAN) for smart sensing and command
execution. The proposed architecture makes the fall detection and prediction system into three distributed sub-
systems connected to each other by the wireless network. We expect this architecture to be efficient regarding
computational capacity and energy management. BAN is indeed a platform consists sensors like wearable
cameras, gyroscopes, accelerometers, GPS sensors, and other related sensors and is pictured as a wearable
glass, but we can accept any platform like a watch, wearable around the waist, helmet on the head, smart-
phone on the pocket or similar. This layer performs data collection, feature extraction, communication to the
edge node, and decision execution. The mobile edge layer consists of a rational edge node and contrasting
processing units. User identification, activity recognition, lead time calculation for fall prediction, and fall
detection are major responsibilities of this tier. It will adapt protocols, classification networks, and users’ daily
human activity patterns from the cloud. The last tier is the cloud, the intellect of this technique. Deep learning,
special machine learning approach, and human intervention (HI) are used to implement this tier. It develops a
user identification model and activity classification model. It should update the mobile edge node with updated
reformed system parameters in a frequent cycle. This architecture can again refer to diverse areas related to
health care.
Figure 1. Proposed architecture of EdgeFall architecture
In this script, we will present the essence of this recommended architecture by breaking up the com-
putational capacity among mobile edge nodes and the cloud. Here, we have simply focused on various human
activity recognition. Because it will contribute to strong fall detection and prediction.
Int J Elec & Comp Eng, Vol. 13, No. 4, August 2023: 4721-4733
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 4725
3. PROPOPSED ALGORITHM
This section will discuss the methodologies named FuseFall and ActDec-SysOpt algorithm. These two
approaches will demonstrate the validity of data fusion, cloud-edge-end type architecture, and the classification
algorithm based on recurrent neural networks. This section will help the readers to visualize the approaches
used in this paper.
3.1. FuseFall algorithm
Fall detection takes advantage of two sources (a wearable camera and a tri-axial accelerometer) of
information. Two different sources detect human activity and fall separately. The block diagram in Figure 2(a)
explains the FuseFall algorithm. In this algorithm, the information sources are loosely coupled and analyzed
separately. We can observe that both sources use a different suitable algorithm to define the same activity
separately and compare it to the final decision from the block diagram, as shown in Figure 2(a). This algorithm
is explained in [31]. Two separate processes are used for two data types to reduce the complexity. Both
processes run simultaneously but on two different platforms. We assume that the resource required to analyze
inertia sensors data is lesser than image processing. So, the inertia sensor’s data processing process will run in
the edge node, while image processing will occur in the cloud. An optimization algorithm will be necessary to
enhance the efficiency of the system. In the detection phase, inertia data will be analyzed to detect a fall using
SVM machine learning technique. The reason behind using SVM technique is to perform binary classification
between AD and fall. A signal window of 3 seconds is considered to determine an event. In the case of
detection, image processing techniques are used to determine the feature (i,e, dissimilarity distance between
odd frames) from the captured images between 1.5 seconds backward and 1.5 seconds forward from the instant
of fall occurrence. This dissimilarity distance will verify the fall event determined in the detection phase. The
overall fall detection algorithm is shown in algorithm 1. Finally, another process will be used to optimize
the system presented in algorithm 2. This process automatically selects the optimal trained model and system
parameters to overwrite the old network.
3.2. ActDec-SysOpt algorithm
Figure 2(b) displays the block diagram of ActDec-SysOpt. We train the classification model with long
short term memory network (LSTM). This algorithm includes two parts. We adopt ActDec to recognize the
human activity, whereas SysOpt is for training classification networks. These two parts also depict mobile edge
nodes and medical cloud appropriately. We explain the details in [34].
(a) (b)
Figure 2. Block diagram of proposed algorithms (a) block diagram of FuseFall algorithm for fall detection and
(b) simple block diagram of ActDec-SysOpt algorithm
EdgeFall: a promising cloud-edge-end architecture for elderly fall care (Kazi Md Shahiduzzaman)
4726 ❒ ISSN: 2088-8708
Now the information sources are tightly coupled. This algorithm represents a centralized trained
model of human activity recognition for all users. SysOpt generates the trained classification model (ActDec)
on the cloud server and pushes it to the edge server. A multivariate time series signal is fed to the bi-directional
LSTM. A signal window of 1.5 seconds is used here. A softmax layer matches the best probability fit for a
certain human activity. With a fully connected network, each human activity model is trained in the cloud layer.
The point should be mentioned that the SysOpt algorithm optimizes the parameters like learning rate, epoch,
and number of layers. According to the incoming data and the patient’s nature of performing the activities. This
trained network is then passed to the edge layer, where the incoming pre-processed raw data from the edge layer
will be analyzed to determine the particular activities and types of falls. The overall activity detection algorithm
is shown in algorithm 3. Finally, another algorithm will be used to initialize and optimize the system presented
in algorithm 4. The test data will be labelled and added to the training data to train the even detection network.
If the result of the newly trained system is better than the old one, the parameter e,g, threshold value used in the
verification phase will be changed. Otherwise, the old parameters are kept. The fall detection and verification
algorithm runs in an intelligent access point, whereas the optimization algorithm and process data will be stored
in the cloud.
Algorithm 1: Detection process of FuseFall algorithm
Detection Phase
Data: Inertia Sensor’s Data
Result: Detection of an activity
Load SVM Model
initialization
while Unless detecting a Fall do
Extract feature F(accl) from signal window
%3 sec data
if F(accl) ∼ ADL then
go to initialization
else
if F(accl) ∼ Fall then
Verification Phase
end
end
end
Verification Phase
Data: Image data
Result: Verification of Fall
while Fall Detected do
Load images between t-1.5 to t+1.5
%Images correspond to 3 sec
Extract feature I(dissimilarity)
if I(dissimilarity) ∼ Fall then
Fall verified
else
Consider as ADL
end
end
Algorithm 2: FuseFall optimization process
Data: Test Data
Result: Optimization of the System
begin
Test data −→ Training Data
Evaluate Both trained model
if ResultNew > ResultOld then
Pass Paranew to Detection process
Modify Both Phase
else
Keep previous State
Remove Test data
end
end
Int J Elec & Comp Eng, Vol. 13, No. 4, August 2023: 4721-4733
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 4727
Algorithm 3: ActDect algorithm
Data: Inertia sensor’s data
Result: Event Detection
Load Trained LSTM Classification Model
initialization
while Unless detecting an event do
Extract feature F(accl & gyro) from signal window
%1.5 sec data
if F(accl & gyro) ∼ Daily Activity then
go to initialization
else
if F(accl & gyro) ∼ Fall then
Alarm, Protection
end
end
end
Algorithm 4: SysOpt algorithm
Data: Test Data
Result: Optimization of detection system
begin
Test data −→ Training Data
Run LSTM with TrainingDataNew
if V alidationNew > V alidationOld then
Modify Classification Network
Pass Paranew to Event Detection Algorithm
else
Keep previous State
Remove Test data
end
end
4. EVALUATION
This section presents our experiments and results using FuseFall and ActDec-SysFall architecture.
The findings will justify our claims outlined in 1. First, the performance evaluating metrics will be deliberated,
and results and findings from studied algorithms will be later.
4.1. Metrics
Accuracy, recall, precision, F1-score, FAR, training time, and execution time are the performance
evaluation metrics. The definitions are given below:
− Accuracy is stated precisely the percentage of correctly detected activities from the tested one.
− Recall is interpreted as the percentage of proper activities that are exactly verified. It provides information
about the accuracy row-wise of the confusion matrix.
− Precision is elucidating as the segment of equitable classified activities i,e, accuracy along column wise of
the confusion matrix.
− F1 score is the harmonic mean of precision and recall, which measures the accuracy of the test.
− False alarm rate (FAR) acquaints with how the system gets confused between any activity and other con-
sidered activities.
− Training-time is the time required by the cloud to generate a global classification network with the proposed
algorithm.
− Execution time is the time the global classifier recognizes human activities at the mobile edge node.
4.2. Experiments with FuseFall
We have simulated and tested loosely coupled two different data in this experiment. The results are
generated by the wearable camera y1 and the raw accelerometer data y2 separately at this stage, as shown
in Figure 3(a). This experiment demonstrates the real-life daily human activities and falls related to elderly
people.
EdgeFall: a promising cloud-edge-end architecture for elderly fall care (Kazi Md Shahiduzzaman)
4728 ❒ ISSN: 2088-8708
4.2.1. Experimental setup and results
Figure 3(a) presents the simulation scheme. A forward fall will take place from the walking position.
We present the details about how to evaluate it in [31]. We calculate the accuracy, precision, recall, F1-score,
and execution time of y2 from 4.1. We consider the k-NN method for comparison. The results are presented
in Table 2. We have utilized the frontal camera of Xiaomi S2 around the head height to perform the same
activities by taking video, similarly to [34]. As illustrated in Figure 3(b), the dissimilarity distance for the
video frames during the fall is considerably larger than other frames, marked by two red circles. From the
binary classification between falls and daily human activity, we can determine the effectiveness of FuseFall.
Nevertheless, the challenge persists in performing this FuseFall with multi-class classification to recognize
the daily human activity. Please point out to [34] what we mean by multi-class activity. Table 3 gives the
performance of the FuseFall algorithm for accelerometer data with k-NN. During video processing, we have
observed high picks like Figure 3(b), which can cause fall alarms. Table 3 and Figure 3(b) show that a loosely
coupled information source with the FuseFall algorithm is unsuitable for multi-class activity detection. That
is why we hold our experiments with y1 and y2. Besides, the time involved in generating results with video
processing is large in our experiments. The challenges ascribed to the multi-class activity classification are
managed more accurately by the ActDec-SysOpt algorithm [34], which will be demonstrated in the next part.
(a)
(b)
Figure 3. Experimental outcome of FuseFall (a) experimental setup to evaluate FuseFall algorithm and
(b) FAR during multiclass activity detection with FuseFall (video processing part)
Table 2. Partial evaluation performance of FuseFall to differentiate fall from daily activities
Metrics FuseFall for y2 k-NN
Accuracy 0.9667 0.900
Recall 0.9667 0.900
Precision 0.9839 0.9545
F1-score 0.9752 0.9265
FAR 0.033 0.10
Execution time (sec) 2.282 1.043
Int J Elec & Comp Eng, Vol. 13, No. 4, August 2023: 4721-4733
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 4729
Table 3. Performance of FuseFall for multi-class type human activity recognition
Metrics FuseFall for y2 k-NN
Accuracy 0.8468 0.8022
Recall 0.8198 0.8222
Precision 0.8518 0.8595
F1-score 0.8355 0.8405
False alarm rate (FAR) 0.0396 0.0547
4.3. Experiments with ActDec-SysOpt
As we have previously noticed, that single source of information and classical machine learning-based
models are not up to the adequate standard for realizing different human activities. Here we will adopt ActDec-
SysOpt [34] algorithm and data fusion technique to justify some remarkable objectives, i) demand of proposed
EdgeFall as shown in Figure 1, ii) the performance of multiple sources, use of features and LSTM, and iii) the
prerequisite of tight coupling.
4.3.1. Experimental setup
For the experimental setup we have utilized four original data sources, named, UniMIB SHAR [23],
MobiAct [24], UCI HAR [25] and HAPT [26] to simulate BAN of the EdgeFall. These sources consist of
diverse human activities for several volunteers. We have transformed the original datasets in such a manner
that it mirrors the BAN of the proposed architecture which relays data to the edge node regularly in time
windows. The preliminary results will confirm the potency of tightly coupled data fusion. ActDec represents
the mobile edge node while SysOpt is for the medical cloud. Here, we want to assess how dramatically this
algorithm can recognize different human activities i,e, the classification process.
4.3.2. Results
First, we will present the operation of EdgeFall through ActDec-SysOpt algorithm. The supporting
chart (Table 4 accumulates the simulation result which outperforms classical machine learning type algorithm.
We have considered a traditional data split (70-30) between training and test data. Results indicate that ActDec-
SysOpt outperforms the other two machines learning-based classification models (where one of them is Fuse-
Fall). Accordingly, we will weigh the performance of ActDec-SysOpt on different datasets. These different
datasets represent BAN of the recommended EdgeFall architecture (i,e, several users with unfamiliar weights,
heights, ages, and gender).
Table 4. Performance of ActDec-SysOpt over FuseFall and k-NN algorithm
Methods FuseFALL k-NN (k=2) ActDec-SysOpt
Accuracy 0.8468 0.8022 0.9187
Recall 0.8198 0.8222 0.8303
Precision 0.8518 0.8595 0.8511
F1-score 0.8355 0.8405 0.8406
FAR 0.0396 0.0547 0.025
Training-time (minute) 2.282 1.043 112
To conduct the simulation further sensibly, we have considered a 50-50 approach for training and test
data to appraise the validity of data fusion. Figure 4 displays the simulation results and reveals that data fu-
sion from numerous sources with proposed ActDec-SysOpt have stronger achievement than information from
a single source and loosely coupled multiple information sources. We can observe that feature extraction can
develop a far better result than raw data fusion. The accuracy of raw data fusion is remarkably unsatisfac-
tory around 59.65% with a considerable FAR 10.66%. Whereas the feature extraction (tightly coupled) from
multiple information sources forms a sharp accuracy of 92.01% with FAR of 1.69%.
We present the expected training time and classification execution time in Table 5. The time of training
classification network and execution time to realize an activity is higher for tightly coupled multiple information
sources. As we have divided the classification model development and execution process into cloud and edge,
the execution time may not be a perturbing factor. This time can be boosted by adopting a hardware/software
co-design concept.
EdgeFall: a promising cloud-edge-end architecture for elderly fall care (Kazi Md Shahiduzzaman)
4730 ❒ ISSN: 2088-8708
Figure 4. Advantages of data fusion from multiple information sources
From Tables 4 and 5, Figure 4 we come up with the following decisions:
− Tightly coupled information sources with an LSTM-based classification model are more adequate. The
accuracy is around 92% with FAR being 1.6%. The high value of recall, precision, and F1-score confirm
that the classification model is competent in recognizing human activity precisely.
− The execution time compelled to recognize activity from one window of information is noticeably less,
around 15 ms with data fusion from multiple sources (check execution time of Table 4.
− RNN type machine learning-based algorithms are better applicable to recognizing human daily activities.
− Features are better reasonable as they can decrease the FAR. 1.69% with HAPT and 1.87% with UCI HAR.
− However, we can learn that training the classification model requires a lot of time, 1,094 minutes for HAPT
and 691 minutes for UCI HAR. Medical cloud is good enough to deal with the training process because
there are no issues with the computational capability and cache.
− The train network later can be passed to the edge node which can execute activity recognition.
− Data collection and feature extraction can be managed conveniently in BAN.
Table 5. Effectiveness of EdgeFall architecture in computation
Methods UniMIB SHAR Mobi Act HAPT UCI HAR
Training-time (minute) 84 161 1094 691
Execution time (ms) ∼ 4.52 - 7.065 ∼ 0.6 - 0.73 ∼ 14.16 - 15.74 ∼ 15.20 - 16.71
So far, ActDec-SysOpt is a definite stride toward fall prediction and detection using EdgeFall. Nev-
ertheless, there are some challenges that need to figure out. For example, accuracy, FAR, a simple algorithm
for fall prediction and detection, and user identification. Though, we are a little behind in accomplishing the
coveted accuracy of over 95%.
5. CONCLUSION
In this study, we have recommended a cloud-edge-end architecture, EdgeFall, which can distribute
the full function of a single board structure to three tiers. A particular contribution of this script is that we
can uphold our claims through proper simulation results. We have sought to figure out challenges related to
establishing this system through appropriate experimental results. The ActDec-SysOpt algorithm represents
not only the EdgeFall architecture but also validates the potency of such architecture. The ActDec-SysOpt
Int J Elec & Comp Eng, Vol. 13, No. 4, August 2023: 4721-4733
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 4731
algorithm also opens up the opportunity for exploiting data fusion for this system for further enhancements in a
term of accuracy and FAR for human activity recognition. We are able to obtain 92% accuracy while lowering
the FAR to 1.69%. Therefore, this EdgeFall architecture can be regarded as being beneficial for reducing
computational capacity and power utilization on BAN, reinforcing the classification performance by specific
recognition of an activity. This architecture can also be useful in the management of mobile edge nodes, user-
profile management (patient-specific), and quality of service (QoS) which are our future research directions.
A suitable single algorithm requires to promote which can predict and detect falls. Visual inertial odometry-
based algorithms may have the capacity because they can produce trajectory. User identification is another issue
that opens a possible research trend. We can develop a modern machine learning-based algorithm to solve it
with privacy protection measures. Another prospective future task for the smart anti-fall system is to build
up a secure, reliable, low latency-based network. Software-defined network (SDN) can be the most practical
solution for not only meeting the QoS requirements but also providing security to the network. Speeding
up the execution time is another demand in implementing this EdgeFall architecture into a prototype device.
Hardware-software co-design approach can overcome this challenge. This paper not only reviews but also
evaluates different approaches toward developing smart anti-fall systems. Concepts, algorithms, experimental
setups, results and decisions, provide comprehensive guidance for deeper investigation, particularly in this
emerging area.
ACKNOWLEDGEMENT
This project is partially funded by Jatiya Kabi Kazi Nazrul Islam University’s Annual (2021-22) Re-
search Grunt.
REFERENCES
[1] Z. Wang, V. Ramamoorthy, U. Gal, and A. Guez, “Possible life saver: a review on human fall detection technology,”
Robotics, vol. 9, no. 3, Jul. 2020, doi: 10.3390/robotics9030055.
[2] E. C. Dinarevic, J. B. Husic, and S. Barakovic, “Issues of human activity recognition in healthcare,”
in 2019 18th
International Symposium INFOTEH-JAHORINA (INFOTEH), Mar. 2019, pp. 1–6, doi: 10.1109/IN-
FOTEH.2019.8717749.
[3] J. A. Álvarez-Garcı́a, L. M. Soria Morillo, M. Á. Á. de La Concepción, A. Fernández-Montes, and J. A. O. Ramı́rez,
“Evaluating wearable activity recognition and fall detection systems,” in IFMBE Proceedings of the European
Conference of the International Federation for Medical and Biological Engineering, vol. 45, 2015, pp. 653–656.
[4] K. Chaccour, R. Darazi, A. H. El Hassani, and E. Andres, “From fall detection to fall prevention: a
generic classification of fall-related systems,” IEEE Sensors Journal, vol. 17, no. 3, pp. 812–822, Feb. 2017,
doi: 10.1109/JSEN.2016.2628099.
[5] WHO, “Fall,” World Health Organization, 2018. https://www.who.int/news-room/fact-sheets/detail/falls (accessed
Aug. 20, 2018).
[6] S. Liang, Y. Liu, G. Li, and G. Zhao, “Elderly fall risk prediction with plantar center of force using ConvL-
STM algorithm,” in 2019 IEEE International Conference on Cyborg and Bionic Systems (CBS), 2019, pp. 36–41,
doi: 10.1109/CBS46900.2019.9114487.
[7] T. Xu, Y. Zhou, and J. Zhu, “New advances and challenges of fall detection systems: a survey,” Applied Sciences,
vol. 8, no. 3, Mar. 2018, doi: 10.3390/app8030418.
[8] R. Rajagopalan, I. Litvan, and T.-P. Jung, “Fall prediction and prevention systems: recent trends, challenges, and
future research directions,” Sensors, vol. 17, no. 11, Nov. 2017, doi: 10.3390/s17112509.
[9] T.R. Frieden, D. Houry, G. Baldwin, A. Dellinger, R. Lee, “Preventing falls: a guide to implementing effective
community-based fall prevention programs,” National Center for Injury Prevention and Control of the Centers Disease
Control and Prevention, 2015.
[10] J. Hamm, A. G. Money, A. Atwal, and I. Paraskevopoulos, “Fall prevention intervention technologies: A conceptual
framework and survey of the state of the art,” Journal of Biomedical Informatics, vol. 59, pp. 319–345, Feb. 2016,
doi: 10.1016/j.jbi.2015.12.013.
[11] “Global Ageing,” ageinternational, 2018. https://www.ageinternational.org.uk/policy-research/statistics/global-
ageing/ (accessed Nov. 22, 2018).
[12] L. Montanini, A. Del Campo, D. Perla, S. Spinsante, and E. Gambi, “A footwear-based methodology for fall detec-
tion,” IEEE Sensors Journal, vol. 18, no. 3, pp. 1233–1242, Feb. 2018, doi: 10.1109/JSEN.2017.2778742.
[13] A. Clark, “The best medical alert systems 2019,” TheSeniorList, Sep. 2019. https://www.theseniorlist.com/medical-
alert-systems/best/ (accessed Sep. 21, 2018).
EdgeFall: a promising cloud-edge-end architecture for elderly fall care (Kazi Md Shahiduzzaman)
4732 ❒ ISSN: 2088-8708
[14] W. Saadeh, S. A. Butt, and M. A. Bin Altaf, “A patient-specific single sensor IoT-based wearable fall predic-
tion and detection system,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 27, no. 5,
pp. 995–1003, May 2019, doi: 10.1109/TNSRE.2019.2911602.
[15] T. R. Mauldin, M. E. Canby, V. Metsis, A. H. H. Ngu, and C. C. Rivera, “SmartFall: a smartwatch-based fall detection
system using deep learning,” Sensors, vol. 18, no. 3363, Oct. 2018, doi: 10.3390/s18103363.
[16] M. J. Al Nahian et al., “Towards an accelerometer-based elderly fall detection system using cross-disciplinary time
series features,” IEEE Access, vol. 9, pp. 39413–39431, 2021, doi: 10.1109/ACCESS.2021.3056441.
[17] L. Rachakonda, S. P. Mohanty, and E. Kougianos, “Good-eye: a device for automatic prediction and detection of
elderly falls in smart homes,” in 2020 IEEE International Symposium on Smart Electronic Systems (iSES) (Formerly
iNiS), Dec. 2020, pp. 202–203, doi: 10.1109/iSES50453.2020.00051.
[18] Y. Zhang, M. Qiu, C.-W. Tsai, M. M. Hassan, and A. Alamri, “Health-CPS: healthcare cyber-physical
system assisted by cloud and big data,” IEEE Systems Journal, vol. 11, no. 1, pp. 88–95, Mar. 2017,
doi: 10.1109/JSYST.2015.2460747.
[19] A. Paneerselvam, R. Yaakob, T. Perumal, and E. Marlisah, “Fall detection framework for smart home,”
in 2018 IEEE 7th
Global Conference on Consumer Electronics (GCCE), Oct. 2018, pp. 351–352,
doi: 10.1109/GCCE.2018.8574617.
[20] N. Fares, R. S. Sherratt, and I. H. Elhajj, “Directing and orienting ICT healthcare solutions to address the needs of
the aging population,” Healthcare, vol. 9, no. 2, Feb. 2021, doi: 10.3390/healthcare9020147.
[21] S. Chaudhuri, H. Thompson, and G. Demiris, “Fall detection devices and their use with older adults:
a systematic review,” Journal of Geriatric Physical Therapy, vol. 37, no. 4, pp. 178–196, Oct. 2014,
doi: 10.1519/JPT.0b013e3182abe77.
[22] K. Ozcan, S. Velipasalar, and P. K. Varshney, “Autonomous fall detection with wearable cameras by using relative
entropy distance measure,” IEEE Transactions on Human-Machine Systems, vol. 47, no. 1, pp. 31–39, Feb. 2017,
doi: 10.1109/THMS.2016.2620904.
[23] D. Micucci, M. Mobilio, and P. Napoletano, “UniMiB SHAR: a dataset for human activity recognition using acceler-
ation data from smartphones,” Applied Sciences, vol. 7, no. 10, Oct. 2017, doi: 10.3390/app7101101.
[24] V. George, C. Chariklei, M. Thodoris, P. Matthew, and T. Manolis, “The MobiAct dataset: recognition of activities of
daily living using smartphones,” in International Conference on Information and Communication Technologies for
Ageing Well and e-Health (ICT4AWE), Apr. 2016, pp. 143–151.
[25] A. Davide, G. Alessandro, O. Luca, P. Xavier, and L. R.-O. Jorge, “A public domain dataset for human activity
recognition using smartphones,” in European Symposium on Artificial Neural Networks, Computational Intelligence
and Machine Learning (ESANN), Apr. 2013.
[26] J.-L. Reyes-Ortiz, L. Oneto, A. Samà, X. Parra, and D. Anguita, “Transition-aware human activity recognition using
smartphones,” Neurocomputing, vol. 171, pp. 754–767, Jan. 2016, doi: 10.1016/j.neucom.2015.07.085.
[27] M. Li, G. Xu, B. He, X. Ma, and J. Xie, “Pre-impact fall detection based on a modified zero moment point
criterion using data from kinect sensors,” IEEE Sensors Journal, vol. 18, no. 13, pp. 5522–5531, Jul. 2018,
doi: 10.1109/JSEN.2018.2833451.
[28] J. Howcroft, J. Kofman, and E. D. Lemaire, “Prospective fall-risk prediction models for older adults based on wear-
able sensors,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 25, no. 10, pp. 1812–1820,
Oct. 2017, doi: 10.1109/TNSRE.2017.2687100.
[29] W. Engel and W. Ding, “Reliable and practical fall prediction using artificial neural network,” in 2017 13th
Inter-
national Conference on Natural Computation, Fuzzy Systems and Knowledge Discovery (ICNC-FSKD), Jul. 2017,
pp. 1867–1871, doi: 10.1109/FSKD.2017.8393052.
[30] Y. Wang, K. Wu, and L. M. Ni, “WiFall: device-free fall detection by wireless networks,” IEEE Transactions on
Mobile Computing, vol. 16, no. 2, pp. 581–594, Feb. 2017, doi: 10.1109/TMC.2016.2557792.
[31] K. M. Shahiduzzaman, X. Hei, C. Guo, and W. Cheng, “Enhancing fall detection for elderly with smart helmet in a
cloud-network-edge architecture,” in IEEE International Conference on Consumer Electronics- Taiwan (ICCE-TW),
May 2019.
[32] N. Otanasap, “Pre-impact fall detection based on wearable device using dynamic threshold model,” in Parallel
and Distributed Computing, Applications and Technologies, Dec. 2016, vol. 0, pp. 362–365, doi: 10.1109/PD-
CAT.2016.083.
[33] L. Tong, Q. Song, Y. Ge, and Ming Liu, “HMM-based human fall detection and prediction method using tri-axial
accelerometer,” IEEE Sensors Journal, vol. 13, no. 5, pp. 1849–1856, May 2013, doi: 10.1109/JSEN.2013.2245231.
[34] K. M. Shahiduzzaman, J. Peng, Y. Gao, X. Hei, and W. Cheng, “Towards accurate and robust fall detection for the
elderly in a hybrid cloud-edge architecture,” in IEEE Smart World Congress (SmartWorld), Aug. 2019.
Int J Elec & Comp Eng, Vol. 13, No. 4, August 2023: 4721-4733
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 4733
BIOGRAPHIES OF AUTHORS
Kazi Md Shahiduzzaman received the B.Sc.Eng. degree in Electrical and Electronic
Engineering (EEE) from the Khulna University of Engineering and Technology (KUET), Bangladesh,
in 2010 and the M.Sc. from Hochschule Bremen (University of Applied Sciences Bremen), Germany,
in 2013. Now, He is pursuing a Ph.D. degree in EEE from KUET. He is an Associate Professor at the
Department of EEE, Jatiya Kabi Kazi Nazrul Islam University, Bangladesh. His research interests
include smart health care, elderly daily activity recognition, machine learning techniques, free space
optical communication, cloud-edge-end-based distributed platforms, artificial intelligent load, and
power production forecasting. His research mainly focuses on application-oriented issues. He can be
contacted at email: shahiduzzaman2103751@stud.kuet.ac.bd and shahiduzzaman@jkkniu.edu.bd.
Md. Salah Uddin Yusuf received the B.Sc., M.Sc., and Ph.D. degree in Electri-
cal and Electronic Engineering from Khulna University of Engineering and Technology (KUET),
Bangladesh, in 2001, 2005, and 2016 respectively. Since 2001, he has been with the same depart-
ment of the same University and currently serving as Professor. His research interests mainly focus
on deep learning-based signal, image, and video processing with quality assessment, Face liveness
detection and authentication, and GAN-based Human activity analysis for the modern healthcare
systems. He can be contacted at email: suyusuf@eee.kuet.ac.bd.
EdgeFall: a promising cloud-edge-end architecture for elderly fall care (Kazi Md Shahiduzzaman)

Recomendados

A FALL DETECTION SMART WATCH USING IOT AND DEEP LEARNING por
A FALL DETECTION SMART WATCH USING IOT AND DEEP LEARNINGA FALL DETECTION SMART WATCH USING IOT AND DEEP LEARNING
A FALL DETECTION SMART WATCH USING IOT AND DEEP LEARNINGIRJET Journal
3 visualizações5 slides
A DEVICE FOR AUTOMATIC DETECTION OF ELDERLY FALLS por
A DEVICE FOR AUTOMATIC DETECTION OF ELDERLY FALLSA DEVICE FOR AUTOMATIC DETECTION OF ELDERLY FALLS
A DEVICE FOR AUTOMATIC DETECTION OF ELDERLY FALLSIRJET Journal
2 visualizações4 slides
Internet of things based fall detection and heart rate monitoring system for... por
Internet of things based fall detection and heart rate monitoring  system for...Internet of things based fall detection and heart rate monitoring  system for...
Internet of things based fall detection and heart rate monitoring system for...IJECEIAES
42 visualizações13 slides
NEW CORONA VIRUS DISEASE 2022: SOCIAL DISTANCING IS AN EFFECTIVE MEASURE (COV... por
NEW CORONA VIRUS DISEASE 2022: SOCIAL DISTANCING IS AN EFFECTIVE MEASURE (COV...NEW CORONA VIRUS DISEASE 2022: SOCIAL DISTANCING IS AN EFFECTIVE MEASURE (COV...
NEW CORONA VIRUS DISEASE 2022: SOCIAL DISTANCING IS AN EFFECTIVE MEASURE (COV...IRJET Journal
8 visualizações8 slides
IRJET- Review on: A Wireless IoT System for Gait Detection in Stroke Patient por
IRJET- Review on: A Wireless IoT System for Gait Detection in Stroke PatientIRJET- Review on: A Wireless IoT System for Gait Detection in Stroke Patient
IRJET- Review on: A Wireless IoT System for Gait Detection in Stroke PatientIRJET Journal
75 visualizações6 slides
F010222529 por
F010222529F010222529
F010222529IOSR Journals
110 visualizações5 slides

Mais conteúdo relacionado

Similar a EdgeFall: a promising cloud-edge-end architecture for elderly fall care

IRJET- A Survey on Vision based Fall Detection Techniques por
IRJET-  	  A Survey on Vision based Fall Detection TechniquesIRJET-  	  A Survey on Vision based Fall Detection Techniques
IRJET- A Survey on Vision based Fall Detection TechniquesIRJET Journal
45 visualizações6 slides
IRJET- IoT based Smart Sensing Wheelchair to Assist in Healthcare por
IRJET- IoT based Smart Sensing Wheelchair to Assist in HealthcareIRJET- IoT based Smart Sensing Wheelchair to Assist in Healthcare
IRJET- IoT based Smart Sensing Wheelchair to Assist in HealthcareIRJET Journal
39 visualizações7 slides
IRJET- Development and Monitoring of a Fall Detection System through Wear... por
IRJET-  	  Development and Monitoring of a Fall Detection System through Wear...IRJET-  	  Development and Monitoring of a Fall Detection System through Wear...
IRJET- Development and Monitoring of a Fall Detection System through Wear...IRJET Journal
10 visualizações6 slides
WIISEL Final Report - 1- Publishable Report Final por
WIISEL Final Report - 1- Publishable Report FinalWIISEL Final Report - 1- Publishable Report Final
WIISEL Final Report - 1- Publishable Report FinalElisenda Reixach
247 visualizações31 slides
energies-15-00828.pdf por
energies-15-00828.pdfenergies-15-00828.pdf
energies-15-00828.pdfASHOKKUMARARJANBHAIP
6 visualizações24 slides
Design and development of fall detector using fall por
Design and development of fall detector using fallDesign and development of fall detector using fall
Design and development of fall detector using falleSAT Publishing House
488 visualizações5 slides

Similar a EdgeFall: a promising cloud-edge-end architecture for elderly fall care(20)

IRJET- A Survey on Vision based Fall Detection Techniques por IRJET Journal
IRJET-  	  A Survey on Vision based Fall Detection TechniquesIRJET-  	  A Survey on Vision based Fall Detection Techniques
IRJET- A Survey on Vision based Fall Detection Techniques
IRJET Journal45 visualizações
IRJET- IoT based Smart Sensing Wheelchair to Assist in Healthcare por IRJET Journal
IRJET- IoT based Smart Sensing Wheelchair to Assist in HealthcareIRJET- IoT based Smart Sensing Wheelchair to Assist in Healthcare
IRJET- IoT based Smart Sensing Wheelchair to Assist in Healthcare
IRJET Journal39 visualizações
IRJET- Development and Monitoring of a Fall Detection System through Wear... por IRJET Journal
IRJET-  	  Development and Monitoring of a Fall Detection System through Wear...IRJET-  	  Development and Monitoring of a Fall Detection System through Wear...
IRJET- Development and Monitoring of a Fall Detection System through Wear...
IRJET Journal10 visualizações
WIISEL Final Report - 1- Publishable Report Final por Elisenda Reixach
WIISEL Final Report - 1- Publishable Report FinalWIISEL Final Report - 1- Publishable Report Final
WIISEL Final Report - 1- Publishable Report Final
Elisenda Reixach247 visualizações
Design and development of fall detector using fall por eSAT Publishing House
Design and development of fall detector using fallDesign and development of fall detector using fall
Design and development of fall detector using fall
eSAT Publishing House488 visualizações
Design and development of fall detector using fall por eSAT Publishing House
Design and development of fall detector using fallDesign and development of fall detector using fall
Design and development of fall detector using fall
eSAT Publishing House229 visualizações
Analysis of Fall Detection Systems: A Review por ijtsrd
Analysis of Fall Detection Systems: A ReviewAnalysis of Fall Detection Systems: A Review
Analysis of Fall Detection Systems: A Review
ijtsrd81 visualizações
IoT Based Human Activity Recognition and Classification Using Machine Learning por IRJET Journal
IoT Based Human Activity Recognition and Classification Using Machine LearningIoT Based Human Activity Recognition and Classification Using Machine Learning
IoT Based Human Activity Recognition and Classification Using Machine Learning
IRJET Journal4 visualizações
Social Distancing Detection, Monitoring and Management Using OpenCV por IRJET Journal
Social Distancing Detection, Monitoring and Management Using OpenCVSocial Distancing Detection, Monitoring and Management Using OpenCV
Social Distancing Detection, Monitoring and Management Using OpenCV
IRJET Journal80 visualizações
Real-time Activity Recognition using Smartphone Accelerometer por ijtsrd
Real-time Activity Recognition using Smartphone AccelerometerReal-time Activity Recognition using Smartphone Accelerometer
Real-time Activity Recognition using Smartphone Accelerometer
ijtsrd20 visualizações
IRJET- A Comprehensive Survey on Smart Healthcare Monitoring of Patients usin... por IRJET Journal
IRJET- A Comprehensive Survey on Smart Healthcare Monitoring of Patients usin...IRJET- A Comprehensive Survey on Smart Healthcare Monitoring of Patients usin...
IRJET- A Comprehensive Survey on Smart Healthcare Monitoring of Patients usin...
IRJET Journal55 visualizações
IRJET- A Real Time, Automated, Medical Emergency Service for Informing th... por IRJET Journal
IRJET-  	  A Real Time, Automated, Medical Emergency Service for Informing th...IRJET-  	  A Real Time, Automated, Medical Emergency Service for Informing th...
IRJET- A Real Time, Automated, Medical Emergency Service for Informing th...
IRJET Journal10 visualizações
Real Time Social Distance Detector using Deep learning por IRJET Journal
Real Time Social Distance Detector using Deep learningReal Time Social Distance Detector using Deep learning
Real Time Social Distance Detector using Deep learning
IRJET Journal42 visualizações
Fall Detection System for the Elderly based on the Classification of Shimmer ... por Moiz Ahmed
Fall Detection System for the Elderly based on the Classification of Shimmer ...Fall Detection System for the Elderly based on the Classification of Shimmer ...
Fall Detection System for the Elderly based on the Classification of Shimmer ...
Moiz Ahmed5.9K visualizações
IRJET- Disaster Detection & Early Warning System por IRJET Journal
IRJET- Disaster Detection & Early Warning SystemIRJET- Disaster Detection & Early Warning System
IRJET- Disaster Detection & Early Warning System
IRJET Journal35 visualizações
Deep Tricorder por Anand Sekar
Deep TricorderDeep Tricorder
Deep Tricorder
Anand Sekar156 visualizações
A Survey on Smart Devices for Object and Fall Detection por IRJET Journal
A Survey on Smart Devices for Object and Fall DetectionA Survey on Smart Devices for Object and Fall Detection
A Survey on Smart Devices for Object and Fall Detection
IRJET Journal5 visualizações
Complications, Complication Rates, And 90 Day Readmission... por Serena Faye
Complications, Complication Rates, And 90 Day Readmission...Complications, Complication Rates, And 90 Day Readmission...
Complications, Complication Rates, And 90 Day Readmission...
Serena Faye3 visualizações
Challenges of Data Management & Analytics in the Future Grid por Power System Operation
Challenges of Data Management & Analytics in the Future GridChallenges of Data Management & Analytics in the Future Grid
Challenges of Data Management & Analytics in the Future Grid
Power System Operation49 visualizações

Mais de IJECEIAES

Predictive fertilization models for potato crops using machine learning techn... por
Predictive fertilization models for potato crops using machine learning techn...Predictive fertilization models for potato crops using machine learning techn...
Predictive fertilization models for potato crops using machine learning techn...IJECEIAES
3 visualizações9 slides
Computer-aided automated detection of kidney disease using supervised learnin... por
Computer-aided automated detection of kidney disease using supervised learnin...Computer-aided automated detection of kidney disease using supervised learnin...
Computer-aided automated detection of kidney disease using supervised learnin...IJECEIAES
2 visualizações10 slides
Reliable and efficient webserver management for task scheduling in edge-cloud... por
Reliable and efficient webserver management for task scheduling in edge-cloud...Reliable and efficient webserver management for task scheduling in edge-cloud...
Reliable and efficient webserver management for task scheduling in edge-cloud...IJECEIAES
2 visualizações10 slides
Reinforcement learning-based security schema mitigating manin-the-middle atta... por
Reinforcement learning-based security schema mitigating manin-the-middle atta...Reinforcement learning-based security schema mitigating manin-the-middle atta...
Reinforcement learning-based security schema mitigating manin-the-middle atta...IJECEIAES
2 visualizações14 slides
Resource-efficient workload task scheduling for cloud-assisted internet of th... por
Resource-efficient workload task scheduling for cloud-assisted internet of th...Resource-efficient workload task scheduling for cloud-assisted internet of th...
Resource-efficient workload task scheduling for cloud-assisted internet of th...IJECEIAES
2 visualizações10 slides
Breast cancer classification with histopathological image based on machine le... por
Breast cancer classification with histopathological image based on machine le...Breast cancer classification with histopathological image based on machine le...
Breast cancer classification with histopathological image based on machine le...IJECEIAES
2 visualizações13 slides

Mais de IJECEIAES(20)

Predictive fertilization models for potato crops using machine learning techn... por IJECEIAES
Predictive fertilization models for potato crops using machine learning techn...Predictive fertilization models for potato crops using machine learning techn...
Predictive fertilization models for potato crops using machine learning techn...
IJECEIAES3 visualizações
Computer-aided automated detection of kidney disease using supervised learnin... por IJECEIAES
Computer-aided automated detection of kidney disease using supervised learnin...Computer-aided automated detection of kidney disease using supervised learnin...
Computer-aided automated detection of kidney disease using supervised learnin...
IJECEIAES2 visualizações
Reliable and efficient webserver management for task scheduling in edge-cloud... por IJECEIAES
Reliable and efficient webserver management for task scheduling in edge-cloud...Reliable and efficient webserver management for task scheduling in edge-cloud...
Reliable and efficient webserver management for task scheduling in edge-cloud...
IJECEIAES2 visualizações
Reinforcement learning-based security schema mitigating manin-the-middle atta... por IJECEIAES
Reinforcement learning-based security schema mitigating manin-the-middle atta...Reinforcement learning-based security schema mitigating manin-the-middle atta...
Reinforcement learning-based security schema mitigating manin-the-middle atta...
IJECEIAES2 visualizações
Resource-efficient workload task scheduling for cloud-assisted internet of th... por IJECEIAES
Resource-efficient workload task scheduling for cloud-assisted internet of th...Resource-efficient workload task scheduling for cloud-assisted internet of th...
Resource-efficient workload task scheduling for cloud-assisted internet of th...
IJECEIAES2 visualizações
Breast cancer classification with histopathological image based on machine le... por IJECEIAES
Breast cancer classification with histopathological image based on machine le...Breast cancer classification with histopathological image based on machine le...
Breast cancer classification with histopathological image based on machine le...
IJECEIAES2 visualizações
Early detection of dysphoria using electroencephalogram affective modelling por IJECEIAES
Early detection of dysphoria using electroencephalogram affective modelling Early detection of dysphoria using electroencephalogram affective modelling
Early detection of dysphoria using electroencephalogram affective modelling
IJECEIAES4 visualizações
A novel defect detection method for software requirements inspections por IJECEIAES
A novel defect detection method for software requirements inspections A novel defect detection method for software requirements inspections
A novel defect detection method for software requirements inspections
IJECEIAES2 visualizações
Convolutional auto-encoded extreme learning machine for incremental learning ... por IJECEIAES
Convolutional auto-encoded extreme learning machine for incremental learning ...Convolutional auto-encoded extreme learning machine for incremental learning ...
Convolutional auto-encoded extreme learning machine for incremental learning ...
IJECEIAES2 visualizações
Brain cone beam computed tomography image analysis using ResNet50 for collate... por IJECEIAES
Brain cone beam computed tomography image analysis using ResNet50 for collate...Brain cone beam computed tomography image analysis using ResNet50 for collate...
Brain cone beam computed tomography image analysis using ResNet50 for collate...
IJECEIAES2 visualizações
Channel and spatial attention mechanism for fashion image captioning por IJECEIAES
Channel and spatial attention mechanism for fashion image captioning Channel and spatial attention mechanism for fashion image captioning
Channel and spatial attention mechanism for fashion image captioning
IJECEIAES2 visualizações
Regional feature learning using attribute structural analysis in bipartite at... por IJECEIAES
Regional feature learning using attribute structural analysis in bipartite at...Regional feature learning using attribute structural analysis in bipartite at...
Regional feature learning using attribute structural analysis in bipartite at...
IJECEIAES2 visualizações
Adaptive traffic lights based on traffic flow prediction using machine learni... por IJECEIAES
Adaptive traffic lights based on traffic flow prediction using machine learni...Adaptive traffic lights based on traffic flow prediction using machine learni...
Adaptive traffic lights based on traffic flow prediction using machine learni...
IJECEIAES2 visualizações
Content-based product image retrieval using squared-hinge loss trained convol... por IJECEIAES
Content-based product image retrieval using squared-hinge loss trained convol...Content-based product image retrieval using squared-hinge loss trained convol...
Content-based product image retrieval using squared-hinge loss trained convol...
IJECEIAES2 visualizações
A deep convolutional structure-based approach for accurate recognition of ski... por IJECEIAES
A deep convolutional structure-based approach for accurate recognition of ski...A deep convolutional structure-based approach for accurate recognition of ski...
A deep convolutional structure-based approach for accurate recognition of ski...
IJECEIAES2 visualizações
IPv6 flood attack detection based on epsilon greedy optimized Q learning in s... por IJECEIAES
IPv6 flood attack detection based on epsilon greedy optimized Q learning in s...IPv6 flood attack detection based on epsilon greedy optimized Q learning in s...
IPv6 flood attack detection based on epsilon greedy optimized Q learning in s...
IJECEIAES2 visualizações
Analysis and prediction of seed quality using machine learning por IJECEIAES
Analysis and prediction of seed quality using machine learning Analysis and prediction of seed quality using machine learning
Analysis and prediction of seed quality using machine learning
IJECEIAES2 visualizações
Explainable extreme boosting model for breast cancer diagnosis por IJECEIAES
Explainable extreme boosting model for breast cancer diagnosis Explainable extreme boosting model for breast cancer diagnosis
Explainable extreme boosting model for breast cancer diagnosis
IJECEIAES2 visualizações
Predicting active compounds for lung cancer based on quantitative structure-a... por IJECEIAES
Predicting active compounds for lung cancer based on quantitative structure-a...Predicting active compounds for lung cancer based on quantitative structure-a...
Predicting active compounds for lung cancer based on quantitative structure-a...
IJECEIAES4 visualizações
U-Net transfer learning backbones for lesions segmentation in breast ultrasou... por IJECEIAES
U-Net transfer learning backbones for lesions segmentation in breast ultrasou...U-Net transfer learning backbones for lesions segmentation in breast ultrasou...
U-Net transfer learning backbones for lesions segmentation in breast ultrasou...
IJECEIAES2 visualizações

Último

Machine Element II Course outline.pdf por
Machine Element II Course outline.pdfMachine Element II Course outline.pdf
Machine Element II Course outline.pdfodatadese1
6 visualizações2 slides
Thermal aware task assignment for multicore processors using genetic algorithm por
Thermal aware task assignment for multicore processors using genetic algorithm Thermal aware task assignment for multicore processors using genetic algorithm
Thermal aware task assignment for multicore processors using genetic algorithm IJECEIAES
29 visualizações12 slides
performance uploading.pptx por
performance uploading.pptxperformance uploading.pptx
performance uploading.pptxSanthiS10
7 visualizações18 slides
Digital Watermarking Of Audio Signals.pptx por
Digital Watermarking Of Audio Signals.pptxDigital Watermarking Of Audio Signals.pptx
Digital Watermarking Of Audio Signals.pptxAyushJaiswal781174
8 visualizações25 slides
Investor Presentation por
Investor PresentationInvestor Presentation
Investor Presentationeser sevinç
16 visualizações26 slides
9_DVD_Dynamic_logic_circuits.pdf por
9_DVD_Dynamic_logic_circuits.pdf9_DVD_Dynamic_logic_circuits.pdf
9_DVD_Dynamic_logic_circuits.pdfUsha Mehta
21 visualizações32 slides

Último(20)

Machine Element II Course outline.pdf por odatadese1
Machine Element II Course outline.pdfMachine Element II Course outline.pdf
Machine Element II Course outline.pdf
odatadese16 visualizações
Thermal aware task assignment for multicore processors using genetic algorithm por IJECEIAES
Thermal aware task assignment for multicore processors using genetic algorithm Thermal aware task assignment for multicore processors using genetic algorithm
Thermal aware task assignment for multicore processors using genetic algorithm
IJECEIAES29 visualizações
performance uploading.pptx por SanthiS10
performance uploading.pptxperformance uploading.pptx
performance uploading.pptx
SanthiS107 visualizações
Digital Watermarking Of Audio Signals.pptx por AyushJaiswal781174
Digital Watermarking Of Audio Signals.pptxDigital Watermarking Of Audio Signals.pptx
Digital Watermarking Of Audio Signals.pptx
AyushJaiswal7811748 visualizações
Investor Presentation por eser sevinç
Investor PresentationInvestor Presentation
Investor Presentation
eser sevinç16 visualizações
9_DVD_Dynamic_logic_circuits.pdf por Usha Mehta
9_DVD_Dynamic_logic_circuits.pdf9_DVD_Dynamic_logic_circuits.pdf
9_DVD_Dynamic_logic_circuits.pdf
Usha Mehta21 visualizações
SWM L1-L14_drhasan (Part 1).pdf por MahmudHasan747870
SWM L1-L14_drhasan (Part 1).pdfSWM L1-L14_drhasan (Part 1).pdf
SWM L1-L14_drhasan (Part 1).pdf
MahmudHasan74787048 visualizações
Literature review and Case study on Commercial Complex in Nepal, Durbar mall,... por AakashShakya12
Literature review and Case study on Commercial Complex in Nepal, Durbar mall,...Literature review and Case study on Commercial Complex in Nepal, Durbar mall,...
Literature review and Case study on Commercial Complex in Nepal, Durbar mall,...
AakashShakya1245 visualizações
SWM L15-L28_drhasan (Part 2).pdf por MahmudHasan747870
SWM L15-L28_drhasan (Part 2).pdfSWM L15-L28_drhasan (Part 2).pdf
SWM L15-L28_drhasan (Part 2).pdf
MahmudHasan74787028 visualizações
LFA-NPG-Paper.pdf por harinsrikanth
LFA-NPG-Paper.pdfLFA-NPG-Paper.pdf
LFA-NPG-Paper.pdf
harinsrikanth40 visualizações
SEMI CONDUCTORS por pavaniaalla2005
SEMI CONDUCTORSSEMI CONDUCTORS
SEMI CONDUCTORS
pavaniaalla200519 visualizações
Performance of Back-to-Back Mechanically Stabilized Earth Walls Supporting th... por ahmedmesaiaoun
Performance of Back-to-Back Mechanically Stabilized Earth Walls Supporting th...Performance of Back-to-Back Mechanically Stabilized Earth Walls Supporting th...
Performance of Back-to-Back Mechanically Stabilized Earth Walls Supporting th...
ahmedmesaiaoun12 visualizações
cloud computing-virtualization.pptx por RajaulKarim20
cloud computing-virtualization.pptxcloud computing-virtualization.pptx
cloud computing-virtualization.pptx
RajaulKarim2082 visualizações
CHI-SQUARE ( χ2) TESTS.pptx por ssusera597c5
CHI-SQUARE ( χ2) TESTS.pptxCHI-SQUARE ( χ2) TESTS.pptx
CHI-SQUARE ( χ2) TESTS.pptx
ssusera597c520 visualizações
13_DVD_Latch-up_prevention.pdf por Usha Mehta
13_DVD_Latch-up_prevention.pdf13_DVD_Latch-up_prevention.pdf
13_DVD_Latch-up_prevention.pdf
Usha Mehta9 visualizações
How I learned to stop worrying and love the dark silicon apocalypse.pdf por Tomasz Kowalczewski
How I learned to stop worrying and love the dark silicon apocalypse.pdfHow I learned to stop worrying and love the dark silicon apocalypse.pdf
How I learned to stop worrying and love the dark silicon apocalypse.pdf
Tomasz Kowalczewski23 visualizações

EdgeFall: a promising cloud-edge-end architecture for elderly fall care

  • 1. International Journal of Electrical and Computer Engineering (IJECE) Vol. 13, No. 4, August 2023, pp. 4721∼4733 ISSN: 2088-8708, DOI: 10.11591/ijece.v13i4.pp4721-4733 ❒ 4721 EdgeFall: a promising cloud-edge-end architecture for elderly fall care Kazi Md Shahiduzzaman, Md. Salah Uddin Yusuf Department of Electrical and Electrical Engineering, Khulna University of Engineering and Technology, Khulna, Bangladesh Article Info Article history: Received Jul 4, 2022 Revised Oct 26, 2022 Accepted Nov 6, 2022 Keywords: Accelerometer Elderly care End-edge-cloud architecture Gyroscope Human activity Long term short memory Wearable camera ABSTRACT Elder citizens face sudden fall, which can lead to injuries of both destructive and non-virulent. These sudden falls are later more precarious than diseases like heart attack, blood sugar, blood pressure because these can be untreated for a lengthy time which can lead to death. Elder citizen who experiences a precipi- tous fall, carry out their communal life narrowed. Therefore, a shrewd and ad- equate anti-fallen system is required for aiding elderly health care, specifically to those who live individually. So, it can identify and anticipate a precipitous fall through appropriate human activity recognition. In this study, we have sug- gested an end-edge-cloud based wearable EdgeFall architecture for elderly care. We have performed simulation setups to clarify the query of why we need such a strategy, and its validity. We have achieved maximum 91.87% accuracy with 1.6% false alarm rate (FAR). These empirical results indicate the superiority of using tightly couple multiple information for recognizing human activity. We can accomplish a low FAR with an enhanced accuracy. We can observe that our proposed end-edge-cloud based architecture can reduce the execution time to millisecond range (ms) of 14.16 to 15.74. This work serves as the starting mark for future related research activities. This is an open access article under the CC BY-SA license. Corresponding Author: Md. Salah Uddin Yusuf Department of Electrical and Electronic Engineering, Khulna University of Engineering and Technology Khulna, Bangladesh Email: suyusuf@eee.kuet.ac.bd 1. INTRODUCTION Any abnormality in everyday human activity brings an unusual fall. An adequate routine human activity recognition can identify, anticipate and inhibit many diseases. Walking, running (jogging), jumping, walking stairs, sitting, standing, and laying are closely related to elderly fall-related service [1]–[3]. Besides, human activity recognition is an elemental part of most explorations concerned to fall detection and prediction [3]. That is why we choose to lead our investigation with human activity recognition from the context of fall detection and prediction. We can express the interpretation of fall as the sudden change of state from a higher post ion like standing to a lower position like laying or sitting [4]. People over 65 years of age are at significant risk of falling, corresponding to many published articles. A survey from the World Health Organization (WHO) shows that there are around 650 thousand injuries occur that are linked to falls every year [1], [5]–[7]. These falls induce curable or incurable injuries from flesh cramping, fractured extremities, posterior damage, and brain injury to life expiry. Every inmate has to put in an abundance of payment in medicine, therapy, and treatment [4], [8]–[10]. The after-effects of falls are likewise severe as they lead to falls, depression, restraint Journal homepage: http://ijece.iaescore.com
  • 2. 4722 ❒ ISSN: 2088-8708 of regular living movements, fewer personal tasks, and poorer conditions of life, specifically for indepen- dent elderly citizens [8]. A surprisingly striking factor is the aged populations will explode in the upcoming 20 years [11]. A survey was done on people equal to and over 65 years to answer the requirement of a fall- related service in [12]. The number of participants was 125. The analysis shows that falling on the floor for more than an hour leads to death within a six months period [12]. Over eight hundred thousand old individu- als, 65 years or above, are using medical alert systems like MobileHelp, Bay alarm medical, LifeFone, Philips Lifeline, and QMedic, with the lowest cost of USD 19.99 [13]. These services offer wearable or mobile devices to bring mobility to the users. Besides the subscription fees, these devices still require user aid for pressing the SOS button during an accident. This is the communal demand to promote a dynamic solution that can predict and detect a precipitous fall. We crave to build up such a structure that can be self-employed in a necessary and affordable. Now, we recommend the interpretation of a wearable smart anti-fall from the point of its performance and the opportunity of fitting into the smart environment. We can suggest the definition as a wearable device capable of predicting and detecting an abrupt fall. For example, the system designed and developed in [14] which can detect and predict a sudden fall while Mauldin et al. [15], Naihan et al. [16], [17] illustrate the opportunity of enforcing end-edge-cloud approach for fall detection. So, we should develop the architecture with various sensors like wearable cameras, accelerometers, gyroscopes, global positioning system (GPS), and alarms. The functions of such architecture should not restrict to predicting and detecting falls but continu- ing the auxiliary functions like position detection and sending alarms and messages to the involved parties. Comfortable to carry, small battery utilization and significant reliability are the elementary properties of such a system. Any wearable platform like spectacles on the eyes, helmet on the head, and gear around the waist can build up an anti-fall system. This system should be transparent to fuse with a smart environment like smart health care and smart home system [18], [19]. This intelligent system can be an integral part of information and communications technology (ICT)-based healthcare solutions for elderly people in the smart home concept similar to [20]. It is significant to define two more terms which are: − Fall prediction: A process that forecasts a sudden fall and tries to restrain it. It provides a clue to the user about an abrupt fall in progress so that the user can become alert and escape the fall for further emotional and health consequences. − Fall detection: A process that identifies a fall and gives alarms to the corresponding third parties. The third parties include the relative of the sufferers, caretakers and near medical amenities so that the effect of falls can be diminished and the medication can be set up as promptly as possible. So, a robust prediction process is required before falling down, and a detection process is after a fall occurs. In [7], the authors reviewed sensors, algorithms, and performance on the 20 most authoritative and often cited papers among 6,830 papers with keywords linked to falling detection. The report reveals that an accelerometer is usually employed for data collection, machine learning-based algorithms have become mainstream in this area, and accuracy is the key performance evaluation factor. There is a positive sugges- tion of applying camera and accelerometer for data collection i,e, data fusion. This study shows nothing about the impact of false alarm rate (FAR) on fall detection or prediction. While the review on fall detection in [21], we see that wearable-based development is more suitable for real-world implementation. Researches like Saadeh et al. [14], Mualdin et al. [15], Rachakonda et al. [17], Ozcan et al. [22], Micucci et al. [23], George et al. [24], Davide et al. [25], Reyes-Ortiz et al. [26], Li et al. [27], Howcroft [28], Engel and Ding [29], Wang et al. [30], and others have developed fall detection systems with a wearable camera, kinit camera sen- sors, accelerometer, gyroscope, internal measurement unit (IMU), Wi-Fi system with either machine-learning interfaced or simply threshold-based techniques. Almost all instances apply experimental setup to classify a binary type grouping, i,e, defining fall or not fall. Classical machine learning techniques like support vector machine (SVM), k- neighbor nearest (k-NN), and artificial neural network (ANN), are not up to the level for defining different human activities of daily living (ADL) [23]. We have noticed that threshold-based algorithms can lead to a false alarm in recognizing multiclass human activity through our own empirical results [31]. In the case of fall prediction, a few researches like Li et al. [27], Engel and Ding [29], Otanasap [32], Tong et al. [33] have developed methods using wearable and ambient devices with both dynamic threshold and machine learning techniques. These algorithms can maintain adequate lead time to the users before falling down so they can be aware. Therefore the algorithms developed strategies for such kind of employment should recognize human activities to separate them from falls. Int J Elec & Comp Eng, Vol. 13, No. 4, August 2023: 4721-4733
  • 3. Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 4723 One of the dominant concerns regarding these developments is the on-chip or on-board system, which have defined power sources and computational capacity. It is worthwhile to investigate end-edge-cloud-based architecture for prediction. Because this architecture can hold larger and perpetual power sources and compu- tational capacity. Therefore, our technical interest in this work is to reinforce the accuracy, lower the FAR, and raise the performance of the classification network so it can realize different human activities, lower latency in classifying tasks, less power utilization, and adequate lead time prediction. In this paper, we are mainly focus- ing on the classification process. We crave to estimate the usefulness of end-edge-cloud architecture using data fusion and modern machine learning approaches in ADLs recognition. Because this recognition is the initial stride toward fall detection and prediction. Thus, we propose EdgeFall, an end-edge-cloud-based architecture where different data sources are fused, and a recurrent neural network (RNN) approach will analyze human activity. We have set up our simulation through four datasets which are UniMIB SHAR [23], MobiAct [24], UCI HAR [25] and HAPT [26] for demonstrating the potency of data fusion in human activity recognition (summary is given in Table 1). We can reproduce various daily activities of a group of volunteers with these datasets. We have further studied the classic machine learning-based activity recognition model with recurrent neural network (RNN) based model on these datasets. The performance evaluating metrics will help us design human activity relevant to the real world. Table 1. Characteristics of considered public datasets Dataset UniMIB SHAR [23] MobiAct [24] UCI HAR [25] HAPT [26] Size of original data 255 MB 1.24 GB 241 MB 7.54 KB Representation Both ADL and fall Both ADL and fall Daily Human Daily Human activities activities No. of sensors 1 3 2 2 Sensors used Accelerometer Accelerometer, Accelerometer and Accelerometer and Gyroscope and Gyroscope Gyroscope Orientation sensor Device used to Samsung Galaxy Nexus Samsung Galaxy S3 Samsung Galaxy S II Samsung Galaxy S collect data I9250 with Bosh with LSM330DLC II with HARApp BMA220 module Position of the Trouser front pocket (half Not Specified Waist-mounted Different body parts device of the time in the left one (trunk, upper and and the remaining time lower extremities) in the right one) No of Subjects 30 (24 Females, 6 males) 57 (15 Females, 30 (the number of 17 (the number of 42 males) female and male is) female and male is) not specified) not specified) Physical Age: 18-60, Age: 20-47, Age: 19-48 Not specified information about Height: 160-190, Height: 160-189, the subjects Weight: 50-82 Weight: 50-120 No. of activities 9 types of ADL and 9 types of ADL and 6 type of daily 12 type of daily 8 type of fall 4 type of fall activities activities Nature of data Raw accelerometer Raw accelerometer, Extracted feature Feature from data gyroscope and accelerometer and accelerometer and orientation data gyroscope data gyroscope data Sampling frequency 50 Hz 20 Hz 50 Hz 50 Hz in original dataset Samples in modified 4,632 67,348 10,299 10,411 dataset Depicted activities walk, walk up-stair, walk, walk down-stair, walk, walk up-stair, walk, walk up-stair, walk down-stair, sitting, sitting, standing, walk down-stair, sitting, walk down-stair, sitting, standing, laying down jogging standing, laying down standing,laying down Signal window 3 second 1 second 1.5 second 1.5 second These results likewise indicate how adequately EdgeFall architecture can analyze different human activities. We observe from the assessment that the model developed by ActDec-SysOpt outperforms classical- based recognition models. So, our significant contributions are: − Designing and resembling the end-edge-cloud-based EdgeFall architecture to figure out the usefulness in human activity detection. EdgeFall: a promising cloud-edge-end architecture for elderly fall care (Kazi Md Shahiduzzaman)
  • 4. 4724 ❒ ISSN: 2088-8708 − Spot the challenges like FAR, accuracy, precision, and recall in loosely coupled data fusion for recognizing human activities. − Performance improvement in classifying and realizing human daily activity through tightly coupled multiple information sources and ActDec-SysOpt algorithm. The remaining paper dwells on four more sections. The next section will present the proposed Edge- Fall paradigm for activity recognition. Algorithms used for this application are depicted in section 4 with respective flowcharts. The EdgeFall is simulated using the ActDec-SysOpt algorithm for human activity detec- tion and described there. Then, the empirical setups for evaluating EdgeFall architecture with data fusion are analyzed in section 5. The performance evaluating metrics are also enlisted in this section. Research findings and highlights on prospective research directions are mentioned in the conclusion. 2. EdgeFall: THE PROPOSED ARCHITECTURE The proposed EdgeFall system is a three-tiered system as depicted in Figure 1 which are the artificial intelligence (AI)-aware medical cloud for model development, the mobile edge for human activity recognition, user identification, fall detection, prediction, and the body area network (BAN) for smart sensing and command execution. The proposed architecture makes the fall detection and prediction system into three distributed sub- systems connected to each other by the wireless network. We expect this architecture to be efficient regarding computational capacity and energy management. BAN is indeed a platform consists sensors like wearable cameras, gyroscopes, accelerometers, GPS sensors, and other related sensors and is pictured as a wearable glass, but we can accept any platform like a watch, wearable around the waist, helmet on the head, smart- phone on the pocket or similar. This layer performs data collection, feature extraction, communication to the edge node, and decision execution. The mobile edge layer consists of a rational edge node and contrasting processing units. User identification, activity recognition, lead time calculation for fall prediction, and fall detection are major responsibilities of this tier. It will adapt protocols, classification networks, and users’ daily human activity patterns from the cloud. The last tier is the cloud, the intellect of this technique. Deep learning, special machine learning approach, and human intervention (HI) are used to implement this tier. It develops a user identification model and activity classification model. It should update the mobile edge node with updated reformed system parameters in a frequent cycle. This architecture can again refer to diverse areas related to health care. Figure 1. Proposed architecture of EdgeFall architecture In this script, we will present the essence of this recommended architecture by breaking up the com- putational capacity among mobile edge nodes and the cloud. Here, we have simply focused on various human activity recognition. Because it will contribute to strong fall detection and prediction. Int J Elec & Comp Eng, Vol. 13, No. 4, August 2023: 4721-4733
  • 5. Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 4725 3. PROPOPSED ALGORITHM This section will discuss the methodologies named FuseFall and ActDec-SysOpt algorithm. These two approaches will demonstrate the validity of data fusion, cloud-edge-end type architecture, and the classification algorithm based on recurrent neural networks. This section will help the readers to visualize the approaches used in this paper. 3.1. FuseFall algorithm Fall detection takes advantage of two sources (a wearable camera and a tri-axial accelerometer) of information. Two different sources detect human activity and fall separately. The block diagram in Figure 2(a) explains the FuseFall algorithm. In this algorithm, the information sources are loosely coupled and analyzed separately. We can observe that both sources use a different suitable algorithm to define the same activity separately and compare it to the final decision from the block diagram, as shown in Figure 2(a). This algorithm is explained in [31]. Two separate processes are used for two data types to reduce the complexity. Both processes run simultaneously but on two different platforms. We assume that the resource required to analyze inertia sensors data is lesser than image processing. So, the inertia sensor’s data processing process will run in the edge node, while image processing will occur in the cloud. An optimization algorithm will be necessary to enhance the efficiency of the system. In the detection phase, inertia data will be analyzed to detect a fall using SVM machine learning technique. The reason behind using SVM technique is to perform binary classification between AD and fall. A signal window of 3 seconds is considered to determine an event. In the case of detection, image processing techniques are used to determine the feature (i,e, dissimilarity distance between odd frames) from the captured images between 1.5 seconds backward and 1.5 seconds forward from the instant of fall occurrence. This dissimilarity distance will verify the fall event determined in the detection phase. The overall fall detection algorithm is shown in algorithm 1. Finally, another process will be used to optimize the system presented in algorithm 2. This process automatically selects the optimal trained model and system parameters to overwrite the old network. 3.2. ActDec-SysOpt algorithm Figure 2(b) displays the block diagram of ActDec-SysOpt. We train the classification model with long short term memory network (LSTM). This algorithm includes two parts. We adopt ActDec to recognize the human activity, whereas SysOpt is for training classification networks. These two parts also depict mobile edge nodes and medical cloud appropriately. We explain the details in [34]. (a) (b) Figure 2. Block diagram of proposed algorithms (a) block diagram of FuseFall algorithm for fall detection and (b) simple block diagram of ActDec-SysOpt algorithm EdgeFall: a promising cloud-edge-end architecture for elderly fall care (Kazi Md Shahiduzzaman)
  • 6. 4726 ❒ ISSN: 2088-8708 Now the information sources are tightly coupled. This algorithm represents a centralized trained model of human activity recognition for all users. SysOpt generates the trained classification model (ActDec) on the cloud server and pushes it to the edge server. A multivariate time series signal is fed to the bi-directional LSTM. A signal window of 1.5 seconds is used here. A softmax layer matches the best probability fit for a certain human activity. With a fully connected network, each human activity model is trained in the cloud layer. The point should be mentioned that the SysOpt algorithm optimizes the parameters like learning rate, epoch, and number of layers. According to the incoming data and the patient’s nature of performing the activities. This trained network is then passed to the edge layer, where the incoming pre-processed raw data from the edge layer will be analyzed to determine the particular activities and types of falls. The overall activity detection algorithm is shown in algorithm 3. Finally, another algorithm will be used to initialize and optimize the system presented in algorithm 4. The test data will be labelled and added to the training data to train the even detection network. If the result of the newly trained system is better than the old one, the parameter e,g, threshold value used in the verification phase will be changed. Otherwise, the old parameters are kept. The fall detection and verification algorithm runs in an intelligent access point, whereas the optimization algorithm and process data will be stored in the cloud. Algorithm 1: Detection process of FuseFall algorithm Detection Phase Data: Inertia Sensor’s Data Result: Detection of an activity Load SVM Model initialization while Unless detecting a Fall do Extract feature F(accl) from signal window %3 sec data if F(accl) ∼ ADL then go to initialization else if F(accl) ∼ Fall then Verification Phase end end end Verification Phase Data: Image data Result: Verification of Fall while Fall Detected do Load images between t-1.5 to t+1.5 %Images correspond to 3 sec Extract feature I(dissimilarity) if I(dissimilarity) ∼ Fall then Fall verified else Consider as ADL end end Algorithm 2: FuseFall optimization process Data: Test Data Result: Optimization of the System begin Test data −→ Training Data Evaluate Both trained model if ResultNew > ResultOld then Pass Paranew to Detection process Modify Both Phase else Keep previous State Remove Test data end end Int J Elec & Comp Eng, Vol. 13, No. 4, August 2023: 4721-4733
  • 7. Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 4727 Algorithm 3: ActDect algorithm Data: Inertia sensor’s data Result: Event Detection Load Trained LSTM Classification Model initialization while Unless detecting an event do Extract feature F(accl & gyro) from signal window %1.5 sec data if F(accl & gyro) ∼ Daily Activity then go to initialization else if F(accl & gyro) ∼ Fall then Alarm, Protection end end end Algorithm 4: SysOpt algorithm Data: Test Data Result: Optimization of detection system begin Test data −→ Training Data Run LSTM with TrainingDataNew if V alidationNew > V alidationOld then Modify Classification Network Pass Paranew to Event Detection Algorithm else Keep previous State Remove Test data end end 4. EVALUATION This section presents our experiments and results using FuseFall and ActDec-SysFall architecture. The findings will justify our claims outlined in 1. First, the performance evaluating metrics will be deliberated, and results and findings from studied algorithms will be later. 4.1. Metrics Accuracy, recall, precision, F1-score, FAR, training time, and execution time are the performance evaluation metrics. The definitions are given below: − Accuracy is stated precisely the percentage of correctly detected activities from the tested one. − Recall is interpreted as the percentage of proper activities that are exactly verified. It provides information about the accuracy row-wise of the confusion matrix. − Precision is elucidating as the segment of equitable classified activities i,e, accuracy along column wise of the confusion matrix. − F1 score is the harmonic mean of precision and recall, which measures the accuracy of the test. − False alarm rate (FAR) acquaints with how the system gets confused between any activity and other con- sidered activities. − Training-time is the time required by the cloud to generate a global classification network with the proposed algorithm. − Execution time is the time the global classifier recognizes human activities at the mobile edge node. 4.2. Experiments with FuseFall We have simulated and tested loosely coupled two different data in this experiment. The results are generated by the wearable camera y1 and the raw accelerometer data y2 separately at this stage, as shown in Figure 3(a). This experiment demonstrates the real-life daily human activities and falls related to elderly people. EdgeFall: a promising cloud-edge-end architecture for elderly fall care (Kazi Md Shahiduzzaman)
  • 8. 4728 ❒ ISSN: 2088-8708 4.2.1. Experimental setup and results Figure 3(a) presents the simulation scheme. A forward fall will take place from the walking position. We present the details about how to evaluate it in [31]. We calculate the accuracy, precision, recall, F1-score, and execution time of y2 from 4.1. We consider the k-NN method for comparison. The results are presented in Table 2. We have utilized the frontal camera of Xiaomi S2 around the head height to perform the same activities by taking video, similarly to [34]. As illustrated in Figure 3(b), the dissimilarity distance for the video frames during the fall is considerably larger than other frames, marked by two red circles. From the binary classification between falls and daily human activity, we can determine the effectiveness of FuseFall. Nevertheless, the challenge persists in performing this FuseFall with multi-class classification to recognize the daily human activity. Please point out to [34] what we mean by multi-class activity. Table 3 gives the performance of the FuseFall algorithm for accelerometer data with k-NN. During video processing, we have observed high picks like Figure 3(b), which can cause fall alarms. Table 3 and Figure 3(b) show that a loosely coupled information source with the FuseFall algorithm is unsuitable for multi-class activity detection. That is why we hold our experiments with y1 and y2. Besides, the time involved in generating results with video processing is large in our experiments. The challenges ascribed to the multi-class activity classification are managed more accurately by the ActDec-SysOpt algorithm [34], which will be demonstrated in the next part. (a) (b) Figure 3. Experimental outcome of FuseFall (a) experimental setup to evaluate FuseFall algorithm and (b) FAR during multiclass activity detection with FuseFall (video processing part) Table 2. Partial evaluation performance of FuseFall to differentiate fall from daily activities Metrics FuseFall for y2 k-NN Accuracy 0.9667 0.900 Recall 0.9667 0.900 Precision 0.9839 0.9545 F1-score 0.9752 0.9265 FAR 0.033 0.10 Execution time (sec) 2.282 1.043 Int J Elec & Comp Eng, Vol. 13, No. 4, August 2023: 4721-4733
  • 9. Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 4729 Table 3. Performance of FuseFall for multi-class type human activity recognition Metrics FuseFall for y2 k-NN Accuracy 0.8468 0.8022 Recall 0.8198 0.8222 Precision 0.8518 0.8595 F1-score 0.8355 0.8405 False alarm rate (FAR) 0.0396 0.0547 4.3. Experiments with ActDec-SysOpt As we have previously noticed, that single source of information and classical machine learning-based models are not up to the adequate standard for realizing different human activities. Here we will adopt ActDec- SysOpt [34] algorithm and data fusion technique to justify some remarkable objectives, i) demand of proposed EdgeFall as shown in Figure 1, ii) the performance of multiple sources, use of features and LSTM, and iii) the prerequisite of tight coupling. 4.3.1. Experimental setup For the experimental setup we have utilized four original data sources, named, UniMIB SHAR [23], MobiAct [24], UCI HAR [25] and HAPT [26] to simulate BAN of the EdgeFall. These sources consist of diverse human activities for several volunteers. We have transformed the original datasets in such a manner that it mirrors the BAN of the proposed architecture which relays data to the edge node regularly in time windows. The preliminary results will confirm the potency of tightly coupled data fusion. ActDec represents the mobile edge node while SysOpt is for the medical cloud. Here, we want to assess how dramatically this algorithm can recognize different human activities i,e, the classification process. 4.3.2. Results First, we will present the operation of EdgeFall through ActDec-SysOpt algorithm. The supporting chart (Table 4 accumulates the simulation result which outperforms classical machine learning type algorithm. We have considered a traditional data split (70-30) between training and test data. Results indicate that ActDec- SysOpt outperforms the other two machines learning-based classification models (where one of them is Fuse- Fall). Accordingly, we will weigh the performance of ActDec-SysOpt on different datasets. These different datasets represent BAN of the recommended EdgeFall architecture (i,e, several users with unfamiliar weights, heights, ages, and gender). Table 4. Performance of ActDec-SysOpt over FuseFall and k-NN algorithm Methods FuseFALL k-NN (k=2) ActDec-SysOpt Accuracy 0.8468 0.8022 0.9187 Recall 0.8198 0.8222 0.8303 Precision 0.8518 0.8595 0.8511 F1-score 0.8355 0.8405 0.8406 FAR 0.0396 0.0547 0.025 Training-time (minute) 2.282 1.043 112 To conduct the simulation further sensibly, we have considered a 50-50 approach for training and test data to appraise the validity of data fusion. Figure 4 displays the simulation results and reveals that data fu- sion from numerous sources with proposed ActDec-SysOpt have stronger achievement than information from a single source and loosely coupled multiple information sources. We can observe that feature extraction can develop a far better result than raw data fusion. The accuracy of raw data fusion is remarkably unsatisfac- tory around 59.65% with a considerable FAR 10.66%. Whereas the feature extraction (tightly coupled) from multiple information sources forms a sharp accuracy of 92.01% with FAR of 1.69%. We present the expected training time and classification execution time in Table 5. The time of training classification network and execution time to realize an activity is higher for tightly coupled multiple information sources. As we have divided the classification model development and execution process into cloud and edge, the execution time may not be a perturbing factor. This time can be boosted by adopting a hardware/software co-design concept. EdgeFall: a promising cloud-edge-end architecture for elderly fall care (Kazi Md Shahiduzzaman)
  • 10. 4730 ❒ ISSN: 2088-8708 Figure 4. Advantages of data fusion from multiple information sources From Tables 4 and 5, Figure 4 we come up with the following decisions: − Tightly coupled information sources with an LSTM-based classification model are more adequate. The accuracy is around 92% with FAR being 1.6%. The high value of recall, precision, and F1-score confirm that the classification model is competent in recognizing human activity precisely. − The execution time compelled to recognize activity from one window of information is noticeably less, around 15 ms with data fusion from multiple sources (check execution time of Table 4. − RNN type machine learning-based algorithms are better applicable to recognizing human daily activities. − Features are better reasonable as they can decrease the FAR. 1.69% with HAPT and 1.87% with UCI HAR. − However, we can learn that training the classification model requires a lot of time, 1,094 minutes for HAPT and 691 minutes for UCI HAR. Medical cloud is good enough to deal with the training process because there are no issues with the computational capability and cache. − The train network later can be passed to the edge node which can execute activity recognition. − Data collection and feature extraction can be managed conveniently in BAN. Table 5. Effectiveness of EdgeFall architecture in computation Methods UniMIB SHAR Mobi Act HAPT UCI HAR Training-time (minute) 84 161 1094 691 Execution time (ms) ∼ 4.52 - 7.065 ∼ 0.6 - 0.73 ∼ 14.16 - 15.74 ∼ 15.20 - 16.71 So far, ActDec-SysOpt is a definite stride toward fall prediction and detection using EdgeFall. Nev- ertheless, there are some challenges that need to figure out. For example, accuracy, FAR, a simple algorithm for fall prediction and detection, and user identification. Though, we are a little behind in accomplishing the coveted accuracy of over 95%. 5. CONCLUSION In this study, we have recommended a cloud-edge-end architecture, EdgeFall, which can distribute the full function of a single board structure to three tiers. A particular contribution of this script is that we can uphold our claims through proper simulation results. We have sought to figure out challenges related to establishing this system through appropriate experimental results. The ActDec-SysOpt algorithm represents not only the EdgeFall architecture but also validates the potency of such architecture. The ActDec-SysOpt Int J Elec & Comp Eng, Vol. 13, No. 4, August 2023: 4721-4733
  • 11. Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 4731 algorithm also opens up the opportunity for exploiting data fusion for this system for further enhancements in a term of accuracy and FAR for human activity recognition. We are able to obtain 92% accuracy while lowering the FAR to 1.69%. Therefore, this EdgeFall architecture can be regarded as being beneficial for reducing computational capacity and power utilization on BAN, reinforcing the classification performance by specific recognition of an activity. This architecture can also be useful in the management of mobile edge nodes, user- profile management (patient-specific), and quality of service (QoS) which are our future research directions. A suitable single algorithm requires to promote which can predict and detect falls. Visual inertial odometry- based algorithms may have the capacity because they can produce trajectory. User identification is another issue that opens a possible research trend. We can develop a modern machine learning-based algorithm to solve it with privacy protection measures. Another prospective future task for the smart anti-fall system is to build up a secure, reliable, low latency-based network. Software-defined network (SDN) can be the most practical solution for not only meeting the QoS requirements but also providing security to the network. Speeding up the execution time is another demand in implementing this EdgeFall architecture into a prototype device. Hardware-software co-design approach can overcome this challenge. This paper not only reviews but also evaluates different approaches toward developing smart anti-fall systems. Concepts, algorithms, experimental setups, results and decisions, provide comprehensive guidance for deeper investigation, particularly in this emerging area. ACKNOWLEDGEMENT This project is partially funded by Jatiya Kabi Kazi Nazrul Islam University’s Annual (2021-22) Re- search Grunt. REFERENCES [1] Z. Wang, V. Ramamoorthy, U. Gal, and A. Guez, “Possible life saver: a review on human fall detection technology,” Robotics, vol. 9, no. 3, Jul. 2020, doi: 10.3390/robotics9030055. [2] E. C. Dinarevic, J. B. Husic, and S. Barakovic, “Issues of human activity recognition in healthcare,” in 2019 18th International Symposium INFOTEH-JAHORINA (INFOTEH), Mar. 2019, pp. 1–6, doi: 10.1109/IN- FOTEH.2019.8717749. [3] J. A. Álvarez-Garcı́a, L. M. Soria Morillo, M. Á. Á. de La Concepción, A. Fernández-Montes, and J. A. O. Ramı́rez, “Evaluating wearable activity recognition and fall detection systems,” in IFMBE Proceedings of the European Conference of the International Federation for Medical and Biological Engineering, vol. 45, 2015, pp. 653–656. [4] K. Chaccour, R. Darazi, A. H. El Hassani, and E. Andres, “From fall detection to fall prevention: a generic classification of fall-related systems,” IEEE Sensors Journal, vol. 17, no. 3, pp. 812–822, Feb. 2017, doi: 10.1109/JSEN.2016.2628099. [5] WHO, “Fall,” World Health Organization, 2018. https://www.who.int/news-room/fact-sheets/detail/falls (accessed Aug. 20, 2018). [6] S. Liang, Y. Liu, G. Li, and G. Zhao, “Elderly fall risk prediction with plantar center of force using ConvL- STM algorithm,” in 2019 IEEE International Conference on Cyborg and Bionic Systems (CBS), 2019, pp. 36–41, doi: 10.1109/CBS46900.2019.9114487. [7] T. Xu, Y. Zhou, and J. Zhu, “New advances and challenges of fall detection systems: a survey,” Applied Sciences, vol. 8, no. 3, Mar. 2018, doi: 10.3390/app8030418. [8] R. Rajagopalan, I. Litvan, and T.-P. Jung, “Fall prediction and prevention systems: recent trends, challenges, and future research directions,” Sensors, vol. 17, no. 11, Nov. 2017, doi: 10.3390/s17112509. [9] T.R. Frieden, D. Houry, G. Baldwin, A. Dellinger, R. Lee, “Preventing falls: a guide to implementing effective community-based fall prevention programs,” National Center for Injury Prevention and Control of the Centers Disease Control and Prevention, 2015. [10] J. Hamm, A. G. Money, A. Atwal, and I. Paraskevopoulos, “Fall prevention intervention technologies: A conceptual framework and survey of the state of the art,” Journal of Biomedical Informatics, vol. 59, pp. 319–345, Feb. 2016, doi: 10.1016/j.jbi.2015.12.013. [11] “Global Ageing,” ageinternational, 2018. https://www.ageinternational.org.uk/policy-research/statistics/global- ageing/ (accessed Nov. 22, 2018). [12] L. Montanini, A. Del Campo, D. Perla, S. Spinsante, and E. Gambi, “A footwear-based methodology for fall detec- tion,” IEEE Sensors Journal, vol. 18, no. 3, pp. 1233–1242, Feb. 2018, doi: 10.1109/JSEN.2017.2778742. [13] A. Clark, “The best medical alert systems 2019,” TheSeniorList, Sep. 2019. https://www.theseniorlist.com/medical- alert-systems/best/ (accessed Sep. 21, 2018). EdgeFall: a promising cloud-edge-end architecture for elderly fall care (Kazi Md Shahiduzzaman)
  • 12. 4732 ❒ ISSN: 2088-8708 [14] W. Saadeh, S. A. Butt, and M. A. Bin Altaf, “A patient-specific single sensor IoT-based wearable fall predic- tion and detection system,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 27, no. 5, pp. 995–1003, May 2019, doi: 10.1109/TNSRE.2019.2911602. [15] T. R. Mauldin, M. E. Canby, V. Metsis, A. H. H. Ngu, and C. C. Rivera, “SmartFall: a smartwatch-based fall detection system using deep learning,” Sensors, vol. 18, no. 3363, Oct. 2018, doi: 10.3390/s18103363. [16] M. J. Al Nahian et al., “Towards an accelerometer-based elderly fall detection system using cross-disciplinary time series features,” IEEE Access, vol. 9, pp. 39413–39431, 2021, doi: 10.1109/ACCESS.2021.3056441. [17] L. Rachakonda, S. P. Mohanty, and E. Kougianos, “Good-eye: a device for automatic prediction and detection of elderly falls in smart homes,” in 2020 IEEE International Symposium on Smart Electronic Systems (iSES) (Formerly iNiS), Dec. 2020, pp. 202–203, doi: 10.1109/iSES50453.2020.00051. [18] Y. Zhang, M. Qiu, C.-W. Tsai, M. M. Hassan, and A. Alamri, “Health-CPS: healthcare cyber-physical system assisted by cloud and big data,” IEEE Systems Journal, vol. 11, no. 1, pp. 88–95, Mar. 2017, doi: 10.1109/JSYST.2015.2460747. [19] A. Paneerselvam, R. Yaakob, T. Perumal, and E. Marlisah, “Fall detection framework for smart home,” in 2018 IEEE 7th Global Conference on Consumer Electronics (GCCE), Oct. 2018, pp. 351–352, doi: 10.1109/GCCE.2018.8574617. [20] N. Fares, R. S. Sherratt, and I. H. Elhajj, “Directing and orienting ICT healthcare solutions to address the needs of the aging population,” Healthcare, vol. 9, no. 2, Feb. 2021, doi: 10.3390/healthcare9020147. [21] S. Chaudhuri, H. Thompson, and G. Demiris, “Fall detection devices and their use with older adults: a systematic review,” Journal of Geriatric Physical Therapy, vol. 37, no. 4, pp. 178–196, Oct. 2014, doi: 10.1519/JPT.0b013e3182abe77. [22] K. Ozcan, S. Velipasalar, and P. K. Varshney, “Autonomous fall detection with wearable cameras by using relative entropy distance measure,” IEEE Transactions on Human-Machine Systems, vol. 47, no. 1, pp. 31–39, Feb. 2017, doi: 10.1109/THMS.2016.2620904. [23] D. Micucci, M. Mobilio, and P. Napoletano, “UniMiB SHAR: a dataset for human activity recognition using acceler- ation data from smartphones,” Applied Sciences, vol. 7, no. 10, Oct. 2017, doi: 10.3390/app7101101. [24] V. George, C. Chariklei, M. Thodoris, P. Matthew, and T. Manolis, “The MobiAct dataset: recognition of activities of daily living using smartphones,” in International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AWE), Apr. 2016, pp. 143–151. [25] A. Davide, G. Alessandro, O. Luca, P. Xavier, and L. R.-O. Jorge, “A public domain dataset for human activity recognition using smartphones,” in European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN), Apr. 2013. [26] J.-L. Reyes-Ortiz, L. Oneto, A. Samà, X. Parra, and D. Anguita, “Transition-aware human activity recognition using smartphones,” Neurocomputing, vol. 171, pp. 754–767, Jan. 2016, doi: 10.1016/j.neucom.2015.07.085. [27] M. Li, G. Xu, B. He, X. Ma, and J. Xie, “Pre-impact fall detection based on a modified zero moment point criterion using data from kinect sensors,” IEEE Sensors Journal, vol. 18, no. 13, pp. 5522–5531, Jul. 2018, doi: 10.1109/JSEN.2018.2833451. [28] J. Howcroft, J. Kofman, and E. D. Lemaire, “Prospective fall-risk prediction models for older adults based on wear- able sensors,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 25, no. 10, pp. 1812–1820, Oct. 2017, doi: 10.1109/TNSRE.2017.2687100. [29] W. Engel and W. Ding, “Reliable and practical fall prediction using artificial neural network,” in 2017 13th Inter- national Conference on Natural Computation, Fuzzy Systems and Knowledge Discovery (ICNC-FSKD), Jul. 2017, pp. 1867–1871, doi: 10.1109/FSKD.2017.8393052. [30] Y. Wang, K. Wu, and L. M. Ni, “WiFall: device-free fall detection by wireless networks,” IEEE Transactions on Mobile Computing, vol. 16, no. 2, pp. 581–594, Feb. 2017, doi: 10.1109/TMC.2016.2557792. [31] K. M. Shahiduzzaman, X. Hei, C. Guo, and W. Cheng, “Enhancing fall detection for elderly with smart helmet in a cloud-network-edge architecture,” in IEEE International Conference on Consumer Electronics- Taiwan (ICCE-TW), May 2019. [32] N. Otanasap, “Pre-impact fall detection based on wearable device using dynamic threshold model,” in Parallel and Distributed Computing, Applications and Technologies, Dec. 2016, vol. 0, pp. 362–365, doi: 10.1109/PD- CAT.2016.083. [33] L. Tong, Q. Song, Y. Ge, and Ming Liu, “HMM-based human fall detection and prediction method using tri-axial accelerometer,” IEEE Sensors Journal, vol. 13, no. 5, pp. 1849–1856, May 2013, doi: 10.1109/JSEN.2013.2245231. [34] K. M. Shahiduzzaman, J. Peng, Y. Gao, X. Hei, and W. Cheng, “Towards accurate and robust fall detection for the elderly in a hybrid cloud-edge architecture,” in IEEE Smart World Congress (SmartWorld), Aug. 2019. Int J Elec & Comp Eng, Vol. 13, No. 4, August 2023: 4721-4733
  • 13. Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 4733 BIOGRAPHIES OF AUTHORS Kazi Md Shahiduzzaman received the B.Sc.Eng. degree in Electrical and Electronic Engineering (EEE) from the Khulna University of Engineering and Technology (KUET), Bangladesh, in 2010 and the M.Sc. from Hochschule Bremen (University of Applied Sciences Bremen), Germany, in 2013. Now, He is pursuing a Ph.D. degree in EEE from KUET. He is an Associate Professor at the Department of EEE, Jatiya Kabi Kazi Nazrul Islam University, Bangladesh. His research interests include smart health care, elderly daily activity recognition, machine learning techniques, free space optical communication, cloud-edge-end-based distributed platforms, artificial intelligent load, and power production forecasting. His research mainly focuses on application-oriented issues. He can be contacted at email: shahiduzzaman2103751@stud.kuet.ac.bd and shahiduzzaman@jkkniu.edu.bd. Md. Salah Uddin Yusuf received the B.Sc., M.Sc., and Ph.D. degree in Electri- cal and Electronic Engineering from Khulna University of Engineering and Technology (KUET), Bangladesh, in 2001, 2005, and 2016 respectively. Since 2001, he has been with the same depart- ment of the same University and currently serving as Professor. His research interests mainly focus on deep learning-based signal, image, and video processing with quality assessment, Face liveness detection and authentication, and GAN-based Human activity analysis for the modern healthcare systems. He can be contacted at email: suyusuf@eee.kuet.ac.bd. EdgeFall: a promising cloud-edge-end architecture for elderly fall care (Kazi Md Shahiduzzaman)