A NEURAL NETWORK-BASED SENSOR FOR ELDER

FALLING DETECTION

Jiann-I Pan

, Cheng-Jie Yung

Department of Medical Informatics, Tzu-Chi University, Hua-Lien, Taiwan

Chung Chao Liang

M.D., Tzu-Chi Hospital, Hua-Lien, Taiwan

Keywords: Fall detection, accelerometer, neural network, back-propagation model.

Abstract: Falling down is going to be a crucial problem to an elder today. In many countries, unintentional injury was

being one of the leading causes of death in persons over age 65 years. As the society now, there are more

and more solitary elders of life alone and because of the isolation, it is necessary to design an intelligent and

sensitive falling detector for the elderly people. In this paper, we present an intelligent and portable fall

detection device based on artificial neural network technology. This fall detector consists of two main

components: accelerometer and microprocessor. The tri-axis accelerometer is used to continuously measure

the variation of elder’s 3 ways acceleration. The microprocessor reads the signals from the accelerometer

and performs the fall activity recognition through a back-propagation neural network model. This device is

integrated in a small box which can be holding on the belt for elder.

1 INTRODUCTION

Approximately 20% to 30% of elderly persons over

age 65 years fall each year, and this rate of falling

increases with age (Tinette, 1994). Therefore, the

falling has become a crucial problem of public

health and clinical problem among elderly persons.

As reported by (Doughty, 2000; Blake, 1992), falls

may lead to injury, fear of falling again, or traumatic

death. In many countries, unintentional injury was

being one of the leading causes of death in persons

over age 65 years (Haga, 1986; Tinette, 1994). In

general, a typical homecare emergency service

model for elder fall consists of four basic phases: (1)

discovering the fall event; (2) communicating the

call centre or hospital for requesting an emergency

service; (3) the call centre dispatch an ambulance to

the elder’s house; and (4) rescuing the elder. It is

obviously that all the activities of emergency service

are derived by discovering the elder falling down.

Therefore, as soon as possible to discover the fall

accident can effectively reduce the risk of death for

elder.

However, as the society now, there are more

and more elders choose solitary life or have only

oneself at home alone in the daytime. It makes the

discovery of fall accident more difficult. Therefore,

the objective of this paper is to develop an automatic

and intelligent fall detector. This fall detector

consists of two main components. First, the tri-axis

accelerometer is used to continuously measure the

variation of elder’s 3 ways acceleration. Second, the

microprocessor reads the signals from the

accelerometer and performs the fall activity

recognition through a back-propagation neural

network model. The incident that if fall, the

microprocessor unit will immediately alarm the

nearby persons to help quickly.

2 RELATED WORK

The simplest detection approach is that passive to

find the fall event. That is, when the elder falling

down, it needs the elder oneself to push a button on

the device to alarm the nearby person. The main

drawback of such devices is that need a human to

initiate the alarm. It will be useless when the elder is

203

Pan J., Yung C. and Chao Liang C. (2006).

A NEURAL NETWORK-BASED SENSOR FOR ELDER FALLING DETECTION.

In Proceedings of the Third International Conference on Informatics in Control, Automation and Robotics, pages 203-206

DOI: 10.5220/0001210002030206

 SciTePress

unconsciousness.

Another type of detection approaches is that

active to detect the fall event. The active detection

approaches can be further classified as portable

(personal worn device) and positioned

(environmental build-in):

1. Positioned fall detector: Equipments such as

video cameras or sensors are positioned on particular

places to detect the fall event. For examples,

Williams et al (Williams, et al, 1998) put the

vibration detector and posture sensor on the floor to

detect the fall event. Yamaguchi (Yamaguchi, 1998)

has place the infrared sensors on the door of

bathroom to detect the time spent in bathroom.

However, due to the sensors are fixed on the

environment, it will unworkable when the elder is

out of the detectable area.

2. Portable fall detector: The sensors integrated

with the microprocessor which having the fall

decision algorithms embedded into a small box or on

a cloth. For examples, Noury (Noury, 2002)

measures the vertical acceleration, posture, and

physiological vibrations by different sensors. After a

level triggering, a Boolean data is generated for the

position, fall, and vibration which is further

interpreted by an algorithm for the decision of the

fall. Degen et al (Degen and Jaeckel, 2003) designed

a fall detector in the form of a wrist watch. Their

approach is to integrate the norm of the three axes

acceleration vector and the two integrals. In these

approaches, the decision of fall event is based on the

predefined thresholds.

The main advantage of positioned detector is

that the device will not uncomfortable to the elder,

but there is restriction of activity area. The design

relative to positioned type, the portable device

makes the activity area not limit. According to the

considerations of privacy and portability, we focus

on developing a portable fall detector. Compared our

approach with the others portable detectors, the main

different is that the use of neural network classifier

to instead of a predefined falling thresholds.

3 MATERIAL

3.1 Principle of Fall Detection

The first step of our approach is to measure the three

axes acceleration vector via an accelerometer. As

shown in Figure 1, a free falling down will result the

accelerations of body’s three axes have changes in

various degree. For example, the y axis and z axis

will have more huge acceleration variation during a

forward fall.

Figure 1: During fall the accelerations of three directions

have changes in various degrees.

Artificial neural network technique was used to

perform the second step, i.e. the fall confidence

pattern recognition. There are two phases in neural

network: learning phase and recalling phase (see

Figure 2).

Figure 2: A neural network based approach to classifying

fall activities.

In the learning phase, a supervised learning

algorithm is adapted to train the neural network from

positive examples and negative examples. The

learning rule to adjust the weights and biases of the

network is based on the back-propagation training

algorithm. By the iterative training process, the

connective weights and bias in the neural network

can be refined to fit the generation of expect output.

Thus, in the recalling phase (also called predicate

phase) the input elder activities can be classified or

predicted by the trained neural network.

3.2 Architecture of the Intelligent

Fall Detector

The intelligent fall detector that proposed in this

paper comprised two key components: an

accelerometer module and a processing unit. Figure

3 shows the functional block diagram for the

intelligent fall detector.

ICINCO 2006 - INTELLIGENT CONTROL SYSTEMS AND OPTIMIZATION

204

Figure 3: The block diagram for intelligent fall detector.

The accelerometer module (implemented by

two bi-axes accelerometers, ADXL 202JE, Analog

Device Inc.) senses the elder activities continuously.

The output signal from accelerometer is modulated

by pulse width modulation (PWM). The processing

unit (implemented by MSP430, TI) handles three

activities: reads the incoming data from the

accelerometer module (the PWM signals that output

from accelerometers are calculated by a counter in

order to get the numerical value of acceleration);

performs neural network-based classification to

determine the fall activities; and produces an alarm

to indicate a fall event in happening.

3.3 The Structure of Neural

Network Classifier

An artificial neural network (ANN) based on

back-propagation algorithm is trained for classifying

falling activities. The structure of the

back-propagation network model comprises four

layers: one input layer, two hidden layer, and one

output layer. The input layer which consists of three

neurons reads data from the three axes accelerations.

The hidden layer consists of two layers which

arranged as 10 neurons and 7 neurons respectively.

The output layer represents the fall patterns, i.e. fall,

and not fall. This structure is gained after numerous

explorations of different hidden layers and neurons.

This network is a fully-connected architecture. The

structure of ANN is shown in Figure 4.

The activity output from this detector can be

classified into two possible patterns: fall and not fall.

If the “fall” pattern is recognized, it will trigger an

alarm to indicate the accident to nearby persons. The

“not fall” pattern means nothing of accident.

Figure 4: The structure of the back-propagation neural

network classifier.

4 RESULTS

A prototype of the intelligent fall detector has been

prepared, which implements the methodology

explained above. According to the safety

consideration, the experimentation is took place in

the laboratory and only young people chosen as the

subject in the experimentation. It involved six

subjects whose aged between 21 and 23.

The initial setup for back-propagation neural

network model is summarized in Table 1.

Table1: The setup for Back-propagation neural network.

Numbers of neurons in

each layer

Input / Output layers: 3 / 2

Hidden layers H / P: 10 / 7

Learning cycle 20000 times

Learning rate 0.5

Momentum 0.01

Activation function

Hyperbolic tangent function

−

)(

In the experimentation, each subject simulates

different possible activities which included 12

normal and 6 fall activities. Here, the basic normal

activities of daily living, where extended from (Wu,

2000), included walking, slow running, sitting down

and rising from a chair or sofa, climbing and

descending the stairs, picking up an object from the

floor, transferring in and out of a tub, and getting up

and lying down to a bed. The fall activities can be

simply classified according to the posture of elder’s

normal activities and ways of falling. We defined six

possible ways of falling activities:

z The body vertical (stand, walking, or running)

and falling forward,

z The body vertical and falling backward,

z The body vertical and falling right side or left

side,

z The body crooked (sit, squat, or stoop) and

A NEURAL NETWORK-BASED SENSOR FOR ELDER FALLING DETECTION

205

falling forward,

z The body crooked and falling backward, and

z The body crooked and falling right side or left

side.

These activities are simulated by the MatLab7

neural network tool in the learning phase. Through

20,000 times iterative learning, the back propagation

network is converged. In the recall phase, we

evaluate 90 activities which are the same activity

types as in the learning phase. In the evaluation, two

types of failure are adapted in our testing

(Yamaguchi, 1998). The first type of failure is that

the device indicating a fall event, but the real

situation is nothing. The next type of failure is the

fall accident is real in happened, but not detected by

the detector. The evaluation results are shown in

Table 2.

Table 2: The evaluation result for detecting normal and fall

activities.

Activitie

Num. of

evaluation

Success

detection

Error

detection

Correct

rate

Normal

activitie

No alarmed

Alarmed

95%

Fall

activitie

Alarmed

No Alarmed

93.3%

The similar experiment is repeated several

times. We have found that detection errors in testing

normal activities are almost distributed in “sit down”

activity. The errors in testing fall activities are

distributed in the “body crooked and falling right

side”. Next, the factors of subject’s agility, the

distance from detector to ground, and the numbers of

learning cycle in back-propagation neural network

are also infected the correct rate in the experiment.

5 CONCLUSIONS

An intelligent fall detector that based on artificial

neural network technology is introduced in this

paper. Comparing with other approaches, the correct

rate of recognition of fall activity by neural network

learning approach is better than the predefined

threshold value approach. In the future research, we

will explore the others learning approaches, such as

Hidden Markov Model classifier, to compare with

the neural network approach.

ACKNOWLEDGEMENTS

This project was funded by the National Science

Council, Taiwan (grant number NSC94-2218-E-

320- 003).

REFERENCES

Blake, A.J., 1992. Fall in the elderly. Br J Hosp Med, 47,

268-72.

Degen, T. and Jaeckel, H., 2003. SPEEDY: a fall detector

in a wrist watch. In Proceedings of 7th IEEE

International Symposium on Wearable Computers.

21-23 Oct.. pp.184 – 187

Doughty, K., 2000. Fall prevention and management

strategies based on intelligent detection, monitoring

and assessment. Presented at New Technologies in

Medicine for the Elderly, Charing Cross Hospital, 30

Nov.

Haga, H., Shibata, H., Mitsuzaki, T., and Hatano, S., 1986.

Falls in the institutionalized elderly in Japan. Arch

Gerentol. Geriatr. Vol.5, pp.1-9

Mathie, M.J., Basilakis, J., and Celler, B.G., 2001. A

system for monitoring posture and physical actitity

using accelerometers. In Proceedings of the 23rd

Annual EMBS International Conference, 25-28 Oct.

Turkey. Pp.3654-3657

Noury, N., 2002. A smart sensor for the remote follow up

of activity and fall detection of the elderly. 2

Annual

International IEEE-EMBS Special Topic Conference

on Microtechnologies in Medicine & Biology. May

Madison, Wisconsin. USA. pp.314-317

Tinetti, M.E., 1994. Prevention of falls and fall injuries in

elderly persons: a research agenda. Preventive

Medicine. Vol 23. pp.756-762

Williams, G., 1998. Doughty, K., Cameron,K. and

Bradley, D.A.. A smart fall and activity monitor for

telecare applications. In International Conference of

IEEE-EMBS, HongKong, pp.1151-1154

Wu, G., 2000. Distinguishing fall activities from normal

activities by velocity characteristics. Journal of

Biomechanics, Vol. 33. pp.1497-1500

Yamaguchi, A., 1998. Monitoring behavior in home using

positioning sensors. In International Conference of

IEEE-EMBS, HongKong, pp.1977-1979

ICINCO 2006 - INTELLIGENT CONTROL SYSTEMS AND OPTIMIZATION

206