FALL DETECTION SYSTEM FOR ELDERLY PEOPLE

A Neural Network Approach

Getúlio Igrejas

, Joana S. Amaral

1,2

and Pedro J. Rodrigues

Polytechnic Institute of Bragança, Campus de Santa Apolónia, Apartado 1134, 5301-857 Bragança, Portugal

REQUIMTE, Pharmacy Faculty, University of Porto, Rua Aníbal Cunha, Porto, Portugal

Keywords: Fall Detection, Neural Network, Inertial Sensors, Features Selection, Embedded Devices.

Abstract: In this work a new approach for a fall detection system is proposed. The device integrates a 3-axis

accelerometer and a 3-axis gyroscope to measure linear acceleration and angular velocities, respectively.

Information from both sensors is used to characterize movements through selected features extracted from

raw data. A classification system based on a Feedforward Backpropagation Neural Network is then trained,

based on the extracted features. The performed tests present low false positives and low false negatives rates

with good specificity and sensitivity values.

1 INTRODUCTION

Elderly people are generally affected by different

problems such as the diminution of muscle strength,

decreased balance, vision difficulties and

neurodegenerative diseases, among others. Due to

these problems, aged people frequently have less

mobility and autonomy, as well as increased

difficulties to perform normal daily activities,

making them a particular group prone to suffer fall

events. In average, each year, one in every three

adults over 65 years older experiences a fall event

and this ratio increases to one in every two adults

aging more than 80 years (Hausdorff et al., 2001;

Hornbrook et al., 1994). Falls in the elderly can

cause physical damage with moderate to severe

injuries, such as soft tissue wounds, hip fractures

and head traumas, which remarkably deteriorate the

health status of elderly people and can even increase

the risk of early death (CDC, 2011).

After a fall event, the individual can become

unconsciousness or immobilized and unable to raise

or ask for help. In both cases, the individual will

remain without medical assistance. In fact, it has

been reported that half of the elderly population that

suffers a long-lie (involuntarily remaining on the

floor for an hour or more) after a fall, dies within six

months, even if no sever injuries result from the fall

(Noury et al., 2008). Thus, it is very important to

provide for assistance as soon as possible when a fall

event takes place. Having in mind the described

context, the development of intelligent systems able

to detect a fall event and send an alert would

contribute for a higher quality of life and

independent living of the elderly people. The main

objective of this work is to introduce a fall detection

system based on a machine learning paradigm. The

proposed scheme presented good specificity and

sensibility rates, which makes it possible to be used

in a real scenario situation. Additionally, the used

strategy introduces a learning and adaption ability

not found in other proposed schemes for the same

objective.

2 STATE-OF-THE-ART

To the present date, different solutions have been

proposed for fall detection of elderly individuals.

The simplest approach consists in using an alarm

button that, in case of an emergency, should be

pressed to send an alert to a relative or a care given

institution. Since an automatic fall detection system

would resolve this problem, recently several

different technologies and approaches have been

proposed with this aim.

According to Noury et al. (2008) a fall event can

be described as a series of four stages: the pre-fall

phase, the critical phase, the post-fall and the

recovery phase. During the pre-fall stage, the

individual is performing its normal ADL, including

some occasional sudden movements, like sitting or

355

Igrejas G., S. Amaral J. and J. Rodrigues P..

FALL DETECTION SYSTEM FOR ELDERLY PEOPLE - A Neural Network Approach.

DOI: 10.5220/0003792003550358

In Proceedings of the International Conference on Biomedical Electronics and Devices (BIODEVICES-2012), pages 355-358

ISBN: 978-989-8425-91-1

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

lying down quickly, which must be distinguished

from a fall. The critical phase is a very brief phase

(about 200ms) consisting in rapid movements of the

body toward the ground, finalized by a strong shock

on the floor. In the post-fall stage the person remains

inactive, most of the times in the floor, and this stage

could last several minutes or hours. In the recovery

phase the person gets up by his own means or with

the help from another person (Noury et al., 2008).

When a fall happens, there is a brief period of “free

fall” during which the vertical speed increases

linearly with time due to gravitational acceleration

(Noury et al., 2008). In this period, an early

detection of the critical phase can be achieved. For

this purpose, different approaches have been

suggested being the most relevant ones the use of

video-based monitoring systems coupled to

computer vision techniques, the use of

accelerometers and gyroscopes (Noury et al., 2008).

Fall detection can also be achieved at the end of the

critical phase by detecting the impact shock that

happens when the body hits the floor or an obstacle

(table, bathtub, stair steps, etc.). This can be

achieved by using an accelerometer or a shock

detector (e.g. piezoelectric sensor).

Most automatic fall detection systems reported

on literature consist in wearable sensor-based

devices targeting the critical phase of a fall (either its

early detection or the detection of its end). Among

the different sensor technologies used, inertial

sensors, mainly accelerometers and gyroscopes, are

the most referred on the literature (Benocci et al.,

2010; Bourke et al., 2007a; Laguna et al., 2010) with

other types, such as barometric pressure sensors

(Bianchi et al., 2010), also being described.

Fall detection systems based on inertial sensors

are probably the most effective ones, since they can

better describe movement dynamics.

Although it is possible to identify a fall just

recurring to one type of inertial sensors (using only

accelerometers or only gyroscopes) some authors

refer that, to improve the effectiveness of the system

(less false positives and false negatives), a

combination of inertial sensors is advisable (Nyan et

al., 2008).

3 PROPOSED STRATEGY

Although with different implementations, most

works generally use one or several threshold values

to identify a fall. This methodology presents some

drawbacks. An important one is that the threshold

values are highly dependent on the test measurement

group.

In this work a fall detection device based on an

Artificial Intelligence paradigm is proposed. Instead

of using threshold values, a neural network was

trained with fall and non-fall data obtained from a 3-

axis accelerometer and a 3-axis gyroscope. The main

advantage of this implementation is that the system

has a learning capability that allows inferring

correlations between the input data and the specific

movement (fall or non-fall).

3.1 System Architecture

The proposed device is composed by a 32 bit

microcontroller platform and two sensors, namely a

3-axis accelerometer and a 3-axis gyroscope. It is

based on the 32 bit ARM Cortex-M3 processor

(MBED) running at 96MHz, 512KB of flash

memory, 64KB of RAM.

The accelerometer used was the ADXL345. It is

a low power 3-axis digital accelerometer with 13-bit

resolution able to measure up to ±16 g.

The gyroscope used was the ITG3200. It is a low

power digital 3-axis gyroscope with a 16-bit

resolution able to measure angular velocities up to

±2000º/s with a sensitivity of 14.375 LSBs per º/sec.

Along with the described sensors the device also

includes an emergency button, a buzzer and a wifi

module. The wifi module is used to send the

emergency alert to a server. This server is

responsible to send a notification/message to a

phone number or email account previously defined.

Although wearable devices can be used attached

on different body locations, such as the wrist, the

neck, the waist, etc., the proposed system is intended

to be used at the waist level because when used on

this location, the system is positioned near the

body’s center of gravity, thus providing reliable

information on the subject body movements (Kangas

et al., 2008). The system is continuously obtaining

samples from the accelerometer and gyroscope in 4

seconds windows, corresponding to a sampling rate

of 200 Hz. In this period about 800 samples of the

acceleration and angular velocity are collected. From

the obtained samples ten features are extracted (the

features used are described in the next section) and

feed to the implemented neural network, which was

previously trained to discriminate fall situations

from non-fall situations.

3.2 Feature Extraction

To reduce the input data dimension we choose to

BIODEVICES 2012 - International Conference on Biomedical Electronics and Devices

356

compute, from the raw values of linear acceleration

and angular velocity, features that can be used to

describe the movement dynamics and, consequently,

be used to discriminate a fall from a non-fall

situation. There are ten features extracted from the

linear acceleration and angular velocity sample data,

namely:

()

iii

ωω

=++

∑

(1)

where N is the number of samples,

is the

acceleration for axes, x, y and z, and

is the sample

index.

()

iii

θθ

=++

∑

(2)

where

is the angular velocity for each axis.

()

(

)

()

Var Var Var

ωω

(3)

where Var() represents the variance.

()

(

)

()

Var Var Var

θθ

(4)

(

)

(

)

(

)

(

)

xx yy zz

ii i

F Max Max Max Max

ωω ωω ωω

=−−−

(5)

where Max() is the maximum operator and

is the

mean value, and

is the sample index.

()

(

) ()

()

xx yy zz

ii i

F Max Max Max Max

θθ θθ θθ

=−−−

(6)

()

iii

FMax

ωω

(7)

()

iii

FMax

θθ

(8)

The last two features represent a dynamic range

operator. They can be expressed as follows:

{

}

500S

ωω

(9)

where

is the acceleration of any axis.

# S

(10)

{

}

500S

θθ

(11)

where

is the angular velocity of any axis.

# S

(12)

3.3 Neural Classifier

The Feedforward Backpropagation (Rumelhart et

al., 1986) is an artificial neural system that has been

widely used in pattern recognition applications. In

most of the supervised applications, this neural

network, based in the perceptron neuron, presents a

good generalization behavior and shows good

computational performance. Nevertheless, there are

several issues that make this neural system not

appropriated for all applications. The

backpropagation algorithm used for training

feedforward networks is based on the gradient

descent. This method allows us to automatically set

the weights of the neurons in a way to minimize the

error between the target pattern set and the output

pattern set. After ending the training stage, the

neural network must map the input vector into an

output vector with a minor error. In this application,

the input vector contains the values from the 10

chosen features and the target/output vector will

represent the fall/non-fall classes in a binary way,

establishing one logical bit, “0” to the non-fall

situation and “1” to the fall situation. The number of

examples that were produced to establish the

movement patterns set was 631. The first half of this

set is formed by patterns concerning the fall class

and the other half concerns the non-fall class.

Several patterns were acquired asking to ten

volunteers to perform different motion situations.

The neural network was trained using 70% of the

movement/pose situations that were acquired

concerning several movement scenarios. The

remaining situations (30%) were used to test the

neural classifier. The training and the test sets were

different and randomly obtained from the total set of

fall and non-fall patterns. Several runs were

performed using different structures for the neural

network model. The experiments were prepared

using a 3-fold (cross-validation) x 20 runs. The

correct average classification rate over the validation

patterns was 96%. The trial performed using the test

set showed a generalization performance which

indicates that the system could be used in a real

scenario.

4 RESULTS

The device was tested in several contexts that

included normal ADL and fall simulated situations.

Tests were performed for the same kind of cases

used for pattern acquisition, presented on Table 2,

by four different individuals (different heights and

FALL DETECTION SYSTEM FOR ELDERLY PEOPLE - A Neural Network Approach

357

weights). The device was used on the waist. The

four volunteers were invited to act as natural as

possible in all situations. Fall simulations were

performed with the help of a protection mattress so

that the volunteers could reproduce a fall without

any constrain of injury.

Each individual performed 3 tests for each

situation. The obtained results evidence a very good

overall performance of the device. For a total of 180

tests only 9 tests were misclassified, which

represents a 95% of correctly classified cases. More

important, no false negatives were found, meaning

that the device is able to detect all the fall events

tested. The failed cases can be explained by the

acceleration and angular velocities profiles. The

sitting down, jumping and to go on all fours

movements could have similar profiles to a fall.

Following the proposal of Noury et al. (2008) for

the classification and evaluation of fall detection

systems, Sensitivity and Specificity criteria were

used to assess the performance of the proposed

scheme.

Sensitivity

TP FN

(13)

Specificity

TN FP

(14)

where TP, FN, TN and FP are the True Positive,

False Negative, True Negative and False Positive

cases, respectively.

The obtained Sensitivity and Specificity

coefficients were 100% and 91.67%, respectively.

5 CONCLUSIONS

In this work a fall detection device based on a neural

network was proposed. This approach revealed to be

very effective in identifying falls, presenting a

Sensitivity coefficient of 100%. It was not so

efficient classifying non-fall events, presenting some

false positives cases.

Considering that not all the tests for the same

type of motion were wrongly classified, this can

indicate an underfitting training. Increasing the

number of training examples could help to improve

the classifier performance in this point.

REFERENCES

Benocci, M., Tacconi, C., Farella, E., Benini, L., Chiari,

L., Vanzago, L., 2010. Accelerometer-based fall detection

using optimized ZigBee data streaming.

Microelectronics Journal, 41(11), 703-710

Bianchi, F., Redmond, S. J., Narayanan, M. R., Cerutti, S.,

Lovell, N.H., 2010. Barometric Pressure and Triaxial

Accelerometry-Based Falls Event Detection. Ieee

Transactions on Neural Systems and Rehabilitation

Engineering, 18(6), 619-627.

Bourke, A. K., O'Brien, J. V., Lyons, G. M., 2007a.

Evaluation of a threshold-based tri-axial accelerometer

fall detection algorithm. Gait & Posture, 26(2),

194-199.

Bourke, A. K., van de Ven, P., Gamble, M., O'Connor, R.,

Murphy, K., Bogan, E., McQuade, E., Finucane, P.,

ÓLaighin, G. Nelson, J., 2010b. Assessment of Waist-

worn Tri-axial Accelerometer Based Fall-detection

Algorithms using Continuous Unsupervised Activities.

2010 Annual International Conference of the Ieee

Engineering in Medicine and Biology Society (Embc),

2782-2785.

CDC. (2011, December 8 2010). Falls among older

Adults: an overview. Retrieved 04-06-2011, from

http://www.cdc.gov/HomeandRecreationalSafety/Falls

/adultfalls.html

Hausdorff, J. M., Rios, D. A., Edelberg, H. K., 2001. Gait

variability and fall risk in community-living older

adults: A 1-year prospective study. Archives of

Physical Medicine and Rehabilitation, 82, 1050-1056.

Hornbrook, M. C., Stevens, V. J., Wingfield, D. J., Hollis,

J. F., Greenlick, M. R., Ory, M. G., 1994. Preventing

Falls among Community-Dwelling Older Persons -

Results from a Randomized Trial. Gerontologist,

34(1), 16-23.

Kangas, M., Konttila, A., Lindgren, P., Winblad, I., &

Jamsa, T., 2008. Comparison of low-complexity fall

detection, algorithms for body attached

accelerometers. Gait & Posture, 28(2), 285-291.

Laguna, M. A., Marques, J. M., Tirado, M. J., Finat, J.,

2010. Fall detection system: a solution based on low

cost sensors. Systems and Information Technologies,

89-94.

Noury, N., Rumeau, P., Bourke, A. K., Olaighin, G.,

Lundy, J. E., 2008. A proposal for the classification

and evaluation of fall detectors. Irbm, 29(6), 340-349.

Nyan, M. N., Tay, F. E. H., Murugasu, E., 2008. A

wearable system for pre-impact fall detection. Journal

of Biomechanics, 41(16), 3475-3481.

Rumelhart, D. E., Hinton, G. E., & Williams, R. J., 1986.

Learning Representations by Back-Propagating Errors.

Nature, 323(6088), 533-536.

BIODEVICES 2012 - International Conference on Biomedical Electronics and Devices

358