Real-World Case Study of a Deep Learning Enhanced Elderly Person

Fall Video-Detection System

Amal El Kaid

1,2,∗ a

, Karim Ba

ına

1 b

, Jamal Ba

ına

3 c

and Vincent Barra

2 d

Universit

e Clermont-Auvergne,CNRS, Mines de Saint-

Etienne, Clermont-Auvergne-INP, LIMOS,

63000 Clermont-Ferrand, France

Alqualsadi Research Team, Rabat IT Center, ENSIAS, Mohammed V University in Rabat, 10112, Rabat, Morocco

Angel Assistance, 57070, Metz, France

∗

Keywords:

Neural Networks, Fall Detection, Fall Classiﬁcation, Real-World Fall Detection System, Reduce False

Positives.

Abstract:

Recent large and rapid growth in the healthcare sector has contributed to an increase in the elderly population

and an increase in life expectancy. One of the important study topics in this ﬁeld is the automatic fall detection

system. Camera-video has been extensively employed recently for applications in surveillance, the home,

and healthcare. Therefore a smart fall detection system is focusing on image and video analysis techniques.

For that, our scientiﬁc work studied an actual vision-based fall detection system. It produces satisfactory

outcomes, but there is still room for improvement. The system has a very high recall rate and can detect all

falls, but it lacks precision and frequently reports false positives (more than 99 per-cent). In fact, due to the

optimum camera quality, several ordinary activities with speciﬁc movements, such as wheelchair mobility, or

the light changing in an empty room, can be mistaken for falls. To address this problem and increase precision,

we propose a post-process approach, hybridizing a CNN model and a Haar Cascade Classiﬁer to determine

whether to conﬁrm or reject an alert that has been identiﬁed as a fall. The system’s effectiveness will increase

while the false positives are decreased.

1 INTRODUCTION

The elderly population is growing more in the world

recently. The elderly are living longer and are getting

more numerous than ever, due to many reasons, in-

cluding the adoption of modern technologies to lessen

health issues. One common and major health prob-

lem that elderly people face is falling. Thousands of

them are victims of fall incidents. Chan et al. (Chan

et al., 2007) estimate that one-third of home-living

adults aged 75 or more experience a fall each year.

Falls cause variable physical consequences depend-

ing on the senior, but they frequently result in serious

injuries including hip fractures or even death. In fact,

falling is one of the ﬁve most common causes of death

among the elderly population (Vishwakarma et al.,

2007). World Health Organization (World Health Or-

ganization, ) estimated that 646 000 fatal falls occur

https://orcid.org/0000-0003-0605-8919

https://orcid.org/0000-0002-4736-1079

https://orcid.org/0000-0003-1933-4708

https://orcid.org/0000-0002-8975-222X

∗

Corresponding author

each year, making it the second leading cause of unin-

tentional injury death, after road trafﬁc injuries. Fur-

thermore, fall victims suffer from severe psychologi-

cal effects, such as the loss of self-conﬁdence (Vignat,

2001), which is one of the most reasons that don’t en-

courage the aged population to live alone and main-

tain an independent lifestyle. As a result, it is highly

recommended to adopt technological innovations like

smart systems to help address these problems and pre-

vent any dangerous injuries by minimizing the time

spent lying down on the cold ﬂoor for several hours

or even days after a fall incidence has occurred. Be-

cause of this, the use of technological advancements,

such as smart systems, is strongly recommended to

help address these problems and to avoid any danger-

ous injuries by reducing the period of laying down

on the cold ﬂoor for several hours or even days af-

ter a fall incident has occurred. This makes devel-

oping an effective fall detection system a major chal-

lenge for the health care research community. That

is the primary focus of this paper. There are several

types of systems for automatic fall detection, as de-

scribed in Section 2. There are several types of sys-

El Kaid, A., Baïna, K., Baïna, J. and Barra, V.

Real-World Case Study of a Deep Learning Enhanced Elderly Person Fall Video-Detection System.

DOI: 10.5220/0011674800003417

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

575-582

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)

575

tems for automatic fall detection, which we will de-

scribe in Section 2. Including wearable devices and

environmental sensors. However, they are unfavor-

able choices for the people, they consider them blind

solutions because they are unable to communicate the

severity level of a fall to provide the required assis-

tance. The camera-based system is the most reliable

and trustful method of detecting falls and delivering

timely aid. Despite the success and the trust of people

in this direction, it still faces many challenges. Over-

coming one of them is our objective in this study. We

based on a real case study in a French company of

fall detection camera-based system, to improve it. We

found out that because of the video’s optimal resolu-

tion and room’s light condition, there are some daily

activities with certain movements such as mobility on

a wheelchair or even the light changing in an empty

room that can be mistaken for falls, which decreases

the system’s precision. We focused on how always

returning positive falls to maximize the system preci-

sion. For this, we proposed to reduce generated false-

positive images. We presented an image processing

algorithm that hybridizes a Convolution Neural Net-

work model and a Haar cascade classiﬁer, which will

be added to the fall detection algorithm to ﬁlter the

alert image before being sent. The organization of

this research paper is as follows. Section 2 will give

an overview of fall human detection systems. Section

3 will present a real-use study of a fall video detec-

tion system in a French company and discuss one of

its weakest points, then we will propose a solution to

overcoming this problem and improving this system.

Section 5 will cover the experimental results obtained.

A conclusion of the work is formulated in section 6.

2 OVERVIEW OF HUMAN FALL

DETECTION SYSTEMS

Recently, a huge amount of research is proposed to

solve the problem of fall detection. This problem can

requires using one of different types of devices in or-

der to collect data when attempting to automatically

detect a fall (Guti

errez et al., 2021; Rastogi and Singh,

2021; Berlin and John, 2021). They can be catego-

rized into invasive devices that must be maintained

by the user and passive devices which continue to op-

erate with minimal maintenance.

In the state-of-the-art of fall detection, researchers

tend to base there approaches on structured predic-

tion, supervised learning, clustering, outlier detection,

dimensional reduction, and neural networks. Through

the last decades, many papers reviewed the evolution

of existing human fall detection systems. The ﬁrst

survey was published in 2008 by Noury et al. (Noury

et al., 2007), they reviewed the sensor-based sys-

tems for automatic fall detection by citing all previous

works that deal with the accelerometer approach as

well as those done with image processing techniques.

Thousands of new techniques have appeared af-

ter the survey published by M.A.Habib et al. (Habib

et al., 2014) in 2014, and yet no survey paper has

been published to group these new approaches un-

til 2018 when T.XU et al. (Xu et al., 2018) updated

the survey of fall detection methods focusing on those

done from 2014. They have selected the twenty most

highly cited articles to discuss profoundly their ideas

and theories and analyse them from three aspects:

sensors, algorithms, and performance. In the same

year, Y. Birku et al. (Birku and Agrawal, 2018) pub-

lished a survey of various fall detection systems and

methods. They detailed the most used approaches in

the fall detection system, which include the wearable,

ambiance, and camera-based devices. Another survey

paper in the human fall detecting domain was written

by S.K. Jarraya et al. (Jarraya, 2018). They consid-

ered only camera-sensor based approaches because of

their performance versus other approaches, they gave

an overview of the related works on fall detection that

used Kinect camera and discussed their limitation and

advantages. Indeed, we will give, in short, a global vi-

sion of these different systems and their limits. First,

the traditional medical alert systems are manual alarm

systems that require the participation of the person.

The alarm buttons, existing in the form of a necklace

or a bracelet to carry on, allow the person to press

on to alert of a fall. The idea behind this type is to

avoid false alarms and to allow the person to feel re-

assured to have at all time a wearing device. But these

devices must be worn to function. This is the main

cause of their lack of effectiveness. In 1981, Wild et

al. (Wild et al., 1981) conducted a study on people

over 65 years old for one year. They found that of

the nine fallers with an alert system, only two man-

aged to use it to alert after falling. This inability to

press the button may be due to the person’s uncon-

sciousness or shock. So even though the senior used

this system, he remains unprotected and rescued in

the event of a sudden fall followed by a loss of con-

sciousness that he could not trigger an alert to the help

desk manually. That is what push the researchers to

develop an automatic alarm system which can be de-

ﬁned as an assistive device whose main objective is to

alert when a fall event has occurred. The most known

commercial fall detection systems are mostly classi-

ﬁed into three categories; those based on wearable

devices, those based on ambiance sensors, and those

based on video processing (Yu, 2008; Mubashir et al.,

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

576

Figure 1: Video fall detection system.

2013). The Wearable device based systems sensi-

tively provide automatic detection, thanks to embed-

ded sensors (accelerometer, triaxial accelerometer, bi-

ological signals) and new algorithms which recognize

the activities. They allow measuring changes in ori-

entation, motion, and location of the body of the sub-

ject to detect the physical activity so that can identify

the fall event. Once a fall is recorded and not inter-

rupted by the user, the system immediately contacts

the health services to allow early intervention. How-

ever, the main drawback of these systems is that they

ﬁxed relative relations with the object, which could

be easily disconnected. In addition, carrying around

devices all the time and wearing them is not conve-

nient, and elderly individuals may simply forget to

put them on. The ambiance sensors came to over-

come this problem, they detect the plantar pressures

exerted by people to ensure that the person is not ly-

ing on the ﬂoor. They rely on proximity and ﬂoor sen-

sors to collect data on the daily living activities of the

senior. They use motion, light, and vibration sensors

and combine visual and sound information to detect

a fall. However, it has the signiﬁcant disadvantage

of sensing the pressure of everything in and around

the object and generating false alarms in the case of

fall detection, which leads to poor detection accuracy.

Furthermore, nothing can guarantee if it is a real fall

or not and the degree of severity if so. The last cate-

gory and not the least is Vision-based systems, which

analyse algorithms for images, sounds, and videos.

Recently, camera-video has been increasingly pop-

ular for surveillance, home, and healthcare applica-

tions. Since they tend to deal with intrusion better

than other approaches. They can integrate many im-

plemented algorithms and open-source libraries that

detect the person’s falls. Only this type of system

raises the doubt whether it is a real fall or not, so

it allows for avoiding unnecessary interventions and

minimizes the overall cost of the service and provides

the exact cause of human falling. And unlike the oth-

ers, they make it possible to have an idea about the

severity level of a fall event, so that, a necessary res-

cue will have done. The Cameras must be installed

in several rooms to cover the whole area of actuation.

When a fall is detected, the system sends the images

of this event to the help desk.

3 REAL CASE STUDY OF

VISION-BASED FALL

DETECTION SYSTEM

3.1 Context and Objective of the Work

Angel Assistance is a French innovative company

launched in 2013 to develop a new technique and a

complete video fall detection system to help depen-

dent elderly in their home from a camera mounted

in a patient room. Once a fall is detected, the sys-

tem sends automatically and immediately an image

alert captured of the senior’s situation to the company

call center to be treated and veriﬁed by human agents,

whether is a true fall or a false one to perform the nec-

essary rescue. The false positives are events similar to

the characteristics that deﬁne the falls.

Real-World Case Study of a Deep Learning Enhanced Elderly Person Fall Video-Detection System

577

3.2 Data Acquisition

An RGB camera and an Edge computer are used for

data acquisition. In fact, the Hidden RGB camera

equipped with infrared LED for night vision used

as presence detector (Ismaili-Alaoui et al., 2022;

Ismaili-Alaoui et al., 2019). And Edge is a local

server that analyzes video feeds from all local cam-

eras. It detects any drop alerts in order to be sent to

the operations center. Indeed, the images captured by

the camera are analysed in the Edge computer to de-

termine whether or not a fall has occurred. The sys-

tem alerts the center of the detected fall, which is then

checked and stored, by a center agent, in one of four

classes: false alerts including empty room, false alerts

including active person, true alerts with average risk

level including seated person, and true alerts with a

high level of risk, which includes people lying down.

Figure 1 illustrates the process of video fall.

We studied data gathered by the system over seven

months, containing about 26 551 images. We have

observed that 99% of received alerts are false posi-

tives, and only seven are real alerts. Four falls with a

seated person, three falls with a lying person, 16972

images of empty rooms and 9572 images of active

people.

3.3 Data Analysis

Various machine learning models were investigated

to increase the accuracy of automatic fall detection.

For example, (

Ozdemir and Barshan, 2014) used

K-Nearest Neighbor classiﬁer or (Karantonis et al.,

2006) used Support Vector Machines (SVM) that ac-

curately detect falls based on wearable motion sen-

sors. However, as mentioned in the section 2, they

pose problems because the elderly must always wear

them, therefore they are insufﬁcient to provide re-

liability in real-world circumstances (Galv

ao et al.,

2021). In the other hand, these machine learning clas-

siﬁers have also been proposed for solutions based on

surveillance cameras and computer vision techniques.

For example, Feng et al. (Feng et al., 2014) clas-

siﬁed human postures identiﬁed from extracted sil-

houettes using a multi-class support vector machine.

Also, Galv

ao et al. (Galv

ao et al., 2017) tested MLP,

KNN and SVM classiﬁers. All these techniques pro-

duced accurate results. However, the current issue is

all these studies results represent only reﬂect the ac-

curacy of sensitivity, which measured the system’s ca-

pacity to detect falls, and completely ignore the accu-

racy of speciﬁcity, which measures the system’s ca-

pacity to prevent false alerts (Aziz et al., 2017).

Indeed, sensitivity refers to the ratio of the true posi-

tives to the total number of falls as shown in equation

1. Conversely, speciﬁcity is determined by the ratio of

true negatives to the total number of discarded trials,

as shown in equation 2, c .

Sensitivity =

TruePositive

TruePositive + FalseNegative

(1)

Speci f icity =

TrueNegative

TrueNegative + FalsePositive

(2)

The false positive rate, which corresponds to the

number of false alerts, could also be considered as a

signiﬁcant evaluation metric 3.

Falsepositiver ate =

FalsePositives

Time

(3)

While Time is given in hours. Aziz O. et al. (Aziz

et al., 2017) is one of a few papers that measured

all these evaluation performance, like the angel sys-

tem, they detect falls in elderly datasets, their algo-

rithm based on SVM classiﬁer showed 80% sensi-

tivity, 99.9% speciﬁcity and false positive rates from

0.05 to 0.15 false alarms per hour. Similarly, the

best performing algorithm reported by Bagal

a et al.

(Bagala et al., 2012) which showed 83% sensitivity,

97% speciﬁcity and 0.21 false alarms per hour. The

Angel system gives, for the real-world data set ob-

tained, as previously mentioned, from elderly people,

a sensitivity of 100%, a speciﬁcity of 95% and 0.02

false alarms per hour.

The primary metrics which determined the sys-

tem performance are precision and sensitivity/recall

(Grandini et al., 2020). The precision is computed as

the proportion of the total relevant results returned to

the total number of results returned. In other words,

the ratio of real falls to the total of falls returned by

the system 4. Such a system, with high recall and

low precision , is referred in a scientiﬁc literature as

a recall -oriented system in contrast to the precision

-oriented system with low recall but with high preci-

sion as they talk a lot but make a lot of mistakes. An

“ideal” system provides all the information required

and nothing more (high recall and high precision).

Precision =

TruePositives

TruePositives + FalsePositives

(4)

Therefore, the main goal of this work is to enhance

the system by increasing precision without reducing

sensitivity. Thus, we propose reducing the number of

false positives.

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

578

Figure 2: Diagram of error minimizing by alerts ﬁltering using Haar cascade and CNN.

4 PROPOSED IMPROVEMENTS

TO THE VISION-BASED FALL

DETECTION SYSTEM

In the previous section 3, we noted that many fall de-

tection systems suffer from low performance (high

false positive rate 3). An example of these systems

is Angel Assistance which suffers from a very high

number of false positives (99% of detected images are

false). In this section, we propose to help improve this

system by reducing the number of these false alarms.

Our proposition is deﬁned by a post-processing step

that is added after the alert is given by the fall detec-

tion system. This approach combines the Haar cas-

cade and a proposed CNN classiﬁer to help determine

if the alert is false or not (see ﬁgure 2). The combi-

nation of the two approaches comes in two integrative

steps. First, the haar cascade is applied to determine if

the received image contains a person or not (the room

is empty). If it does, the image will be transferred

to the CNN classiﬁer to determine ﬁnally if there is

a fall or not.Before we apply this combination of ap-

proaches, a data preparation, and processing step is

realised.

First, we annotated and visualized false positive data.

This phase helped us to more clearly identify the

kinds of circumstances that the system has not been

able to separate from fall-related ones. We manually

labeled roughly 2050 over 9572 images in the cate-

gory ”active person,” sub-classing them into 4 sub-

categories: 10 images of people using walkers, 1080

images of people using wheelchairs, 210 images of

people using chairs, and the remaining 748 images.

In fact, a prior paper (El Kaid et al., 2019b) ex-

amined and offered the idea of reducing false pos-

itives by employing a convolutional neural network

to get rid of those of a person in a wheelchair. But

as indicated in its perspectives, there are still a lot of

false positives. For example, none of the CNN mod-

els evaluated could distinguish empty room images

because of the complexity of the images, variations

in the room’s lighting, and video resolution. We came

up with the notion to further increase the fall detection

algorithm’s accuracy at that point. Indeed, the idea

presented in the earlier paper to address this request

will be improved by the proposed method. Since we

don’t want to affect the main system of the fall de-

tection, we propose an algorithm that will be used as

post-processing on the system after the fall detection

phase.

This algorithm is applied to the outcomes of the

video-based fall detection algorithm or the warnings

that have been identiﬁed. Post-processing thus anal-

yses the static images, and it is divided into two

main components and produces results instantly. The

ﬁrst part uses the Haar-like features method to detect

whether an image contains a human or not. If so, a

CNN model will process the image and assign it to

one of the following two classes: ”person sitting on

a wheelchair”, ”person not sitting on chair”, or ”oth-

ers”. Then, only images from the last class will be

assigned to provide desk assistance. They could be

true falls or false ones (some images of an active per-

son). But still, a large portion of these alerts (”empty

room”, ”a person sitting on a wheelchair”, and ”per-

son sitting on a chair”) are erased.

Human Detection Using Haar-Like Features. The

ﬁrst part of the algorithm, which is based on the

human detection technique, aims to distinguish the

“Empty room” image from others (“False positive:

Active person”, “Real fall”). The goal is to keep

only the alerts that include a human and ignore any

alerts that do not. First, we tested the most pop-

ular algorithm for this task, extracting the features

by HOG (Histogram of Oriented Gradients) descrip-

tor and classifying them using SVM. But this ma-

chine learning algorithm was unable to detect peo-

ple in images well due to noise from the surveil-

lance cameras. Then, we attempted to use pre-trained

OpenCV classiﬁers saved in XML ﬁles to detect hu-

mans using Haar-like features. These ﬁles are haar-

cascade fullbody, haarcascade lowerbody and haar-

cascade upperbody.

Real-World Case Study of a Deep Learning Enhanced Elderly Person Fall Video-Detection System

579

Figure 3: The binary CNN classiﬁer architecture.

Yet, none of these ﬁles are compatible with our

data. The use of these ﬁles necessitates a dis-

tinct persona, which does not satisfy our condi-

tions. In practice, after experimenting with numer-

ous OpenCV Haar Cascade Classiﬁers, we found

three XML ﬁles that we could use to identify the

presence of a person in our images, these ﬁles

are : haarcascade frontalcatface extended, haarcas-

cade frontalcatface and lbpcascade frontalcatface.

We employ the detectMultiScale technique from

CascadeClassi f ier approach to locate the subject in

the image. These three detectors are utilized to deter-

mine whether or not a person is present in the image.

If this classiﬁer indicates that there is no one in the

new alert, it is assumed that this image represents an

empty room and the warning is canceled. By doing

this, we will greatly reduce the number of false posi-

tives and enhance the fall detection system.

Convolution Neural Network Model. To build a

convolution neural network, we follow the architec-

ture outlined in the graph of the ﬁgure 3.

First, we load the CNN model from our previous pa-

per (El Kaid et al., 2019b). It was a binary classiﬁer

that was saved in an HDF5 (Hierarchical Data Format

version 5) ﬁle that was used to determine whether or

not an alert contained a person using a wheelchair,

which provides a good accuracy of 98%. Using this

approach, we can get rid of the images that show a

person in a wheelchair, which accounts for around

17% of false positives. Second, we considered cre-

ating a new multi-class CNN classiﬁer to categorize

alerts into the four categories of ”person sitting on

a wheelchair”, ”person sitting on a chair”, ”empty

room”, and ”other”. Due to a lack of data, we cannot

incorporate the ”images with a walker” class. Several

models have been tested, although due to the similar-

ity of the images of the different classes, they gen-

erally produce less effective results, as shown in our

previous research (El Kaid et al., 2019a). The best

performance, in this case, was 89%, which is why we

considered including the previously described human

detection system. However, we build another multi-

class CNN classiﬁer for the three categories above ex-

cept ”empty room”.

The algorithm 1 explains the proposed post-

processing approach to eliminate false positives.

Algorithm 1: Post-processing to ﬁlter alerts.

Begi n

FilterAlert ( Mat alert image)

f aceExtendedCas ← C a s c a d e Cl a s si f i e r

(" haarc a s c a d e_ fr o nt al ca tf a ce _e xt en d ed . xm l ")

f rontalcat f aceCas ← C a s ca d e C la s s if i e r

(" h aa rc a sc a de _ fr on t al c at fa c e . xm l ")

f rontalcat f aceLbp ← CascadeCla s s i f i e r

(" l bp c as ca d e_ f ro n ta l ca tf a ce . xml ")

detect ← d e t e c t M u lt i S c a le ()

/* detect contains location and size

of the bounding box (x,y,w,h)

if a person is detected

d ← 0

If detect is not None

Retu r n d ← d + 1

EndI f

If d > 3 //A person is detected

model = l oad ( f ilter model ) // Load CNN

img resized = pre - proces s alert image

// Resize the input image

result ← m odel . predict (img resized)

If result[0][0] = 1

i.e: the image contains

a person on a wheelchair

Exi t //Cancel alert sending

Els e

sen d (alert image) // Confirm

alert sending

EndI f

Els e

if no person is detected,

it is an Empty room.

Exi t // Cancel alert sending

EndI f

EndFunction

End

5 EXPERIMENTATION RESULTS

This article outlines how an alert is processed and ﬁl-

tered before it is transmitted. This section contains

some of the outcomes of the proposed post-process.

When the algorithm is tested on 5254 images, the

Haar cascade classiﬁer, which we use as the ﬁrst step

of our approach to ﬁlter out alerts containing ”empty

room”, provided 76% accuracy.

As can be seen from the ﬁgure 5, the Haar cas-

cade classiﬁer succeeds in identifying people in the

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

580

(a) False alarm : A per-

son in a wheelchair

(b) False alarm: An

empty room

son in a chair

(d) False alarm : A per-

son in a wheelchair

Figure 4: detected Alerts by the fall video detection algo-

rithm

(a) Person detected (b) Person detected

Figure 5: The outcomes of a classiﬁer using a Haar cascade

classiﬁer.

images and marks their presence with a green rect-

angle (bounding box) around them while eliminating

the image of an ”empty room”. The second ﬁlter, a bi-

nary classiﬁcation model, uses the output of the ﬁrst

step as input and removes the false-positive category

of ”person sitting in a wheelchair” with an accuracy

of 98%, which enhances the fall detection system.

The algorithm’s ﬁnal ﬁndings are shown in the

ﬁgure 6. The categorical model, on the other hand,

allows us to do away with the categories of ”person

sitting in a wheelchair” and also ”person sitting in a

chair”.

Even if its accuracy, 82%, is less than that of the

binary classiﬁer, it is still respectable. AlexNet model

is the network that was trained for it. We could en-

hance the elderly person fall video-detection algo-

rithm by hybridizing one of the CNN classiﬁers with

a Haar cascade classiﬁer, maximizing its precision

while minimizing the number of false positives sent

by eliminating three categories of false alarms: an

(a) Prediction: person is

on a wheelchair

(b) Prediction: person is

on a wheelchair

not on a wheelchair

Figure 6: The ﬁnal results of Algorithm.

empty room, a person in a wheelchair, and a person

in a chair.

6 CONCLUSION AND

PERSPECTIVES

To summarize, when it comes to fall detection de-

vices and methodologies, all sensors, whether worn

or embedded in people’s environments, are all blind

and send fall alerts just like any other detector. Only

image-based methods can verify the reality of the fall

and avoid unnecessary assistance. However, we rec-

ognize that there is room for improvement.

We could enhance the elderly person fall video-

detection algorithm by hybridizing a Haar cascade

classiﬁer with a CNN classiﬁers. This improve-

ment could maximize the system’s precision by min-

imizing the number of false positives sent. Indeed,

we proposed an algorithm that eliminates three cate-

gories of false alarms: an empty room, a person in a

wheelchair, and a person in a chair.

In the upcoming work, due to the proximity between

images of different classes and to the lighting varia-

tion, we propose to develop a video-based fall detec-

tion system based on our previous 3D pose estimation

approach (El Kaid et al., 2022) to more accurately rec-

ognize the fall events .

ACKNOWLEDGEMENTS

The authors wish to thank Angel Assistance for pro-

viding us with the necessary data to accomplish our

work. This work is ﬁnanced by CIFRE France-

Maroc, ANRT, France and CNRST, Morocco.

Real-World Case Study of a Deep Learning Enhanced Elderly Person Fall Video-Detection System

581

REFERENCES

Aziz, O., Klenk, J., Schwickert, L., Chiari, L., Becker, C.,

Park, E. J., Mori, G., and Robinovitch, S. N. (2017).

Validation of accuracy of svm-based fall detection

system using real-world fall and non-fall datasets.

PLoS one, 12(7):e0180318.

Bagala, F., Becker, C., Cappello, A., Chiari, L., Aminian,

K., Hausdorff, J. M., Zijlstra, W., and Klenk, J.

(2012). Evaluation of accelerometer-based fall de-

tection algorithms on real-world falls. PloS one,

7(5):e37062.

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.

Birku, Y. and Agrawal, H. (2018). Survey on fall detection

systems. International Journal of Pure and Applied

Mathematics, 118(18):2537–2543.

Chan, B. K., Marshall, L. M., Winters, K. M., Faulkner,

K. A., Schwartz, A. V., and Orwoll, E. S. (2007).

Incident fall risk and physical activity and physical

performance among older men: the osteoporotic frac-

tures in men study. American journal of epidemiology,

165(6):696–703.

El Kaid, A., Ba

ına, K., and Ba

ına, J. (2019a). Real

case study of improving elderly person fall video-

detection algorithm. In Conference on Complexity

Analysis of Industrial Systems and Advanced Model-

ing - CAISAM’19, Ben Guerir, Morocco.

El Kaid, A., Ba

ına, K., and Ba

ına, J. (2019b). Reduce false

positive alerts for elderly person fall video-detection

algorithm by convolutional neural network model.

Procedia computer science, 148:2–11.

El Kaid, A., Brazey, D., Barra, V., and Ba

ına, K.

(2022). Top-down system for multi-person 3d abso-

lute pose estimation from monocular videos. Sensors,

22(11):4109.

Feng, W., Liu, R., and Zhu, M. (2014). Fall detection for el-

derly person care in a vision-based home surveillance

environment using a monocular camera. signal, image

and video processing, 8(6):1129–1138.

Galv

ao, Y. M., Albuquerque, V. A., Fernandes, B. J., and

Valenc¸a, M. J. (2017). Anomaly detection in smart

houses: Monitoring elderly daily behavior for fall de-

tecting. In 2017 IEEE Latin American Conference

on Computational Intelligence (LA-CCI), pages 1–6.

IEEE.

Galv

ao, Y. M., Ferreira, J., Albuquerque, V. A., Barros, P.,

and Fernandes, B. J. (2021). A multimodal approach

using deep learning for fall detection. Expert Systems

with Applications, 168:114226.

Grandini, M., Bagli, E., and Visani, G. (2020). Metrics for

multi-class classiﬁcation: an overview. arXiv preprint

arXiv:2008.05756.

Guti

errez, J., Rodr

ıguez, V., and Martin, S. (2021). Com-

prehensive review of vision-based fall detection sys-

tems. Sensors, 21(3):947.

Habib, M. A., Mohktar, M. S., Kamaruzzaman, S. B., Lim,

K. S., Pin, T. M., and Ibrahim, F. (2014). Smartphone-

based solutions for fall detection and prevention: chal-

lenges and open issues. Sensors, 14(4):7181–7208.

Ismaili-Alaoui, A., Ba

ına, K., and Benali, K. (2022). Iodep:

Towards an iot-data analysis and event processing ar-

chitecture for business process incident management.

International journal of advanced computer science

and applications (IJACSA), 13(4).

Ismaili-Alaoui, A., Kasmi, O., Ba

ına, A., Ba

ına, K., Benali,

K., and Bellafkih, M. (2019). Priority-based event

management using fuzzy logic for an iot-bpm archi-

tecture. In 2019 IEEE 12th Conference on Service-

Oriented Computing and Applications (SOCA), pages

111–118. IEEE.

Jarraya, S. K. (2018). Computer vision-based fall detection

methods using the kinect camera: A survey. Inter-

national Journal of Computer Science & Information

Technology (IJCSIT) Vol, 10.

Karantonis, D. M., Narayanan, M. R., Mathie, M., Lovell,

N. H., and Celler, B. G. (2006). Implementa-

tion of a real-time human movement classiﬁer us-

ing a triaxial accelerometer for ambulatory monitor-

ing. IEEE transactions on information technology in

biomedicine, 10(1):156–167.

Mubashir, M., Shao, L., and Seed, L. (2013). A survey on

fall detection: Principles and approaches. Neurocom-

puting, 100:144–152.

Noury, N., Fleury, A., Rumeau, P., Bourke, A. K., Laighin,

G., Rialle, V., and Lundy, J.-E. (2007). Fall detection-

principles and methods. In 2007 29th Annual In-

ternational Conference of the IEEE Engineering in

Medicine and Biology Society, pages 1663–1666.

IEEE.

Ozdemir, A. T. and Barshan, B. (2014). Detecting falls with

wearable sensors using machine learning techniques.

Sensors, 14(6):10691–10708.

Rastogi, S. and Singh, J. (2021). A systematic review on

machine learning for fall detection system. Computa-

tional Intelligence, 37(2):951–974.

Vignat, J. (2001). Aspects psychiques de la chute. Neurol

Psychiatr Geriatr, 1:12–6.

Vishwakarma, V., Mandal, C., and Sural, S. (2007). Au-

tomatic detection of human fall in video. In Interna-

tional conference on pattern recognition and machine

intelligence, pages 616–623. Springer.

Wild, D., Nayak, U., and Isaacs, B. (1981). How dangerous

are falls in old people at home? Br Med J (Clin Res

Ed), 282(6260):266–268.

World Health Organization. Falls. https://www.who.int/en/

news-room/fact-sheets/detail/falls.

Xu, T., Zhou, Y., and Zhu, J. (2018). New advances and

challenges of fall detection systems: A survey. Ap-

plied Sciences, 8(3):418.

Yu, X. (2008). Approaches and principles of fall detection

for elderly and patient. In HealthCom 2008-10th In-

ternational Conference on e-health Networking, Ap-

plications and Services, pages 42–47. IEEE.

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

582