Threshold-based Fall Detection on Smart Phones

Sebastian Fudickar, Alexander Lindemann and Bettina Schnor

Department of Computer Science, University of Potsdam, August-Bebel Str. 89, Potsdam, Germany

Keywords:

Fall detection, Development of Assistive Technology, Sensors-based Applications.

Abstract:

This paper evaluates threshold-based fall detection algorithms which use data from acceleration sensors that

are part of the current smart phone technology. The evaluation was done with sampled fall records where

young people simulate falls. To test the false positive rate of the algorithms, another record set with Activities

of the Daily Living (ADLs) from elderlies was used. The results are very promising and show that smart phone

sensors are suitable for fall detection. This will offer a new opportunity to assist elderlies in their daily living

and extend their period of self-determined living.

1 INTRODUCTION

With physical disorders gaining relevance with in-

creasing age to the elderlies’ daily living, the prob-

ability of falls increases and the probability of full

health-recoverage from falls decreases (especially if

falls are not instantaneously detected) (I. S. Joint

Commission Resources, 2008). Consequently, ﬁrst-

responders must be notiﬁed instantaneously if elder-

lies fall to increase healing chances.

Mobile devices like smart phones are equipped

with accelerometers which may be suited for auto-

matic fall detection. This is an interesting opportunity

for Assistive Living technologies where the automatic

fall detection can initiate an emergency call which

also includes the localization information where the

patient is fallen.

Prior studies have shown that the best results

are achieved when the mobile device is worn at the

hip (Kangas et al., 2008). The optimum should com-

bine a high sensitivity with a high speciﬁcity. The

sensitivity is a measure for the correctly detected falls:

Sensitivity =

TruePositives

Number o f all f alls

where TruePositives is the number of correctly

detected falls.

The speciﬁcity measures the rate of

TrueNegatives, i.e. the percentage of correctly

classiﬁed non-fall situations:

Speci ficity =

TrueNegatives

Number o f all non − f alls

A high speciﬁcity means a low number of false

alarms. This is also a very important feature to make

a fall detection system a suited assistive technology.

Modern smart phones are equipped with a tri-axial

accelerometer sensor which collects periodically a

vector with axis-speciﬁc acceleration. Typically, this

is used to re-orient the screen as a user moves the de-

vice. But this sensor has also potential to be used for

fall detection.

In context of the Assistive Living project

KopAL (Fudickar et al., 2011), we use a mobile de-

vice called Efﬁcient Mobile Unit (EMU) (Fudickar

et al., 2012a) which is equipped with the ADXL345

accelerometer. This is a high-end accelerometer

which is able to sample data with up to 800 Hz (for the

C bus) and which is capable of in-hardware prepro-

cessing. A simulation study has shown that the EMU

running the ADXL345 at 800 Hz has a very high sen-

sitivity of 93% (Fudickar et al., 2012b).

Recently, it was indicated by Mehner et

al. (Mehner et al., 2013) that the algorithm’s

detection rate is even accurate with sampling rates

below 800 Hz. The lower sampling rates are typical

for multi-purpose smart phones that are as well

equipped with acceleration sensors. Smart phone

operating systems such as Android aim to save

energy and typically sample with low rates (of up

to 100 Hz). The question was whether these lower

sampling rates will allow a high sensitivity.

Since the authors evaluate their algorithm with a

different fall set, it is not directly comparable to the

original one from (Fudickar et al., 2012b). Hence,

we extended the simulator to evaluate both algorithms

303

Fudickar S., Lindemann A. and Schnor B..

Threshold-based Fall Detection on Smart Phones.

DOI: 10.5220/0004795803030309

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2014), pages 303-309

ISBN: 978-989-758-010-9

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

(the one optimized in (Fudickar et al., 2012b) and the

one of Mehner et al. (Mehner et al., 2013)) with our

recorded falls and the simulator ﬁrst proposed in (Fu-

dickar et al., 2012b), to assure a meaningful compa-

rability.

Since the initial Activities of the Daily Living

(ADLs) were recorded by young probands that exe-

cuted critical ADLs in a high frequency and in an ex-

aggerated manner, the results may be less realistic for

elderlies. While recording falls of elderlies is still a

challenging and critical task, the recording of ADLs

of the elderly is less problematic. Therefore, ADLs

of elderlies were recorded in a nursing home and en-

abled us to test also the speciﬁcity of both algorithms

under realistic conditions for the intended user group.

The remainder of the article is structured as fol-

lows: In Section 2 related work is discussed. The

basic algorithm of the threshold-based fall detection

is introduced in Section 3. Section 4 presents the ex-

tension of the simulator. It follows the description of

the evaluation environment. The results of the evalu-

ation are presented in Section 6. The article ends with

a conclusion.

2 RELATED WORK

Jia (Jia, 2011) proposed a threshold-based fall detec-

tion algorithm for the accelerometer ADXL345 which

is able to pre-process raw acceleration data itself.

This feature can be used to let the processor remain in

low-power mode until a special event, in our use case

a free fall, was detected. This makes the accelerom-

eter beneﬁcial for the use in mobile devices since it

helps to save energy.

The Efﬁcient Mobile Unit (Fudickar et al., 2012a)

is a mobile device dedicated for Assisted Living sce-

narios which is equipped with the ADXL345. In (Fu-

dickar et al., 2012b), the parameter of Jia’s threshold-

based fall detection were optimized for the EMU.

For this purpose, the authors proposed a fall-detection

simulator which is able to model the threshold-based

fall detection. The simulator uses pre-recorded data

records from fall-situations and ADLs. ADLs were

generated for the simulator from recorded movements

of three probands with an age between 20 and 30

years.

The optimal parameter set for the EMU was iden-

tiﬁed by running several simulations to test the com-

plete parameter space. This resulted in an optimized

algorithm with a sensitivity of 93% compared to a

sensitivity of 33% of the original Jia algorithm.

The work of Mehner et al. (Mehner et al., 2013)

indicated that threshold based fall-detection algo-

Figure 1: States of a fall shown for a frontal fall without

loss of consciousness (Fudickar et al., 2012b).

rithms are as well applicable for smart phones and

that the lower sampling rates (such as 50 Hz) that are

supported by the smart phone’s operating system are

uncritical. Furthermore, they indicated that the exclu-

sion of the free-fall detection may increase the detec-

tion accuracy by 27 % from 56% with free-fall detec-

tion to 83% without free-fall detection. Overall the

proposed algorithm achieved a maximal sensitivity of

83% and a speciﬁcity of 100%.

Sannino et al. (Sannino et al., 2013) have recently

proposed a fall detection algorithm that is based on

supervised knowledge extraction for a windowing

technique and was optimized with the an subset of the

recorded fall set from (Fudickar et al., 2012b). The re-

sulting algorithm achieved a promising sensitivity of

91% and a speciﬁcity of 92 % as the average of 25

runs for a seperate subset (testing set) of the fall set.

3 THRESHOLD BASED FALL

DETECTION

For a tri-axial accelerometer, fall situations are char-

acterized by multiple sequential events, as shown in

Figure 1. The accelerometer collects a vector (x,y,z)

of the axis-speciﬁc acceleration. While the following

description from (Fudickar et al., 2012b) is applicable

to threshold based fall detection algorithms in gen-

eral, the parameter settings are described according

to Jia (Jia, 2011).

Within a fall situation, the falling body experi-

ences zero-gravity during free falls. The free fall is

the initial event of each fall situation and therefore

is identiﬁed ﬁrst. For example, for the accelerom-

eter ADXL345 used in our experiments, the gravity

must drop below the free fall threshold THRESH FF

(as described in Table 1) for a minimal duration of

HEALTHINF2014-InternationalConferenceonHealthInformatics

304

Table 1: The optimal register settings are identical for all evaluated sampling rates and fall-detection algorithms.

Parameter’s Description Optimized algorithm

THRESH FF free-fall threshold 0.75 g (0x0C)

STRIKE THRESHOLD minimal impact per axis 2 g (0x20)

STRIKE WINDOW maximal delay between free fall and impact 500 ms (0x19)

STABLE TIME minimal stable-phase duration 1 s (0x01)

STABLE WINDOW maximal stable-phase duration 3.5 s (0xAF)

NOMOVEMENT TIME duration of critical phase 5s (0x19)

NOMOVEMENT THRESHOLD maximal acceleration during stable-phase 0.4375 g (0x07)

TIME FF minimal free-fall duration 30 ms (0x06)

TIME FF. Other algorithms (such as (A.K.Bourke

et al., 2007)) exclude a minimal duration for free fall

detection but instead detect them only via threshold

value.

Once a free fall occurred, the algorithm tries

to detect an impact, which is given if the acceler-

ation values of all three axis exceed the so-called

STRIKE THRESHOLD. The maximal duration be-

tween free fall and impact recognition may not ex-

ceed the duration deﬁned in STRIKE WINDOW. If

no impact was detected within this period, no fall was

detected and the algorithm resets the accelerometer to

detect further free falls.

In a fall situation the impact is followed by

a stable phase. During the stable phase, all

axis’ acceleration values drop below the NOMOVE-

MENT THRESHOLD, for a minimal duration of

STABLE TIME. A fall is recognized, if a free fall,

an impact and a stable phase were detected sequen-

tially. Therefore, the detection of the stable phase is

essential for the fall detection, but does not indicate a

loss of consciousness.

Instead, loss of consciousness is detected during

the critical phase, which follows the stable phase.

If no movements were recognized during the critical

phase (we chose 5 seconds), a loss of consciousness

can be assumed, caused by the previous fall.

4 SIMULATOR EXTENSION

The fall detection simulator from (Fudickar et al.,

2012b) was extended by the following aspects:

Since we wanted to compare two different thresh-

old based fall detection algorithms (one with and one

without the free fall phase detection), we had to im-

plement the one which ignores the fall’s initial free

fall phase. This altered algorithm was implemented

according to the one presented in (Mehner et al.,

2013) and is used to evaluate the fall detection al-

gorithm for mobile devices with less complex ac-

celerometers.

The original threshold-based fall detection algo-

rithm classiﬁes the following three fall types:

1. Normal Falls: cover falls where the proband

moves again.

2. Critical Falls: describe falls where the proband

does not move after the impact and loss of con-

sciousness (l.o.c.) is assumed.

3. Critical Free Falls: are characterized by multiple

free fall events before the impact, which are e.g.

typical for falling down stairs. Critical free falls

do not differentiate regarding the l.o.c. since they

are in any case critical.

The fall detection algorithm without free fall de-

tection classiﬁes detected falls into normal or critical

falls and can not detect the third fall type. Instead, the

associated falls are detected as normal or critical falls.

The other modiﬁcation was the implementation of

an additional simulation stage which allows us to sim-

ulate different sampling rates (50 Hz, 100 Hz, 200 Hz,

400 Hz, 800 Hz).

5 EVALUATION ENVIRONMENT

The sensitivity and speciﬁcity of the threshold based

fall detection algorithms is evaluated with the fall sets

and ADLS that are described in the following sub-

sections. All data records were sampled with the

EMU (Fudickar et al., 2012a) since it is equipped with

the ADXL345 accelerometer which is capable to sam-

ple data with a high sampling rate of 800 Hz.

5.1 Fall Set

The fall set was taken from (Fudickar et al., 2012b).

Since, the risk of injuries in case of older probands

is obviously much too high, the fall set was recorded

by three young probands with an age between 20 and

30 years including both genders. The fall set covered

frontal falls, backward falls, falls to the left and to the

right, falling after standing up, falling from bending

Threshold-basedFallDetectiononSmartPhones

305

Table 2: Detected falls with and without free fall detection for the analyzed sampling rates. Falls are only counted at the most

critical fall-type. The sum is calculated from the detected falls and critical falls and critical free falls.

Sampling rate with free fall detection without free fall detection

normal falls critical fall critical free fall sum falls normal falls critical fall sum falls

800 Hz 29 43 6 78 (92%) 35 48 83 (99%)

400 Hz 32 41 6 79 (94%) 37 46 83 (99%)

200 Hz 29 45 4 78 (92%) 34 48 82 (98%)

100 Hz 28 44 7 79 (94%) 34 48 82 (98%)

50 Hz 28 45 4 77 (92%) 34 49 83 (99%)

Expected falls 36 36 12 84 42 42 84

down while picking up a book and falling from stairs

as suggested by (Wang et al., 2008). The half of the

allover 84 recorded falls were critical falls, with loss

of consciousness (where the proband did not move af-

ter the falls). The other half were falls after which the

proband moved again.

5.2 ADLs of Elderlies

To evaluate the speciﬁcity of the algorithms, we

recorded the acceleration characteristics of ADLs of

elderlies. Thereby, elderlies were equipped with a

fanny pack, worn at the probands’ hip in which the

EMU was carried.

The recordings took place in the nursing home

Florencehort, Stahnsdorf, Germany on two consecu-

tive days in August 2013. Nine elderlies (four males

and ﬁve females) with an average age of 82 years

(ranging from 70 to 95 years) participated in the

recordings. Among the participants, one was using

a wheelchair and four were using a wheeled walking

frame.

Only recordings of eight probands are used for

the evaluation of the algorithm’s speciﬁcity, since the

recording failed in one case due to a device error.

Allover, more than 41 hours of ADLs were recorded.

The phases of attachment and detachment of the fanny

pack were excluded (cropped) from the recordings.

This results in a considered duration of allover 39

hours in which 37695360 acceleration values were

recorded. The considered recording duration per de-

vice ranged from 3.37 to 5.93 hours and was in av-

erage about 5 hours. The recordings started around

10:15 a clock and therefore covered typical daily ac-

tivities including eating lunch, walks and potential af-

ternoon naps.

Figures 2 shows the recorded acceleration values

(x,y,z) plotted as the length of the vector:

+ y

+ z

The maximal recorded acceleration in the cropped

recordings was at 8.7 g. However, such measurements

over 5 g occurred rarely and its occurence is marked

in the Figures by a red cross at the upper bound. Fur-

ther, some missing samples exist in the recordings

shown in Figure 2 c) and h) due to recording problems

of the devices (which are still in prototype stage). The

acceleration of the recording shown in Figure 2 d) var-

ied less than the one of the others and thereby was

probably sampled with the proband in the wheel chair.

6 RESULTS

With the extended simulator and the additional

recorded ADLs of elderlies, we evaluated both thresh-

old based fall detection algorithms (with and without

free-fall detection) and the lower sampling rates to get

insights on their inﬂuence on the sensitivity and speci-

ﬁcity.

The evaluation results are discussed separately for

the sensitivity and speciﬁcity in the following.

6.1 Sensitivity

Ahead of the evaluation, the threshold parameter’s

settings were optimized for the fall set for all sam-

pling rates and both algorithms (with and without free

fall detection), in respect to potential variations of

the optimum. However, the optimal parameter set-

tings as expected did not vary or change since they

rather represent the fall characteristics, which did not

change. Therefore, the optimal parameter settings are

as shown in Table 1 identical to the ones that were

identiﬁed as optimal in (Fudickar et al., 2012b). Con-

sequently, the evaluated algorithms differed only by

the inclusion of the free-fall detection step and the

used sampling rate.

For the evaluated sampling rates (between 50 Hz

and 800 Hz), the fall detection algorithm’s overall

sensitivity is minor affected by the sampling rate, as

shown in Table 2. The associated confusion matrixes

are shown in Table 3 and in Table 4. For the applied

fall records, a sampling rate of 400 and 50 Hz resulted

in a slightly higher fall-detection rate. In contrast to

the sampling rate, the exclusion of the free-fall detec-

tion signiﬁcantly increased the algorithm’s sensitivity

HEALTHINF2014-InternationalConferenceonHealthInformatics

306

(a) (b)

(e) (f)

(g) (h)

Figure 2: The acceleration in g (black) of the ADL records of elderlies (red crosses indicate the time of acceleration measure

that exceeds 5 g, which occured in a) 17, in c) 8, in e) 12, in f) 20, in g) 7 and in h) 15 times).

Threshold-basedFallDetectiononSmartPhones

307

Table 3: Confusion matrix for fall detection algorithm with

freefall detection at 50 Hz.

Detected as Falls Detected as ADL

Falls 77 7

ADLs 0 ca. 2355960

and increased the detection rate by up to 5% to 83

falls (99%) without free-fall detection instead of 79

falls (94%) with free-fall detection.

6.2 Speciﬁcity

The speciﬁcity of both algorithms was evaluated with

the records of the elderlies’ ADLs. Both algorithms

have optimal speciﬁcity, since no false positive fall

was detected. Hence in our experiments, the free-fall

detection had no inﬂuence on the speciﬁcity.

However, the speciﬁcity of the fall-type detection

differs and is without free fall detection less accurate

(see Table 2). While most falls are correctly detected,

the type of a fall is regularly false detected among

all algorithm conﬁgurations (see Table 2). Typically

normal falls are detected as critical falls. Since in

any fall-situation (even the ones without loss of con-

sciousness) an emergency should be notiﬁed, this as-

pect is unproblematic, but might be further investi-

gated.

6.3 Interpretation of the Results

The results indicate that the exclusion of the free-fall

detection step enhances the sensitivity, while not af-

fecting the algorithms speciﬁcity. Thereby, we could

conﬁrm the results of (Mehner et al., 2013) who pro-

posed the exclusion of this processing step. Further-

more, we conﬁrmed that the data sampling rate has

minor impact on the algorithm’s sensitivity and speci-

ﬁcity in the tested range (50 – 800 Hz). The algo-

rithm’s optimal parameter settings have shown to be

independent from the data sampling rates at all.

7 CONCLUSIONS

We evaluated threshold-based fall detection algo-

rithms which can be used on smart phones. Our

motivation was that multi-purpose smart phones are

equipped with acceleration sensors. Smart phone op-

erating systems such as Android aim to save energy

and typically sample with low rates (typically be-

tween 20 and 100 Hz). The question was whether

these lower sampling rates will allow a high sensitiv-

ity.

Table 4: Confusion matrix for fall detection algorithm with-

out freefall detection at 50 Hz.

Detected as Falls Detected as ADL

Falls 83 1

ADLs 0 ca. 2355960

Therefore, we extended the simulator presented in

(Fudickar et al., 2012b) to compare different sampling

rates and two different fall detection variants: one

with and one without detection of a free fall phase.

Our results show that

1. Fall detection with low sampling rates of at least

50 Hz can be used and have a sensitivity of 99%

for our fall records.

2. The sensitivity of the fall detection algorithm vari-

ant without the detection of a free fall phase is

(with up to 5 %) slightly better. From 84 falls,

the algorithm with free fall phase detection rec-

ognized 77 falls where the algorithm without free

fall detection recognized 83 falls.

3. The speciﬁcity of both algorithms regarding the

false-positive rates is perfect for the recorded

ADLs of the elderly.

The application programming interfaces for the ac-

cess of accelerometers of current smart phone OSs

such as Android and iOS are limited regarding the

conﬁguration of speciﬁc sampling rates and the ac-

cess of interrupts. Therefore, we are looking forward

to propose better extensions and will integrate the fall

detection into the Kernel of the Android OS.

Android based smart phones use in some opera-

tion modes even lower sampling rates (up to 20 Hz)

than the ones used in our simulation to save energy on

the mobile devices. Therefore, we plan to adapt the

simulator to work also for this lower sampling rates.

ACKNOWLEDGEMENTS

The nursing home Florencehort in Stahnsdorf which

belongs to the Landesauschuss f

ur innere Mission

(LAFIM) supports sincerely our studies. We like to

thank the LAFIM team and all of the probands for

their patient support of our work.

REFERENCES

A.K.Bourke, J.V.O’Brien, and G.M.Lyons (1 July 2007).

Evaluation of a threshold-based tri-axial accelerom-

eter fall detection algorithm. Gait & Posture, 26:194–

199.

HEALTHINF2014-InternationalConferenceonHealthInformatics

308

Fudickar, S., Frohberg, M., Taube, S., Mahr, P., and Schnor,

B. (2012a). An energy efﬁcient mobile device for

assisted living applications. In 2012 IEEE Online

Conference on Green Communications (IEEE Green-

Com’12).

Fudickar, S., Karth, C., Mahr, P., and Schnor, B. (2012b).

Fall-detection simulator for accelerometers with in-

hardware preprocessing. In Proceedings of the 5th

International Conference on PErvasive Technologies

Related to Assistive Environments, PETRA ’12, pages

41:1–41:7, New York, NY, USA. ACM.

Fudickar, S., Schnor, B., Felber, J., Neyer, F. J., Lenz, M.,

and Stede, M. (2011). KopAL - An Orientation Sys-

tem For Patients With Dementia. In Bj

orn Gottfried

and Hamid Aghajan, editor, Behaviour Monitoring

and Interpretation - BMI, pages 83–104. IOS Press,

Amsterdam, Netherlands.

I. S. Joint Commission Resources (2008). Reducing the risk

of patient harm resulting from falls. Joint Commission

Resources.

Jia, A. D. N. (2011). AN-1023 - Fall Detec-

tion Application by Using 3-Axis Accelerom-

eter ADXL345. Analog Devices. http:

//www.analog.com/static/imported-files/

application_notes/AN-1023.pdf; accessed June

1, 2013.

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

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

detection algorithms for body attached accelerome-

ters. Gait & posture, 28:285–91.

Mehner, S., Klauck, R., and Koenig, H. (2013). Location-

independent fall detection with smartphone. In Pro-

ceedings of the 6th International Conference on PEr-

vasive Technologies Related to Assistive Environ-

ments, PETRA ’13, New York, NY, USA. ACM.

Sannino, G., Falco, I. D., and Pietro, G. D. (2013). Effec-

tive supervised knowledge extraction for an mhealth

system for fall detection. In Medicon.

Wang, C., Chiang, C., Huang, C., and Chan, C. (2008). De-

velopment of a fall detecting system for the elderly

residents. The Second International Conference of

Bioinformatics and Biomedical Engineering, 0:1359–

1362.

Threshold-basedFallDetectiononSmartPhones

309