Human Fall Detection from Sequences of Skeleton Features using Vision

Transformer

Ali Raza

1,2 a

, Muhammad Haroon Yousaf

1,2 b

, Sergio A. Velastin

3,4 c

and Serestina Viriri

5 d

Department of Computer Engineering, University of Engineering and Technology, Taxila, Pakistan

Swarm Robotics Lab, National Centre of Robotics and Automation (NCRA), Pakistan

School of Electronic Engineering and Computer Science, Queen Mary University of London, London E1 4NS, U.K.

Department of Computer Engineering, University Carlos III, 28911 Legan

es, Spain

School of Mathematics, Statistics & Computer Science, University of KwaZulu-Natal, Durban, 4041, South Africa

Keywords:

Computer Vision, Fall Detection, Vision Transformers, Event Recognition.

Abstract:

Detecting human falls is an exciting topic that can be approached in a number of ways. In recent years, several

approaches have been suggested. These methods aim at determining whether a person is walking normally,

standing, or falling, among other activities. The detection of falls in the elderly population is essential for

preventing major medical consequences and early intervention mitigates the effects of such accidents. How-

ever, the medical team must be very vigilant, monitoring people constantly, something that is time consuming,

expensive, intrusive and not always accurate. In this paper, we propose an approach to automatically identify

human fall activity using visual data to timely warn the appropriate caregivers and authorities. The proposed

approach detects human falls using a vision transformer. A Multi-headed transformer encoder model learns

typical human behaviour based on skeletonized human data. The proposed method has been evaluated on the

UR-Fall and UP-Fall datasets, with an accuracy of 96.12%, 97.36% respectively using RP normalization and

linear interpolation comparable to state-of-the-art methods.

1 INTRODUCTION

Human fall detection is a challenging problem that

may be tackled in several ways. To tackle the prob-

lem, several approaches have been suggested in re-

cent years. The majority of the existing methods in-

tend to recognize human actions or activities but not

limited to walking, standing or falling etc. Elderly

fall detection is vital among these activities. Accord-

ing to the World Health Organization, around 37.3

million people worldwide suffer fall-related injuries

that need medical care each year(Organization et al.,

2008). Falls are consistently among leading causes of

death all around the world. Although elderly people

are most in danger of terrible and deadly falls, chil-

dren and younger people are also at high risk when

they sustain fall injuries. A human fall can be inter-

preted as a sudden and signiﬁcant change in the hu-

man pose.

https://orcid.org/0000-0000-0000-0001

https://orcid.org/0000-0001-8255-1145

https://orcid.org/0000-0001-6775-7137

https://orcid.org/0000-0002-2850-8645

Human pose estimation is currently an area of ac-

tive interest for researchers. There are many recent

studies in the literature in this ﬁeld, especially on the

use of machine learning and computer vision meth-

ods. Various sensor-based approaches can be used

to detect different actions, such as walking, running,

standing, sitting, jumping, and falling. The use of

computer vision algorithms to detect actual falls has

caught interest due to its application in the health-

care sector. Most of the increases in healthcare costs

may be associated to the increases in the aging pop-

ulations in developed countries. This kind of ap-

plication has grown in popularity due to the imple-

mentation of safety measures at high-risk workplaces,

shopping malls, hospitals, nursing homes, and other

places. Falls are one of the most prevalent causes of

injuries, and the elderly are more susceptible to them.

Wearable sensors have been previously used for fall

detection, and barometers, inertial sensors, and gyro-

scopes have been used in some systems to identify

falls. However, the hardware components are expen-

sive, intrusive and not user-friendly to be worn by

people. Computer vision can be considered a more

Raza, A., Yousaf, M., Velastin, S. and Viriri, S.

Human Fall Detection from Sequences of Skeleton Features using Vision Transformer.

DOI: 10.5220/0011678800003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 5: VISAPP, pages

591-598

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

591

ﬂexible approach for fall detection and researchers

have reported many possible techniques and datasets.

Certain studies have demonstrated that videos can be

used to detect human activity in general. Vision-

based on skeletonized data of human joints, derived

from conventional images, are proving popular be-

cause they can protect privacy and be more robust to

varying environmental factors. These methods also

detect falling, walking, running, jumping, jogging,

etc. More people can now be recognized in a single

image. Probabilistic models are used in computer vi-

sion systems to detect falls. A computer vision-based

fall detection system uses deep learning or machine

learning to recognize and predict human falls based

on the human skeleton joint points rather than images

used as training data, reducing training time.

A summary of relevant existing research work in

the domain of human fall detection is given in Table

1. Article references are given in the ﬁrst column.

The second column shows whether a method can de-

tect activities other than falls. The third column indi-

cates whether a pose detection technique with skele-

ton detection process is employed. The fourth col-

umn indicates whether the algorithm has been imple-

mented to identify activities using image sequences.

The ﬁfth column speciﬁes if the detection technique

includes depth sensors. The sixth column shows the

model/technique applied in each work. The last col-

umn indicates which dataset was used in the respec-

tive research work. It can be observed that 60% of

the surveyed literature uses skeleton data for activity

detection, with OpenPose being the most frequently

pose estimator. Only 20% of the work can detect mul-

tiple activities. The most often chosen public datasets

for testing are UR-Fall 30%, UP-Fall 16%, and per-

sonal datasets 5%, while 70% of recent publications

describe the use of sequences for activity detection.

Considering the complex nature of the human fall

detection problem, this paper presents a vision trans-

former technique to solve the human fall detection

problem in a non-intrusive way. For this purpose,

BlazePose is used as a pose extractor and backbone

(although any other similar pose estimation method

may be used). After that, a vision transformer is

used as the head of the network. The proposed so-

lution eliminates the need to wear any physical sen-

sors. The approach takes less training time, helps pre-

serving users’ privacy, and increases machine learn-

ing algorithm efﬁciency and accuracy. The proposed

Figure 1: Fall sequence frames.

methodology uses linear interpolation to compen-

sate for missing joint points. A sequence of frames

is utilised to identify pose estimations that a single

frame cannot identify. Figure 1 shows a fall sequence.

The proposed approach uses UR-Fall and UP-Fall and

gives improved results that are compared with state-

of-the-art methods.

Table 1 show the existing literature of the hu-

man fall detection on visual data. Some of the exist-

ing work done using pose estimation and use LSTMs

and/or CNNs, but to the best of our knowledge vision

transformers have not been reported to solve human

fall detection using pose estimation.

The key contributions in this work are:

• A vision transformer is proposed for human fall

detection. Speciﬁcally, we use a multi-headed

transformer encoder network for the fall detec-

tion.

• Blazepose is used for pose estimation. RGB

videos are used to get skeleton features, which are

then fed into a multi-headed transformer encoder.

This encoder takes a sequence as input, uses a way

to pay attention, and can predict where a person

will fall.

• When it comes to position normalization and lin-

ear interpolation, missing keypoints in the skele-

tons can be ﬁxed with preprocessing techniques.

• The proposed approach is evaluated on the pub-

licly available UR-FALL and UP-FALL datasets,

and results are compared with state-of-the-art

methods.

2 PROPOSED APPROACH

The process ﬂow of the proposed technique is given

in Figure 2. As input, the network gets a series of

video frames, which are then processed by subsequent

blocks. Each of these blocks is explained in more de-

tail below.

2.1 Human Pose Estimation

The ﬁrst step in determining how people fall is to lo-

cate a person in video frames and determine the loca-

tion of their skeleton joints. Human skeletons behave

differently in the fall compared to conventional ac-

tions. The ability to locate these skeletons can aid the

detection of falls in videos. This work uses Blazepose

as a method for estimating human poses or skeletons.

Blazepose is a machine learning (ML) technique that

uses RGB video frames to ﬁgure out 33 landmarks

and a background segmentation mask. This gives a

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

592

Table 1: Summary of Human Fall Detection Works using Visual Data.

Ref

Multi

Activity

Skeleton Sequence

Cam

RGB

Model Dataset

(Chhetri et al., 2021a) No No Yes Yes CNN UR-Fall

(Chen and Duan, 2021) No No Yes Yes NanoDet-Lite UR-Fall

(Cai et al., 2021) No No Yes Yes MCCF UR-Fall

(Sultana et al., 2021) No No Yes Depth 2D CNN-GRU UR-Fall

(Leite et al., 2021) No No Yes Depth CNN-SVM UR-Fall

(Keskes and Noumeir, 2021) No No Yes Depth ST-GCN TST-Fall

(Tran et al., 2021) No No No Yes CNN-YOLO CMD-Fall

(Nguyen et al., 2021) No No No Yes KNN-SVM BOMNI

(Berlin and John, 2021) No own Yes Yes LSTM UP-Fall

(Galv

ao et al., 2021) No own Yes Yes AutoEncoder UP-Fall

(Guan et al., 2021) No OpenPose Yes Yes LSTM UP-Fall

(Kang et al., 2021b) No PoseNet Yes Yes GRU UP-Fall

(Lin et al., 2020) No OpenPose Yes Yes LSTM-GRU UP-Fall

(Chhetri et al., 2021b) No OpenPose Yes Yes SVM UP-Fall

(Dentamaro et al., 2021) No OpenPose Yes Yes SVM UP-Fall

(Chang et al., 2021) No OpenPose Yes Yes LSTM CMU

(Liu et al., 2021) No OpenPose Yes Depth RF SDU-Fall

(Kang et al., 2021a) No own Yes Yes GRU SDU-Fall

(Ramirez et al., 2021a) No AlphaPose No Yes KNN UP-Fall

(Wang et al., 2021) Yes(7) Yolov3 Yes Yes 3D CNN PKU-MMD

(Zhu et al., 2021) Yes(4) OpenPose Yes Depth DNN FDD

(Yin et al., 2021a) Yes(8) own Yes Depth MC-LSTM TST-Fall

(Ramirez et al., 2021b) Yes(12) AlphaPose No Yes

RF-SVM

MLP-KNN

UP-Fall

Figure 2: Proposed Approach Pipeline.

good estimate of the body’s pose. Figure 3 illustrates

the 33 landmarks.

Blazepose follows a two-step mechanism. First,

it detects the person. After the person is detected, it

calculates pose. Figure 4 shows a skeleton detected

from a video frame.

Human Fall Detection from Sequences of Skeleton Features using Vision Transformer

593

Figure 3: 33 full body landmarks.

Figure 4: Skeleton detected from a video frame.

2.2 Data Pre-Processing

Pre-processing is useful because data may not be col-

lected in the same unit or format. As a result, the

format must be changed, outliers must be removed,

missing values must be added, and features must be

scaled. Here are the steps that are taken to correct the

data format, normalize the data, and locate the miss-

ing values.

2.2.1 Relative Position Normalization

To improve the model’s accuracy, we use min-max

normalization to scale the data to the interval [0, 1]

to accelerate the model’s convergence. The new X

scale data are deﬁned as.

nom

X −X

min

max

− X

min

∈ [0, 1] (1)

The original data is represented by X, the lowest value

is represented by X

and the highest value is repre-

sented by X

However, the dataset contains missing points.

Even though not all groups have missing data, the

value 0 indicates that X

min

will equal 0. As a result,

the original data distribution will be changed, result-

ing in ambiguous attributes. Consequently, we sug-

gest the following normalization approach to convert

the initial coordinate position to the relative coordi-

nate positions in the f

frame.

(2)

and

(3)

where W and H are the width and height of the im-

ages, respectively. When x

or y

point equals 0, it is

considered to be a missing point and is not calculated.

The calculated center point will replace the missing

center point if it is a missing point. The distances

, y

) that need to be displaced are deﬁned as

dis

= x

− x

(4)

and

dis

= y

− y

(5)

The displacement is to the center point where the

eighth joint point is (x

, y

) and the central hip

circumference is used as the reference point (x

).When the displacement value is less than zero,

the object moves to the right. When the displacement

value is more than zero, the object goes to the left.

The new joint point coordinates (rx

, ry

) for

computing the displacement to the relative location

are then given as



− x

dis



(6)

and



− y

dis



(7)

Two advantages can be gained from this method.

First, the original data distribution is preserved, and

secondly the missing points do not need to be com-

puted simultaneously. In addition, any unnecessary

features should be eliminated. Aspects of the human

body displacement process, such as left-right or far-

near walking, ﬁx the human skeleton in the same po-

sition and give continuous motions to facilitate the

training of the model.

As shown in Figure 5, the image is resized to 640

by 480 pixels and its center is located. The human

skeleton will be moved to the image’s center after

displacement,as shown in Figure 5, and the central

hip-joint point will be similar to the position of the

center point. Some joints in the skeleton were con-

sidered irrelevant for distinguishing fall and non-fall

events. Thus, they were eliminated because an exces-

sive number of feature parameters may confuse the

process and diminish the recognition rate. So, the

only signiﬁcant points that are retained for training

are: the nose, shoulders, elbows, wrists, neck, hips,

knees, and ankles.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

594

Figure 5: After displacement, the location and number of

joint points.

Figure 6: Missing skeleton detected from a video frame.

2.2.2 Settlement for Missing Keypoints

BlazePose uses 2D image features to identify the

skeleton. Therefore, instability and incorrect predic-

tion errors might arise when these points are absent.

As seen in Figure 6, if the human body is obscured

or concealed, its contour is unclear, or if the upper

and lower limbs of the human body are occluded, it

will result in loss and errors during skeleton construc-

tion. The proposed approach uses linear interpola-

tion to adjust for the missing data. To ensure ade-

quate data, we designate 100 images as a group and

use the initial non-missing value data to interpolate

missing values in succeeding images. The interpo-

lation method may estimate the following data value

from the original data. If the initial number of joint

points is inadequate, the predicted missing value and

the real value will differ signiﬁcantly. Linear inter-

polation calculates the slope between two known data

points using two known data points. The approximate

values for Y are given by,

Y = Y

−Y

(X −X

) (8)

Figure 6 shows that the discrepancy between the es-

timated and real value is not statistically signiﬁcant

when the turning point is not missing. Before RP

normalization, the total dataset’s missing key points

accounted for 22.38%, and after RP-Normalization,

missing key points accounted for 13.4%. Moreover,

after interpolation + RP-Normalization, the missing

key points are only 1.8% of the total.

2.3 Multi-Headed Transformer Encoder

The multi-head self-attention mechanism is the ma-

jor component of the transformer encoder. In the

proposed technique, the transformer perceives the en-

coded representation of the input in the form of a se-

quence vector. Skeleton features are used to encode

the input.

Shape = [S, N, D] (9)

In this case, S represents the sequence dimension

number of each frame in the video, N represents the

stack of multiple videos into a batch, and D represents

the vector dimension, 512 in this case. The multi-

Figure 7: Multi-Head Transformer encoder.

head mechanism processes scaled dot-product atten-

tion numerous times in parallel rather than just once.

The encoder generates an attention-based representa-

tion, which can locate a single piece of data within an

inﬁnitely large context by concatenating and linearly

converting the separate attention outputs.

MultiHead(Q, K, V) = [ head

;. . . ; head

(10)

where head

= Attention



, KW

, VW



(11)

where W

, W

, and W

(12)

2.4 Linear Transformation Layer

In addition to the multiheaded transformer encoder, a

linear transformation layer is applied to the skeleton

regression to predict the output.

3 EXPERIMENTATION AND

RESULTS

3.1 Datasets

The UR-FALL (Alariﬁ and Alwadain, 2021) and UP-

FALL (Mart

ınez-Villase

nor et al., 2019a) datasets are

Human Fall Detection from Sequences of Skeleton Features using Vision Transformer

595

used in this research. Due to varying lighting condi-

tions, occlusion and the size of the dataset, the UR-

Fall dataset is challenging. This dataset is not of

real events but comprises actors performing every-

day activities and falling. It contains 70 sequences

(30 falls and 40 daily life activities). The dataset was

captured using two Microsoft Kinect cameras and an

accelerometer. The UP-FALL dataset is also pub-

licly available and contains 11 activities and different

stages in falls on three of those activities. Both of the

datasets have a frame rate of 25 frames per second,

and the resolution is 320×240 pixels.

3.2 Experimental Setup

The proposed approach is trained and tested on a Core

i7 machine equipped with an NVIDIA Quadro P5000

GPU. The extracted skeletons have 33 joints per hu-

man body and the Pytorch framework is used for the

implementation.

3.3 Results and Discussion

In the proposed approach, skeleton poses are ap-

plied to detect and distinguish between falling and not

falling activities. RP normalization with interpolation

is used to detect missing key points, and it provides

better results than those from the method without nor-

malization and interpolation. Table 3 shows accuracy,

precision, recall, and F1 score without normalization,

min-max normalization, and RP normalization with

interpolation for UR-FALL. In this case, linear inter-

polation + RP normalization increased accuracy, pre-

cision, and F1 score to 96.12%, 96.05%, and 96.04%

respectively.

The proposed approach is also used on the UP-

Fall dataset which gives slightly better results as

compared to the UR-Fall dataset because UR-Fall

dataset include two type of fall sitting from chair

and standing other activities are sitting crouching

down,sitting.On the other hand UP-Fall dataset has

more fall types which include forward using knees

forward,backwards sitting forward using hands other

activities are sitting,picking up an object,jumping lay-

ing,walking,standing and kneel down. Table 4 shows

accuracy, precision, recall, and F1 score without nor-

malization, min-max normalization, and RP normal-

ization with interpolation for UP-FALL. It shows

improvements in accuracy, recall and F1-score to

97.36%, 97.15% and 96.47% respectively.

Figure 9 shows the confusion matrices of all

falls and everyday normal events using different pro-

cesses. For UR-Fall without normalization, accuracy

is 94.87% and after min-max normalization accuracy

Figure 8: Confusion Matrices UR-FALL.

Figure 9: Confusion Matrices UP-FALL.

increases to 95.13% and after RP normalization there

is another small increase to 95.51% and after lin-

ear interpolation and RP normalization accuracy is

higher, 96.12%, as compared to the other methods.

Figure 10 illustrates a prediction of a fall and a daily

life normal activities where Figure 10a shows a ex-

ample of the detection of a human fall and Figure 10b

shows an example of a detection of a normal daily life

activity such as walking.

3.4 Comparison with Existing Work

In Table 4 , fall detection approaches are compared to

other reports in the literature . Chhetr et. al.(Chhetri

et al., 2021a) use a CNN method to detect human

falls. This method was reported to be 95% accurate

on the UR-Fall dataset, which is less accurate than

our method. A multiheaded transformer encoder with

RP normalization and interpolation is used in our pro-

posed approach to estimate human poses from skele-

ton sequences. In comparison with the state of the art

method, this method produces better accuracy results

of 96.12 percent in the UR-Fall dataset and 97.36 per-

cent in the UP-FALL dataset.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

596

Table 2: Evaluation results for the proposed method on UR-FALL Dataset.

Evaluation

Measures

Normalization

Min-Max

Normalization

Linear Interpolation

+RP Normalization

Accuracy 94.87 95.13 95.51 96.12

Precision 95.24 94.76 95.83 96.05

Recall 94.40 95.04 96.08 95.09

F1-score 94.82 95.01 95.57 96.04

Table 3: Evaluation results for the proposed method on UP-FALL Dataset.

Evaluation

Measures

Normalization

Min-Max

Normalization

Linear Interpolation

+RP Normalization

Accuracy 95.89 96.52 96.98 97.36

Precision 95.54 96.30 96.70 96.61

Recall 95.38 96.00 96.70 97.15

F1-score 95.47 96.15 96.26 96.47

Table 4: Comparison with other methods.

Methods Dataset Methodology

Skeleton

Sequence

Accuracy

(Chhetri et al., 2021a)

UR-Fall CNN Yes 95.00

(Chen and Duan, 2021)

UR-Fall NanoDet-Lite X 91.2

(Sultana et al., 2021)

UR-Fall CNN X 94.00

(Leite et al., 2021)

UR-Fall CNN-SVM X 94.96

(Mart

ınez-Villase

nor et al., 2019a). UP-Fall

RF,SVM

MLP,KNN

Yes

32.33,34.40

27.08,34.03

(Yin et al., 2021b)

UR-Fall SVM Yes 91.07

(Mart

ınez-Villase

nor et al., 2019b)

UP-Fall MLP,CNN X 95.0 ,95.1

Our Proposed UR-Fall

Multiheaded

Vision Transformer

Yes 96.12

Our Proposed UP-Fall

Multiheaded

Vision Transformer

Yes 97.36

Figure 10: Prediction results (a) corresponds to a human fall

sequence and (b) a normal sequence.

4 CONCLUSIONS

This paper has presented a computer vision-based fall

detection method using a vision transformer. A multi-

headed transformer uses skeletons to provide the de-

sired prediction. A multiheaded transformer encoder

network is used to model how a person falls based

on a human skeleton feature vector. Missing joint

points are addressed using min-max normalization,

RP normalization, and instate-of-the-artpared to the

work that has been done so far for the detection of hu-

man falls, we achieved state-of-the-art results. Future

work is planned to use Le2i for training and testing the

proposed network. We also plan to extend the work to

realistic (not acted out) environments including multi-

ple people, like shopping malls, public places, homes,

etc. Implementing an edge platform for real-time pro-

cessing is also in the future pipeline.

REFERENCES

Alariﬁ, A. and Alwadain, A. (2021). Killer heuristic op-

timized convolution neural network-based fall detec-

tion with wearable iot sensor devices. Measurement,

167:108258.

Berlin, S. J. and John, M. (2021). Vision based human fall

detection with siamese convolutional neural networks.

Journal of Ambient Intelligence and Humanized Com-

puting, pages 1–12.

Human Fall Detection from Sequences of Skeleton Features using Vision Transformer

597

Cai, X., Liu, X., An, M., and Han, G. (2021). Vision-based

fall detection using dense block with multi-channel

convolutional fusion strategy. IEEE Access, 9:18318–

18325.

Chang, W.-J., Hsu, C.-H., and Chen, L.-B. (2021). A pose

estimation-based fall detection methodology using ar-

tiﬁcial intelligence edge computing. IEEE Access,

9:129965–129976.

Chen, G. and Duan, X. (2021). Vision-based elderly fall

detection algorithm for mobile robot. In 2021 IEEE

4th International Conference on Electronics Technol-

ogy (ICET), pages 1197–1202. IEEE.

Chhetri, S., Alsadoon, A., Al-Dala’in, T., Prasad, P.,

Rashid, T. A., and Maag, A. (2021a). Deep learn-

ing for vision-based fall detection system: Enhanced

optical dynamic ﬂow. Computational Intelligence,

37(1):578–595.

Chhetri, S., Alsadoon, A., Al-Dala’in, T., Prasad, P.,

Rashid, T. A., and Maag, A. (2021b). Deep learn-

ing for vision-based fall detection system: Enhanced

optical dynamic ﬂow. Computational Intelligence,

37(1):578–595.

Dentamaro, V., Impedovo, D., and Pirlo, G. (2021). Fall de-

tection by human pose estimation and kinematic the-

ory. In 2020 25th International Conference on Pattern

Recognition (ICPR), pages 2328–2335. IEEE.

Galv

ao, Y. M., Portela, L., Ferreira, J., Barros, P., Fagun-

des, O. A. D. A., and Fernandes, B. J. (2021). A

framework for anomaly identiﬁcation applied on fall

detection. IEEE Access, 9:77264–77274.

Guan, Z., Li, S., Cheng, Y., Man, C., Mao, W., Wong, N.,

and Yu, H. (2021). A video-based fall detection net-

work by spatio-temporal joint-point model on edge

devices. In 2021 Design, Automation & Test in Eu-

rope Conference & Exhibition (DATE), pages 422–

427. IEEE.

Kang, Y., Kang, H., and Kim, J. (2021a). Fall detection

method based on pose estimation using gru. In Inter-

national Conference on Software Engineering, Artiﬁ-

cial Intelligence, Networking and Parallel/Distributed

Computing, pages 169–179. Springer.

Kang, Y. K., Kang, H. Y., and Weon, D. S. (2021b). Hu-

man skeleton keypoints based fall detection using gru.

Journal of the Korea Academia-Industrial Coopera-

tion Society, 22(2):127–133.

Keskes, O. and Noumeir, R. (2021). Vision-based fall de-

tection using st-gcn. IEEE Access, 9:28224–28236.

Leite, G. V., da Silva, G. P., and Pedrini, H. (2021). Three-

stream convolutional neural network for human fall

detection. In Deep Learning Applications, Volume 2,

pages 49–80. Springer.

Lin, C.-B., Dong, Z., Kuan, W.-K., and Huang, Y.-F.

(2020). A framework for fall detection based on open-

pose skeleton and lstm/gru models. Applied Sciences,

11(1):329.

Liu, C., Lv, J., Zhao, X., Li, Z., Yan, Z., and Shi, X.

(2021). A novel key point trajectory model for fall

detection from rgb-d videos. In 2021 IEEE 24th Inter-

national Conference on Computer Supported Coop-

erative Work in Design (CSCWD), pages 1021–1026.

IEEE.

Mart

ınez-Villase

nor, L., Ponce, H., Brieva, J., Moya-Albor,

E., N

nez-Mart

ınez, J., and Pe

nafort-Asturiano, C.

(2019a). Up-fall detection dataset: A multimodal ap-

proach. Sensors, 19(9):1988.

Mart

ınez-Villase

nor, L., Ponce, H., Brieva, J., Moya-Albor,

E., N

nez-Mart

ınez, J., and Pe

nafort-Asturiano, C.

(2019b). Up-fall detection dataset: A multimodal ap-

proach. Sensors, 19(9):1988.

Nguyen, V. D., Pham, P. N., Nguyen, X. B., Tran, T. M., and

Nguyen, M. Q. (2021). Incorporation of panoramic

view in fall detection using omnidirectional camera.

In The International Conference on Intelligent Sys-

tems & Networks, pages 313–318. Springer.

Organization, W. H., Ageing, W. H. O., and Unit, L. C.

(2008). WHO global report on falls prevention in

older age. World Health Organization.

Ramirez, H., Velastin, S. A., Fabregas, E., Meza, I., Makris,

D., and Farias, G. (2021a). Fall detection using human

skeleton features.

Ramirez, H., Velastin, S. A., Meza, I., Fabregas, E., Makris,

D., and Farias, G. (2021b). Fall detection and activ-

ity recognition using human skeleton features. IEEE

Access, 9:33532–33542.

Sultana, A., Deb, K., Dhar, P. K., and Koshiba, T. (2021).

Classiﬁcation of indoor human fall events using deep

learning. Entropy, 23(3):328.

Tran, T.-H., Nguyen, D. T., and Nguyen, T. P. (2021). Hu-

man posture classiﬁcation from multiple viewpoints

and application for fall detection. In 2020 IEEE

Eighth International Conference on Communications

and Electronics (ICCE), pages 262–267. IEEE.

Wang, K., Li, X., Yang, J., Wu, J., and Li, R. (2021).

Temporal action detection based on two-stream you

only look once network for elderly care service robot.

International Journal of Advanced Robotic Systems,

18(4):17298814211038342.

Yin, J., Han, J., Xie, R., Wang, C., Duan, X., Rong, Y.,

Zeng, X., and Tao, J. (2021a). Mc-lstm: Real-time 3d

human action detection system for intelligent health-

care applications. IEEE Transactions on Biomedical

Circuits and Systems, 15(2):259–269.

Yin, J., Han, J., Xie, R., Wang, C., Duan, X., Rong, Y.,

Zeng, X., and Tao, J. (2021b). Mc-lstm: Real-time 3d

human action detection system for intelligent health-

care applications. IEEE Transactions on Biomedical

Circuits and Systems, 15(2):259–269.

Zhu, N., Zhao, G., Zhang, X., and Jin, Z. (2021). Falling

motion detection algorithm based on deep learning.

IET Image Processing.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

598