FALL DETECTOR BASED ON NEURAL NETWORKS

Rubén Blasco, Roberto Casas, Álvaro Marco, Victorián Coarasa, Yolanda Garrido and Jorge L. Falcó

Instituto de Investigación en Ingeniería de Aragón, Universidad de Zaragoza, María de Luna 1, Zaragoza, Spain

Keywords: Fall detector, neural networks, ZigBee, wearable sensors, pattern recognition.

Abstract: Falls are one of the biggest concerns of elderly people. This paper addresses a fall detection system which

uses an accelerometer to collect body accelerations, ZigBee to send relevant data when a fall might have

happened and a neural network to recognize fall patterns. This method presents improved performance

compared to traditional basic-threshold systems. Main advantage is that fall detection ratio is higher on

neural network based systems. Another important issue is the high immunity to events not being falls, but

with similar patterns (e.g. sitting in a sofa abruptly), usually confused with real falls. Minimization of these

occurrences has big influence on the confidence the user has on the system.

1 INTRODUCTION

Aging of population is a well-known problem in

developed countries. Nowadays, elderly people (+65

years old) represents in Spain more than 16% of the

population (Eurostat, 2007). Falls are one of the

major fears of the elderly and their relatives. Indeed,

some authors estimate the amount of falls of people

aged over 75 to be at least 30 percent per year

(Sixsmith and Johnson, 2004). In the end, people’s

concern about falls and whether there will be

someone there to help them in case of an emergency,

prevent them to age at home (Rodriguez et al, 2005).

As a result, people have to move to residences,

usually causing negative effect in their health and

happiness and resulting in high costs to the

individual, their family or the Social Welfare

System.

Fortunately, many initiatives are going on in

order to increase the time people can stay at home.

We will further see many fall detection systems

enabling people to receive quickly assistance even

when they are not able to request the assistance by

themselves (e.g. immobilized or unconscious). Also,

combination of these systems with telemedicine

allows closer monitoring or collaboration of various

experts in the diagnoses (Tunstall web, 2007).

Various methods have been described in order to

detect falls in the elderly. Those based in a sensing

infrastructure - infrared cameras (Alwan, et al.,

2006), vision systems (Williams et al., 1998) or

smart floors (Williams et al., 1998) - can be hardly

used in many cases. We find wearable systems more

appropriated in real scenarios because people refuse

to have cameras everywhere in their homes and

systems are much more expensive.

Inertial elements are mostly used for mobile

monitoring, but still the perfect detector does not

exist. Main reason is the difficulty in modelling a

fall, it can happen in many different ways; it will not

always be the typical big impact followed by

inactivity and horizontality. Williams et al. use a

shock sensor and a tilt switch to measure the

inclination after the impact (Williams et al., 1998).

Doughty et al. also use two sensors to perform the

same two-stage-analysis (Doughty et al, 2000),

which moreover is concreted in a commercial gadget

from Tunstall (Nait-Charif and McKenna, 2004).

Noury refines the procedure using an accelerometer

to detect the shock, also a tilt switch, and adding a

vibration sensor to estimate the physiological

activity (Noury, 2002). Of course, the more

variables measured, the more accurate the detection

can be, but also the more complicated and expensive

the hardware will be. Many actual works propose

just using accelerometers to carry out the full

detection (Noury, 2002; Degen et al, 2005; Chen et

al., 2005). Main reasons are their low power

consumption, reduced cost and versatility detecting

different events -shocks, inclination and activity-.

The devices presented in these works perform

satisfactory fall detection: more of 80% of falls are

correctly detected (Noury, 2002).

540

Blasco R., Casas R., Marco Á., Coarasa V., Garrido Y. and L. Falcó J. (2008).

FALL DETECTOR BASED ON NEURAL NETWORKS.

In Proceedings of the First International Conference on Bio-inspired Systems and Signal Processing, pages 540-545

DOI: 10.5220/0001066205400545

 SciTePress

2 SYSTEM DESCRIPTION

Users must find fall detection systems trust-worthy

and efficient in order to use them. Systems which

detect all falls but generate many false alarms make

users unconfident about it. Moreover, if we consider

the difficulty of distinguishing between some kind of

falls and ordinary movements in elderly people’s

life, threshold systems (those that generate an alert

when acceleration rise above a fixed value) become

not be reliable enough (Noury, 2002).

Figure 1.a. shows tri-axial acceleration when a

person has a sideward fall. On the other hand, Figure

1.b. shows accelerations when a person sits down on

a sofa abruptly. Both figures were obtained with a

device which measures triaxial accelerations, hanged

around the neck.

Figure 1a: Acceleration in three axes in a sideward fall.

Figure 1b: Acceleration in three axes while sitting abruptly

in a sofa.

As we can see, both figures have similar

acceleration peaks being also the shapes pretty

similar.

Our solution aims to distinguish falls from

movements that have similar acceleration patterns

not being falls using neural networks; that is to say,

separate occurrences into true and false falls.

2.1 Blocks Diagram

The fall detector consists of a mobile device with an

inertial sensor which is able to communicate through

a ZigBee network. The system also needs a

computer that analyzes data using a neural network.

Figure 2 shows the portable device blocks: battery,

sensor, microcontroller (µC), interface and Zigbee

transceiver. Reduced size and low power

consumption had been considered in the design

process of every block.

Vbat

Sensor

BATTERY

uC Transceiver

INTERFACE

Vbat

Figure 2: Blocks of the portable device.

The chosen sensor is MMA7260Q Freescale

accelerometer because of its wide input voltage

range (2,2 V - 3,6 V), current consumption (typically

less than 500 µA and 3 µA in sleep mode) and

reduced size (6x6x1,45mm). It also has three analog

outputs that give the acceleration value in axis X, Y

and Z. Its sensitivity is configured digitally into

some ranges (1,5 g; 2 g; 4 g or 6 g). As some falls

are above 4 g, our application uses the maximum

range (6 g) and minimum sensitivity (200 mV/g).

The chosen microcontroller is Microchip’s

PIC16F688. It has eight A/D channels that can be

configured to 10 bits. As well as working within a

wide voltage supply range (2 V - 5,5 V), it also has

very low current consumption (800 µA in active

mode and 1 nA while sleeping).

Regarding communications, we discarded the

development of a proprietary network for

interoperability reasons. Other standard wireless

protocols such as Bluetooth or WiFi consume too

much energy as they are intended for higher data

rates. We decided to use ZigBee because its

adequate data rate (250 kbps), security (128 bits

AES encryption), low latency (30 ms to join and 15

ms to access the network) and energy efficiency. Its

interoperability with other potential applications

(home control and automation), future projection of

FALL DETECTOR BASED ON NEURAL NETWORKS

541

the protocol, and its consequent cost reduction were

other strategic reasons behind our decision (Geer,

2005).

The chosen ZigBee chip is ETRX2 from

Telegesis. This is a ZigBee module on the 2,4 GHz

ISM band based upon the Ember’s EM250 chip. We

used the development environment proposed by

Ember to develop a ZigBee-compliant network

following mesh topology (ZigBee Alliance, 2007).

The chip consumes 30 mA when receiving or

transmitting data and 10 µA in sleep mode. As we

will use the radio exceptionally, just when are

reasonable indications about a fall (when a threshold

is exceeded), average power consumption due to

communication is reduced.

The user interface consists of a single button and

a buzzer for user interaction. Figure 3 shows the

mobile device prototype. Its size, including battery,

is 58x36x16 mm and it weights 30 gr.

Figure 3: Mobile device prototype.

In order to make the device useful is extremely

important to keep it on working long time using the

same set of batteries. That is why we gave

preference to power-conservative and size of

batteries among other designing requirements like

transmission rate or processing time. Precise battery

life estimation is very difficult because it will

depende on the number of false alarms generated;

every time the threshold is exceeded it sends data via

ZigBee. Anycase, with the battery used (3 V, 1000

mA·h), it can last for several months daily sending

several false falls to analize.

2.2 Software

As we said before, we designed a neural net to detect

falls also aiming to minimize the number of false-

falls compared to simple threshold based detectors.

Figure 4: System’s simplified flow diagram.

In our case, we use an acceleration threshold to

get the “acceleration pattern” of the possible fall to

be studied. Every 32 ms the device stores the current

acceleration measurements. It keeps a buffer with

the last 5 samples (t

≈ 160 ms). If the threshold is

exceeded, a possible fall might have happened. Then

we gather 25 samples more (t

≈ 800 ms) and all the

data (960 ms) is sent via ZigBee to the PC. As we

will see in section 3.2, those times and the threshold

have been empirically set through acceleration

pattern analysis of many falls and false-falls.

The “window time” (t

) represents the time

that the neural net analyzes the data in order to

relation the detected event to a true fall.

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

542

Figure 5: Window time.

In case the neural network detects a real fall, the

PC asks the mobile device to buzz for one minute.

During this time, the user can cancel the fall

situation pressing the button; the user is okay and

does not need assistance. In other case, an alert is

sent to an assistance center asking for help.

3 NEURAL NETWORK DESIGN

We have chosen MLP (MultiLayer Perceptron)

architecture because is the best neural network for

pattern classification (Del Hoyo Alonso, 2003).

MLPs are feedforward neural networks trained with

the standard backpropagation algorithm. They are

networks that learn how to transform input data into

a desired response. As they are supervised, they

require a set of known patterns with known

responses to get trained. With one or two hidden

layers, they can approximate virtually any input-

output map. They have been shown to approximate

the performance of optimal statistical classifiers in

difficult problems (Neurosolution web, 2007).

Every acceleration point, within window time, is

considered as input data to train neural network to

distinguish between true and false falls (figures 1a

and 1b). That is to say, if an event is represented by

30 samples for each axis (X, Y and Z), the number

of inputs will be 90 (30x3). Consequently, it is the

same as we give the net the whole graph to compare

and classify.

We have decided to train the net with one hidden

layer. To check if our choice is convenient or not,

we have designed a test bench with different

numbers of neurons, studying the absolute error in

each case. To accelerate the training, we have

chosen a bipolar sigmoid activation function for

neurons of the hidden layer. The activation function

of the output neuron is unipolar sigmoid so the

output looks like a binary signal (1 = TRUE FALL;

0 = FALSE FALL).

The suitable number of neurons of the hidden

layer is obtained doing simulations of different

neural nets. Finally, we choose the one which

produces the minimum absolute error. To reduce the

number of simulations and to get patterns from the

inputs able to generalize the results, we have defined

a requirement: the number of inputs is greater than

the number of neurons of the hidden layer.

3.1 Input Data Harvesting

Ten people of different ages, weight, height and sex

imitated the movements of elderly people to create a

data base of falls.

Table 1: Volunteers’ characteristics.

Age range 25-40 years

Weight range 44-105 kg

Height range 1.58-1.90 m

To get the data as close to reality as possible, the

volunteers had the acceleration detector hanged

around the neck. Volunteers were asked to simulate

true and false falls situations.

TRUE falls:

Every volunteer falls down 10 times on a

straw mat. The fall intensity changed (rough

and soft) and the way of falling down too

(side, front, backwards), hitting the ground

with their back, hip, knees, etc.

FALSE falls:

Every volunteer flings himself down 5 times

on the center and 5 times on the side of a sofa.

Every volunteer stumbles and hits a wall

without falling down 5 times.

Every volunteer walks around for 2 minutes

doing normal movements like sitting up and

down in chairs, picking up things, etc.

During the test, the fall detector continuously

samples the three acceleration axes each 32 ms

sending them to a PC working as a data logger. In

the end, we get a file with all the acceleration

samples in axis X, Y and Z for every volunteer. The

resulting data base consists of 99 samples of true

falls (we had one error while collecting data) and

150 of false falls.

FALL DETECTOR BASED ON NEURAL NETWORKS

543

3.2 Input Data Analysis

First of all, data analysis has determined the window

time length. After studying all the falls, we decided

that an event could be represented with 30 samples

(tv = 960 ms; t1 = 160 ms; t2 = 800 ms). This means

that the microcontroller has to store always in

memory the last five samples to send, in case the

acceleration threshold is exceeded, the event to the

PC to be analyzed.

With the window time selected, the number of

inputs to the neuronal network is set to 90. In order

to reduce the number of network entries -and

consequently the network size- we have done a PCA

(Principal Component Analysis). This method lies in

referencing input data to a new origin and coordinate

base.

In the new reference, the main components are

chosen to be those with the maximum variance

among samples (those with the highest covariance).

Therefore, if we take the samples representing

more than 95% of covariance, the number of input

will be reduced without losing significant

information. This leads to suppose that the greater is

the variance of an input, the more information it

gives.

The acceleration threshold was decided

experimentally. At first, guided by most of the

bibliography (Chen et al., 2005), we chose a 3 g

value. Then 97 out of 99 true falls and 121 out of

150 false falls surpassed the selected threshold.

Missing true falls is far worse than over-

detecting false falls, thus we reduced the threshold to

2 g to prevent losing any fall. As expected, we got

all the falls, but the number of false falls which

surpassed the threshold, increased to 241 because

even normal movements triggered the detection

process.

After using PCA analysis with the 340 events (99

falls plus 241 false-falls), the number of inputs was

reduced from 90 to 55, keeping the 95% of the

covariance of the original data.

3.3 Network Performance

The network was trained used Levenberg-Marquardt

algorithm (Neural-toolbox in Matlab).

We trained different MLP architectures 55xMx1

(being M the number of neurons in the hidden layer,

5≤M≤35). We repeated this process ten times in

order to ensure the network design and its

performance. Each test randomly selected 80% of

the events for training and 20% for validating. That

is to say, from the whole 340 events (99 falls plus

241 false-falls), the validation group had 20 true

falls and 48 events that could be confused with falls.

In the end, a neural net with 22 hidden neurons was

able to classify falls correctly.

When interpreting the neural net output give

precedence to the fall detection. Thus, we decided

that if the output is above or equal to 0.3, a fall is

detected. On the other hand, if the output is below

0.3, the analyzed event was not a true fall.

In table 2 we can see the network performance for

the ten tests.

Table 2: Validation group detection results.

Network fall

detection /

Fall events

Network fall

detection /

False-fall events

Test 1 20 / 20 0 / 48

Test 2 20 / 20 0 / 48

Test 3 20 / 20 1 / 48

Test 4 20 / 20 0 / 48

Test 5 18 / 20 1 / 48

Test 6 20 / 20 1 / 48

Test 7 16 / 20 1 / 48

Test 8 18 / 20 0 / 48

Test 9 17 / 20 1 / 48

Test 10 15 / 20 0 / 48

We can see how the network is able to detect

92% of all the falls and filter up 99% of the events

that can be confused with falls.

In figure 6, we show the network output we got

for the validating group in third test.

Figure 6: Network output for the third validation group.

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

544

4 CONCLUSIONS

The final results using MLP neural networks for fall

detection have been quite satisfactory. The

application classifies correctly 92% of the validation

group falls, better performance than other detection

methods: 80% in (Chen et al., 2005). Moreover, the

number of false alarms is drastically reduced to 1%,

which leads to enhance users trust on the fall

detector. Nevertheless, a more extensive study with

more users being also elderly has to be developed in

order to gather more data and confirm the results.

Although the portable device can run for months

with the same battery, the system needs a computer

to analyze all the data. In order to reduce costs, it is

possible to analyze the pattern remotely. As the

amount of exchanged data is reduced, it could be

sent via ADSL (if the person is at home), GPRS or

even SMS to a service center. Anyhow our

application gets better performance than others

embedded in a microcontroller but a higher cost and

complexity. To overcome this, we are currently

minimizing the neural network size so it can run in a

microcontroller or FPGA.

ACKNOWLEDGEMENTS

This work was supported by the Spanish MCYT

under project Ambiennet (TIN2006-15617-C03-02)

and by the EU under projects MonAmi (IST-5-

0535147) and EasyLine+ .

REFERENCES

Alwan, M., et al., 2006. ‘A Smart and Passive Floor-

Vibration Based Fall Detector for Elderly’. Proc. The

2nd IEEE International Conference on Information &

Communication Technologies: from Theory to

Applications ICTTA’06, Damascus, Syria, April 2006.

Chen, J., et al., 2005. ‘Wearable Sensors for Reliable Fall

Detection’. Proc. 27th Annual International

Conference of the Engineering in Medicine and

Biology Society, IEEE-EMBS 2005, September 2005,

pp. 3551–3554.

Degen, T., et al, 2005. ‘SPEEDY: a fall detector in a wrist

watch’. Proc 7th IEEE International Symposium on

Wearable Computers, October 2005, pp. 184–187.

Del Hoyo Alonso, R. 2003. ‘Supervised classification with

Associative SOM.’ IWANN 2003: 334-341

Doughty, K., et al, 2000. ‘The design of a practical and

reliable fall detector for community and institutional

telecare’, Journal of Telemedicine and Telecare,

February 2000, 6, (1), pp. 150–154.

Eurostat web: http://epp.eurostat.ec.europa.eu/ last visited

2007

Geer, D., 2005. ‘Users make a Beeline for ZigBee sensor

technology’. IEEE Computer Magazine, 38, pp. 16–

19.

Hansen, T.R., et al., 2005. ‘Using smart sensors and a

camera phone to detect and verify the fall of elderly

persons’. Proc. European Medical & Biological

Engineering Conference (EMBEC 2005), November

2005.

Nait-Charif, H., and McKenna, S. J., 2004. ‘Activity

Summarisation and Fall Detection in a Supportive

Home Environment’. Proc. of the 17th international

Conference on Pattern Recognition, (Icpr'04),

Washington DC, August 2004, 4, pp. 323–326.

Neurosolution web.

http://www.nd.com/definitions/mlp.htm last visited 2007.

Noury, N., 2002. ‘A Smart Sensor for the remote follow

up of activity and fall detection of the elderly’. Proc.

2nd Annual International Special Topic Conference on

Microtechnologies in Medicine & Biology, IEEE-

MMB2002, Madison-USA, May 2002, pp. 314–317.

Rodriguez J.M., et al, 2005. ‘Portable System for Patient

Monitoring With Wireless Technologies’. The

European Journal for the Information Professional,

October 2005, 5, (5), pp. 46–52.

Sixsmith, A., and Johnson, N., 2004. ‘A smart sensor to

detect the falls of the elderly’, IEEE Pervasive

Computing, April-June 2004, 3, (2), pp. 42–47.

Tunstall web. http://www.tunstall.co.uk last visited 2007.

Williams, G., et al., 1998. ‘A smart fall and activity

monitor for telecare applications’. Proc. of the 20th

IEEE Annual International Conference of the

Engineering in Medicine and Biology Society,

October 1998, 3, pp.1151–1154.

ZigBee Alliance: ‘ZigBee Standard’, 2007

FALL DETECTOR BASED ON NEURAL NETWORKS

545