Chameleon Swarm Algorithm with Improved Fuzzy Deep Learning for Fall Detection Approach to Aid Elderly People

Abstract

Over the last few decades, the processes of mobile communications and the Internet of Things (IoT) have been established to collect human and environmental data for a variety of smart applications and services. Remote monitoring of disabled and elderly persons living in smart homes was most difficult because of possible accidents which can take place due to day-to-day work like falls. Fall signifies a major health problem for elderly people. When the condition is not alerted in time, then this causes death or impairment in the elderly which decreases the quality of life. For elderly persons, falls can be assumed to be the main cause for the demise of posttraumatic complications. Therefore, early detection of elderly persons’ falls in smart homes is required for increasing their survival chances or offering vital support. Therefore, the study presents a Chameleon Swarm Algorithm with Improved Fuzzy Deep Learning for Fall Detection (CSA-IDFLFD) technique. The CSA-IDFLFD technique helps elderly persons with the identification of fall actions and improves their quality of life. The CSA-IDFLFD technique involves two phases of operations. In the initial phase, the CSA-IDFLFD technique involves the design of the IDFL model for the identification and classification of fall events. Next, in the second phase, the parameters related to the IDFL method can be optimally selected by the design of CSA. To validate the performance of the CSA-IDFLFD technique in the fall detection (FD) process, a widespread experimental evaluation process takes place. The extensive outcome stated the improved detection results of the CSA-IDFLFD technique.

Main article text

INTRODUCTION

Regular natural processes occasionally place older people at a higher risk of falling. Falls are a common and frequently ignored reason for injuries in elderly people ( Rezaee et al., 2022). Some reasons for the fall in elderly people are muscle weakness, heart disease, hypotension, poor eyesight, etc. Falling is a common cause of medical attention essential for aged people ( Hassan et al., 2019). Elderly people always get injured by falling particularly if they are living alone. After a fall happens, clinical care should be presented quickly for decreasing the risk to victim. Numerous technologies were devised that use web cameras to observe elderly people’s activities ( Kumar et al., 2021). Nevertheless, the cost of installation and operation is not economical and it can be installed indoors ( Karar et al., 2022). Present commercialized devices need users to wear wireless emergency transmitters in the form of wristwatches. This technique would limit the user movements and produces higher false alarm because of frequent movement and swinging of the devices ( Ramachandran et al., 2020). The long lie can give rise to severe health problems that include hypothermia, dehydration, and pneumonia and many cases can result in death within 6 months after a fall. Hence, a fall not aided in time in an elderly individual can adversely affect their QoL and independence ( Yu et al., 2021).

In this context, the growth of real-time Internet of Things (IoT) systems that contribute to proficiently identifying falls and alerting emergency services in the short-lived possible time is a social need ( Tahir et al., 2022). Fall detection (FD) policies use alert systems for providing and identifying emergency support to seniors who are disposed to falls. If one falls, such systems will rapidly trigger the sensors. The built-in technology may be positioned around the wrist, around the neck, or on the waist, dependent on the devices ( Kyriakopoulos et al., 2020). Many clinical alert businesses include the fall recognition ability in their clinical alert structure for premium service charges. Few companies sell FD devices that could be worn distinctly from one’s clinical alert button. During the monthly subscription plan, the cost of the second device can be included ( Pillai et al., 2022). Many presented systems utilize statistical methods that regularly make false alarms during classification and detection. Furthermore, statistical methods are less effective in the occurrence of nonlinear and complex issues ( Gharti, 2020). Generally, gait analysis for FD and avoidance makes noisy data in the acquisition. Statistical approaches are sensitive to noisy information that resulted in degradation performance.

This article introduces a Chameleon Swarm Algorithm with Improved Fuzzy Deep Learning for Fall Detection (CSA-IDFLFD) technique. The CSA-IDFLFD technique helps elderly persons with the identification of fall actions and improves their quality of life. The CSA-IDFLFD technique involves two phases of operations. In the initial phase, the CSA-IDFLFD technique involves the design of the IDFL model for the identification and classification of fall events. Next, in the second phase, the parameters related to the IDFL method can be optimally selected by the design of CSA. To validate the performance of the CSA-IDFLFD technique in the FD process, a widespread experimental evaluation process takes place.

RELATED WORKS

Kong et al. (2019) developed a histogram of oriented gradients (HOG)-support vector machines (SVM)-related FD-IoT mechanism for elderly people. To ensure privacy, deep sensors can be utilized by RGB cameras to acquire the binary imageries of senior people. People are tracked and detected by Microsoft Kinect software development kit (SDK), and the undesirable noises are reduced by noise reduction method. After gaining the denoized binary imageries, histogram of oriented gradient mines the person’s features and the image classification would be achieved to determine the fall status by the liner SVM. Yadav et al. (2022) presented an ARFDNet, activity recognition, and FD system. Now, to abstract skeleton attributes of the user, the new RGB video was sent to the pose prediction network. Such skeleton coordinate is inputted and preprocessed in sliding window way to modeled GRUs and convolutional neural networks (CNNs), to study the spatiotemporal dynamics in the information. The output of GRU is sent to the FC layer for classification.

Khraief et al. (2020) designed weighted multi-stream deep convolutional neural network (DCNN) that uses the rich multimodal dataset presented by the RGB-D camera. This approach finds fall actions and transfers a help application to caregivers. The contribution is three-fold. The author constructs a novel structure containing four separate CNN streams, one for all modalities. Taghvaei and Kosuge (2018) inspect an image-related classification of individual movement while utilizing a walking support structure for enhancing the dependability and safety of this system. The author classifies human behavior while exploiting walker robots into eight states (five falling, sitting, walking, and standing types), and offers two approaches, namely, hidden Markov models (HMMs) and normal distribution for finding and recognizing these states. In Kchouri et al. (2022), different training datasets are allocated to the membership degree. Few data points with higher possibility of falling are allotted a higher amount of membership, earning a higher contribution for making decisions. This fails to attain precise FD but minimizes hesitation in labeling the results and enhances the SVM heuristic transparency.

In Rahman et al. (2020), a prototype of wheelchairs with an FD structure can be advanced that can be controlled by an android application. A nurse or caretaker can control this developed prototype in various directions with speed variations; the patient’s position is observed. The FD mechanism utilizes an android device’s accelerometer and gyroscope. Wang et al. (2020) devise a study on FD of walking training robot (WTR) related to fuzzy reasoning that utilizes a two-dimensional (2D) laser sensor and high-precision posture sensors. By utilizing, one wearable sensor not just evades the issue of lower user comfort, but also ensures the detection effects.

THE PROPOSED MODEL

In this study, a new FD method named CSA-IDFLFD technique has been developed for helping elderly people. The CSA-IDFLFD technique helps elderly persons with the identification of fall actions and improves their quality of life. The CSA-IDFLFD technique involves two phases of operations: FD and parameter tuning. Figure 1 defines the overall flow of the CSA-IDFLFD approach.

Figure 1:

Overall flow of the CSA-IDFLFD approach. Abbreviations: CSA-IDFLFD, Chameleon Swarm Algorithm with Improved Fuzzy Deep Learning for Fall Detection.

IDFL-based FD

In the initial phase, the CSA-IDFLFD technique involves the design of the IDFL model for the identification and classification of fall events. The architecture and mathematical modeling of the improved fuzzy deep learning (IFDL) model are discussed in this section ( Benhari and Hossseini, 2022). Using fuzzy set theory and Dempster combinational rule, the study proposes a new CNN structure with the ability of uncertainty handling a deep architecture. For evidence combination, Belief network uses Dempster–Shafer rule in this study. In order to manage uncertainty in the classification algorithm, the independent evidence is aggregated with the concept of dataset (DS). As well, fuzzy logic (FL) was combined with the Belief network to avoid the object removal of the class with lower probability and improve the efficacy of the system. The feature map in the FC layer is transmitted to distance layer for the computation of Euclidian distance between the prototypes (feature map of representative sample in data) and all the feature maps. It calculates the possibility of all the classes as m and μ vectors, where μ refers to the copy of m that is combined to vector m in the second layer. m and μ are combined using the Dempster rule. The fuzzification layer for weight adjustment was exploited to the DS layer for best prediction and to deal with cell overlapping uncertainty.

During the convolutional layer, the weight of the feature map is updated in this model. The pooling distance function layer calculates the distance between all the feature maps and prototypes. The prototype shows the dataset representation that is designated by the k-means clustering. The output node is the outcome of the distance between the nodes and all the prototypes. This value is considered event and aggregated with the Dempster rule. The FL system is used for making decisions regarding the class of the samples to resolve the problem of unclassified instances in the output.

The input tensor after augmentation is [224×224×3], and the final convolution function tensor beforehand the FC layer in GoogleNet is an inception layer of a [3×3] filter size. The number of filters is 128, the number of channels is 832, padding size is [0×0×0], and stride is [1×001]. Also, the FC layer is [1×1000] in size.

The IFDL model consists of the following layers: convolutional layers, apply Belief function, Dempster combinational rule, and fuzzification process.

Convolutional layers

This layer extracts the feature map of the input dataset. The feature is transmitted to Belief network for measuring the distance.

Apply belief function

The feature map enters into the Belief network. The Euclidian distance of prototypes and all the feature maps are evaluated, and then the probability of all the classes is calculated.

Dempster combinational rule

The probability of dissimilar classes was combined by means of Dempster rules.

Fuzzification process

A fuzzy model of the feature space was employed to the Dempster rule to avoid rejection of the Belief network.

CSA-based parameter tuning

In the second phase, the parameters based on the IDFL method can be optimally selected by the design of CSA. CSA technique focuses on modeling the dynamic behaviors of chameleon as they find food nearby dunes, marshes, and forests ( Zhou et al., 2023). The mathematical modeling stimulates the food searching process of chameleons where they rotate the eye around 360° for capturing prey and prey localization with their sticky tongue. Initialization: like other optimization algorithms, this technique is randomly initialized at the search range.

(1) $$y_t^i=\hspace{0.17em}[y_{t,1^{\prime }}^i{y}_{t,2^{\prime }{\dots }^{\prime }}^i{y}_{t,d}^i]\hspace{0.17em}$$

(2) $$y^i=l_j+\mathrm{(}{u}_j-l_j\mathrm{)}\times r,$$

where t refers to the iteration count, i = {1,2,…, n}, d shows the dimension, and

$$y_{t,d}^i$$ indicates the position of ith individuals in d dimension. In Eq. (2), r signifies the formation of random number within [0,1], u _j and l _j indicate the upper and lower boundaries of jth dimension, and y ⁱ symbolizes the initial level of ith individuals, correspondingly.

• Search for prey: the chameleon begins to find prey for solving the problem of needing food.

(3) $$y_{i,j}^{t+1}=\{\begin{array}{ll}{y}_{i,j}^t+r_2\mathrm{(}{P}_{i,j}^t-G_i^t\mathrm{)}{p}_1+r_1\mathrm{(}{G}_j^t-y_{i,j}^t\mathrm{)}{p}_2\hfill & r_j\ge P_p\hfill \\ y_{i,j}^t+\mu \mathrm{(}{l}_b^j+r_3\mathrm{(}{u}^i-l^i\mathrm{)}\mathrm{)}sgn\mathrm{(}rand-0.5\mathrm{)}\hfill & r_i<P_p\hfill \end{array}\hspace{0.17em}.$$

In Eq. (3),

$$G_j^t$$ indicates the existing optimum individual location, $$y_{i,j}^{t+1}$$ denotes the location of ith chameleons at step ( t + 1) in dimension j, $$P_{i,j}^t$$ is the fittest location of ith chameleon, the two parameters p ₁ and p ₂ are used to control the exploration capability; r ₁, r ₂, and r ₃ denotes the random value within [0,1], and r _j is a random number in the interval [0,1]. P _p = 0.1 shows the likelihood that the chameleon sees the prey sgn( rand − 0.5) is 1 or −1 that particularly affect the development and exploration:

(4) $$\mu =\gamma e^{{\mathrm{(}-\alpha t/T\mathrm{)}}^{\beta }}.$$

In Eq. (4), t shows the existing amount of iterations, α, and β are 1, 3.5, and 3, correspondingly, and T represents the maximal amount of iterations.

• Prey location: the chameleon’s eyes move individually, enabling to spatially explore and search for the prey. The eyes could rotate and view in two dissimilar directions and simultaneously focus which allows them to realize a sweeping view of the environment.

The authors performed four setups to simulate this process: localizing the matrix of rotation that identifies the prey position, changing the original position of the individual to the center of gravity, moving the individual back to its original location, and updating the location using the matrix of rotation at the center of gravity. Figure 2 demonstrates the flowchart of CSA.

Figure 2:

Flowchart of the CSA. Abbreviation: CSA, Chameleon Swarm Algorithm.

This phase simulates the location update that takes place while the chameleon finds the prey via eye rotation.

(5) $$y_{t+1}^i=y{r}_t^i+{\overline{y}}_t^i.$$

In Eq. (5),

$$y_{t+1}^i$$ represents the coordinate after rotation, $$y{r}_t^i$$ shows the coordinate of the center of rotation, and $$y_t^i$$ denotes the average position of the chameleon in every dimension.

(6) $$y{r}_t^i=m\times y{c}_{t^{\prime }}^i,$$

where

$$y{r}_t^i$$ indicates the centering coordinate.

(7) $$y{c}_t^i=y_t^i-{\overline{y}}_t^i$$

(8) $$m=R\mathrm{(}\theta ,\overrightarrow{V}{z}_1{\overrightarrow{z}}_2\mathrm{)},$$

where m represents the rotation matrix of the location,

$$\overrightarrow{V}{z}_1$$ and $$z_2{\overrightarrow{z}}_2$$ signify two quadrature variations in the spatial coordinate, and R signifies the rotation matrix on every axis.

(9) $$\theta =rsgn\mathrm{(}rand-0.5\mathrm{)}\times 180°.$$

In Eq. (9), θ signifies the rotational angle of the chameleon eye, and r denotes the random value within [0,1], such that θ constraints to [−180°, 180°].

• Hunting prey: the method defines that the chameleon closer to the prey is the better position chameleon. This chameleon with its tongue attack the prey. Thus, the location would be somewhat upgraded, meanwhile, its tongue could expand to twice its original length. The speediness of the chameleon’s tongue movement toward the target is given as follows:

(10) $$v_{t+1}^{i,j}=\omega v_t^{i,j}+r_1\mathrm{(}{G}_t^j-y_t^{i,j}\mathrm{)}{c}_1+r_2\mathrm{(}{P}_t^i-y_t^{i,j}\mathrm{)}{c}_2.$$

Most of the parameters are described previously. c1 = c2 = 1.75 control the effect of G and P on the tongue velocity:

(11) $$\omega ={\mathrm{(}1-t/T\mathrm{)}}^{\mathrm{(}\rho \sqrt{\mathrm{(}t/T\mathrm{)}}\mathrm{)}}$$

For ρ = 1 denote the parameter which controlled the exploitation capability. Once the tongue of chameleon can be projected near prey then the speed of the tongue bouncy up is a. The location of the tongue specifies the chameleon location, and the equation can be given as follows:

(12) $$v_{t+1}^{i,j}=y_t^{i,j}+\mathrm{(}{\mathrm{(}{v}_t^{i,j}\mathrm{)}}^2-{\mathrm{(}{v}_{t-1}^{i,j}\mathrm{)}}^2\mathrm{)}/\mathrm{(}2a\mathrm{)}$$

(13) $$a=2590\times \mathrm{(}1-e^{-\log \mathrm{(}t\mathrm{)}}\mathrm{)}.$$

The fitness choice is a crucial aspect of the CSA algorithm. An encoding result can be utilized for evaluating the goodness of candidate performances. At present, the accuracy value is the major condition used to plan a fitness function.

(14) $$Fitness=\text{max}\text{(}P\text{)}$$

(15) $$P=\frac{TP}{TP+FP},$$

in which, FP refers to the false positive and TP exemplifies the true positive values.

RESULTS AND DISCUSSION

In this section, the FD outcomes of the CSA-IDFLFD system are examined by a dataset comprising 314 samples with two classes as displayed in Table 1.

Table 1:

Details on database.

Class	No. of events
Fall	74
Nonfall	240
Total number of events	314

Figure 3 reveals the classifier outcomes of the CSA-IDFLFD method under test dataset. Figure 3a portrays the confusion matrix presented by the CSA-IDFLFD method on 80% of TRP. The figure specified that the CSA-IDFLFD methodology has identified 53 samples under Fall and 197 samples under Nonfall. Moreover, Figure 3b shows the confusion matrix rendered by the CSA-IDFLFD algorithm on 20% of TSP. The figure designated that the CSA-IDFLFD method has identified 20 samples under Fall and 42 samples under Nonfall. In the same way, Figure 3c validates the PR study of the CSA-IDFLFD approach. The results exhibited that the CSA-IDFLFD method has gained maximum PR performance under two classes. Eventually, Figure 3d exhibits the ROC investigation of the CSA-IDFLFD approach. The figure portrays that the CSA-IDFLFD method has productive outcomes with maximum ROC values under two class labels.

Figure 3:

Classifier outcome of (a and b) Confusion matrices, (c) PR-curve, and (d) ROC-curve.

In Table 2 and Figure 4, the fall recognition rate outcomes of the CSA-IDFLFD system is assessed under 80:20 of TRP/TSP. The results confirmed that the CSA-IDFLFD technique detected the Fall and Nonfall events. For instance, on 80% of TRP, the CSA-IDFLFD technique gains average accu _y , prec _n , sens _y , spec _y , and F _score of 99.07, 99.75, 99.07, 99.07, and 99.41%, respectively. Next, on 20% of TSP, the CSA-IDFLFD method gains average accu _y , prec _n , sens _y , spec _y , and F _score of 98.84, 97.62, 98.84, 98.84, and 98.19%, correspondingly.

Table 2:

Fall recognition rate outcome of the CSA-IDFLFD approach on 80:20 of TRP/TSP.

Class	Accu y	Prec n	Sens y	Spec y	F score
Training phase (80%)
Fall	98.15	100.00	98.15	100.00	99.07
Nonfall	100.00	99.49	100.00	98.15	99.75
Average	99.07	99.75	99.07	99.07	99.41
Testing phase (20%)
Fall	100.00	95.24	100.00	97.67	97.56
Nonfall	97.67	100.00	97.67	100.00	98.82
Average	98.84	97.62	98.84	98.84	98.19

Abbreviations: CSA-IDFLFD, Chameleon Swarm Algorithm with Improved Fuzzy Deep Learning for Fall Detection; TRT, training time; TST, testing time.

Figure 4:

Average outcome of the CSA-IDFLFD approach on 80:20 of TRP/TSP. Abbreviations: CSA-IDFLFD, Chameleon Swarm Algorithm with Improved Fuzzy Deep Learning for Fall Detection; TRT, training time; TST, testing time.

Figure 5 inspects the accuracy of the CSA-IDFLFD method in the training and validation on test database. The result specified that the CSA-IDFLFD technique has greater accuracy values over higher epochs. Besides, the higher validation accuracy over training accuracy portrayed that the CSA-IDFLFD method learns productively on test database.

Figure 5:

Accuracy curve of the CSA-IDFLFD approach. Abbreviation: CSA-IDFLFD, Chameleon Swarm Algorithm with Improved Fuzzy Deep Learning for Fall Detection.

The loss analysis of the CSA-IDFLFD technique in training and validation is shown on test database in Figure 6. The results specify that the CSA-IDFLFD method reaches adjacent values of training and validation loss. The CSA-IDFLFD technique learns efficiently on test database.

Figure 6:

Loss curve of the CSA-IDFLFD approach. Abbreviation: CSA-IDFLFD, Chameleon Swarm Algorithm with Improved Fuzzy Deep Learning for Fall Detection.

In Table 3 and Figure 7, the overall comparison study of the CSA-IDFLFD technique with recent approaches in terms of accu _y is illustrated ( Vaiyapuri et al., 2021). The results implied that the CSA-IDFLFD technique gained increased accu _y of 99.07%. On the other hand, the compared methods such as VGG16, VGG19, Depthwise Model, 1D Conv NN, 2D Conv NN, ResNet50, ResNet101, and IMEFD-ODCNN models yield decreased accu _y of 97.60, 98, 98, 92.70, 95, 95.40, 96.20, and 98.87% correspondingly.

Table 3:

Accu _y analysis of the CSA-IDFLFD approach with recent systems.

Methods	Accuracy (%)
VGG16	97.60
VGG19	98.00
Depthwise	98.00
1D Conv NN	92.70
2D Conv NN	95.00
ResNet50	95.40
ResNet101	96.20
IMEFD-ODCNN	98.87
CSA-IDFLFD	99.07

Abbreviation: CSA-IDFLFD, Chameleon Swarm Algorithm with Improved Fuzzy Deep Learning for Fall Detection.

Figure 7:

Accu _y analysis of the CSA-IDFLFD approach with recent systems. Abbreviation: CSA-IDFLFD, Chameleon Swarm Algorithm with Improved Fuzzy Deep Learning for Fall Detection.

In Table 4, the comparative training time (TRT) and testing time (TST) results of the CSA-IDFLFD method are illustrated. Based on TRT, the CSA-IDFLFD technique gains a lower cycle threshold (CT) of 0.10 s while the VGG16, VGG19, Depthwise Model, 1D Conv NN, 2D Conv NN, ResNet50, ResNet101, and IMEFD-ODCNN models obtain increased CT of 0.65, 0.77, 0.30, 0.33, 0.34, 0.39, 0.43, and 0.28 s, respectively. At the same time, based on TST, the CSA-IDFLFD method gains a lower CT of 0.12 s while the VGG16, VGG19, Depthwise Model, 1D Conv NN, 2D Conv NN, ResNet50, ResNet101, and IMEFD-ODCNN models obtain increased CT of 0.31, 0.38, 0.20, 0.23, 0.22, 0.24, 0.26, and 0.19 s correspondingly. These reuslts verified that the CSA-IDFLFD technique reaches improved FD results.

Table 4:

TRT and TST analyses of the CSA-IDFLFD approach with recent systems.

Methods	TRT (seconds)	TST (seconds)
VGG16	0.65	0.31
VGG19	0.77	0.38
Depthwise model	0.30	0.20
1D Conv NN	0.33	0.23
2D Conv NN	0.34	0.22
ResNet50 model	0.39	0.24
ResNet101 model	0.43	0.26
IMEFD-ODCNN	0.28	0.19
CSA-IDFLFD	0.10	0.12

Abbreviations: CSA-IDFLFD, Chameleon Swarm Algorithm with Improved Fuzzy Deep Learning for Fall Detection; TRT, training time; TST, testing time.

CONCLUSION

In this study, a new FD technique named CSA-IDFLFD technique has been developed for helping elderly people. The CSA-IDFLFD technique helps elderly persons with the identification of fall actions and improves their quality of life. The CSA-IDFLFD technique involves two phases of operations. In the initial phase, the CSA-IDFLFD technique involves the design of the IDFL model for the identification and classification of fall events. Next, in the second phase, the parameters related to the IDFL method can be optimally selected by the design of CSA. To validate the performance of the CSA-IDFLFD technique in the FD process, a widespread experimental evaluation process takes place. The extensive outcome stated the improved detection results of the CSA-IDFLFD technique.

CONFLICTS OF INTEREST

The authors declare no conflicts of interest in association with the present study.

REFERENCES

1. Benhari M, Hossseini R. 2022. An improved fuzzy deep learning (IFDL) model for managing uncertainty in classification of pap-smear cell images. Intell. Syst. Appl. Vol. 16:200133

2. Gharti P. 2020. A study of fall detection monitoring system for elderly people through IOT and mobile based application devices in indoor environment2020 5th International Conference on Innovative Technologies in Intelligent Systems and Industrial Applications (CITISIA); IEEE. Sydney, Australia: 25-27 November 2020; p. 1–9

3. Hassan MM, Gumaei A, Aloi G, Fortino G, Zhou M. 2019. A smartphone-enabled fall detection framework for elderly people in connected home healthcare. IEEE Netw. Vol. 33(6):58–63

4. Karar ME, Shehata HI, Reyad O. 2022. A survey of IoT-based fall detection for aiding elderly care: sensors, methods, challenges and future trends. Appl. Sci. Vol. 12(7):3276

5. Kchouri M, Harum N, Obeid A, Hazimeh H. 2022. Fuzzy support vector machine based fall detection method for traumatic brain injuries. Int. J. Adv. Comp. Sci. Appl. Vol. 13(11):302–314

6. Khraief C, Benzarti F, Amiri H. 2020. Elderly fall detection based on multi-stream deep convolutional networks. Multimed. Tools Appl. Vol. 79:19537–19560

7. Kong X, Meng Z, Nojiri N, Iwahori Y, Meng L, Tomiyama H. 2019. A HOG-SVM based fall detection IoT system for elderly persons using deep sensor. Procedia Comput. Sci. Vol. 147:276–282

8. Kumar CR, Rahman MK, Gilchrist ED, Pooja RL, Sruthi C. 2021. Smart band for elderly fall detection using machine learning. Nat. Vol. Essent. Oils J. 8269–8285

9. Kyriakopoulos G, Ntanos S, Anagnostopoulos T, Tsotsolas N, Salmon I, Ntalianis K. 2020. Internet of things (IoT)-enabled elderly fall verification, exploiting temporal inference models in smart homes. Int. J. Environ. Res. Public Health. Vol. 17(2):408

10. Pillai AS, Badgujar S, Krishnamoorthy S. 2022. Wearable sensor and machine learning model-based fall detection system for safety of elders and movement disordersProceedings of Academia-Industry Consortium for Data Science; Springer. Singapore: p. 47–60

11. Rahman MA, Ahsanuzzaman SM, Hasan A, Rahman I, Ahmed T, Kadir MM. 2020. Building a wheelchair controlling and fall detection system using mobile application2020 2nd International Conference on Advanced Information and Communication Technology (ICAICT); IEEE. Dhaka, Bangladesh: 28-29 November 2020; p. 213–218

12. Ramachandran A, Ramesh A, Karuppiah A. 2020. Evaluation of feature engineering on wearable sensor-based fall detection2020 International Conference on Information Networking (ICOIN); Barcelona, Spain: 7-10 January; IEEE. p. 110–114

13. Rezaee K, Khosravi MR, Moghimi MK. 2022. Intelligent elderly people fall detection based on modified deep learning deep transfer learning and IoT using thermal imaging-assisted pervasive surveillanceIntelligent Healthcare. Springer. Singapore: p. 113–132

14. Taghvaei S, Kosuge K. 2018. Image-based fall detection and classification of a user with a walking support system. Front. Mech. Eng. Vol. 13:427–441

15. Tahir A, Taylor W, Taha A, Usman M, Shah SA, Imran MA, et al.. 2022. IoT based fall detection system for elderly healthcareInternet of Things for Human-Centered Design. Springer. Singapore: p. 209–232

16. Vaiyapuri T, Lydia EL, Sikkandar MY, Díaz VG, Pustokhina IV, Pustokhin DA. 2021. Internet of things and deep learning enabled elderly fall detection model for smart homecare. IEEE Access. Vol. 9:113879–113888

17. Wang Y, Xiong W, Yang J, Wang S. 2020. A new fall detection method based on fuzzy reasoning for an omni-directional walking training robot. Int. J. Innov. Compu. Inf. Control. Vol. 16(2):597–608

18. Yadav SK, Luthra A, Tiwari K, Pandey HM, Akbar SA. 2022. ARFDNet: an efficient activity recognition & fall detection system using latent feature pooling. Knowl. Based Syst. Vol. 239:107948

19. Yu S, Chai Y, Chen H, Brown RA, Sherman SJ, Nunamaker JF Jr. 2021. Fall detection with wearable sensors: a hierarchical attention-based convolutional neural network approach. J. Manag. Inf. Syst. Vol. 38(4):1095–1121

20. Zhou W, Wang L, Han X, Wang Y, Zhang Y, Jia Z. 2023. Adaptive density spatial clustering method fusing Chameleon Swarm algorithm. Entropy. Vol. 25(5):782