Conference of the IEEE EMBS
Cité Internationale, Lyon, France
August 23-26, 2007.
Determination of simple thresholds for accelerometry-based
parameters for fall detection
Maarit Kangas, Antti Konttila, Ilkka Winblad and Timo Jämsä
Abstract—The increasing population of elderly people is
mainly living in a home-dwelling environment and needs
applications to support their independency and safety. Falls are
one of the major health risks that affect the quality of life
among older adults. Body attached accelerometers have been
used to detect falls. The placement of the accelerometric sensor
as well as the fall detection algorithms are still under
investigation. The aim of the present pilot study was to
determine acceleration thresholds for fall detection, using
triaxial accelerometric measurements at the waist, wrist, and
head. Intentional falls (forward, backward, and lateral) and
activities of daily living (ADL) were performed by two
voluntary subjects. The results showed that measurements
from the waist and head have potential to distinguish between
falls and ADL. Especially, when the simple threshold-based
detection was combined with posture detection after the fall,
the sensitivity and specificity of fall detection were up to 100 %.
On the contrary, the wrist did not appear to be an optimal site
for fall detection.
HE increasing population of elderly people (aged 65+)
is mainly living in a home-dwelling environment and
needs applications to support their independency and
safety. Falls are one of the major health risks that affect the
quality of life among elderly by producing fear and resulting
in decrease in mobility and activity -. It appears that
elderly people are willing to accept new technologies to
support their independence and safety , .
Body attached accelerometers - and gyroscopes
,  have been used to detect human movement and
especially falls. The placement of an acceleration sensor to
optimize the location of a fall detector has been studied in
some extent. The placement site at the waist has been
suggested to be the most efficient, since at this site the
acceleration signal is similar and evenly distributed between
different fall types . Furthermore, waist attached
accelerometers are located near to the body center of gravity
providing reliable information on subject’s movements, with
the exception of movements of arms and legs .
The sum vector of a triaxial accelerometer signal has been
suggested to be more accurate in fall detection than single
axis thresholds . Start of the fall, before the actual
impact, has been monitored using the norm of the triaxial
signal ,  or velocity , . The actual impact has
been detected with threshold based algorithms , ,
The aim of the present pilot study was to determine
acceleration thresholds for fall detection, using triaxial
accelerometric measurements at the waist, wrist, and head.
II. MATERIALS AND METHODS
A. Subjects, Test Falls and ADL
Intentional falls and activities of daily living (ADL) were
performed by two voluntary subjects (aged 22 and 38 years).
Forward, backward, and lateral falls were performed
towards an air-filled bed or a combination of a tatami and
mattresses. Acceleration signal was measured either from
the waist (falls n=14, ADL n=31), wrist (falls n=12, ADL
n=15), or head (falls n=5, ADL n=12). ADL samples
represented dynamic activities (e.g. walking, walking on the
stairs, picking up object from the floor) and posture
Accelerations during the falls and ADL were measured with
body attached triaxial accelerometers, constructed using
three uniaxial capacitive accelerometers . Each triaxial
accelerometer was attached to a separate data logger (Onset
Computer Corp. Tattletale 8v2) with the sampling frequency
of 400 Hz .
Accelerometers were attached to the non-dominant wrist,
same side of the waist, and in front of the forehead. The
sensitive axes of the head and waist worn accelerometers
were mediolateral, anteroposterior, and vertical. The axes of
the wrist worn accelerometer were identical when the
subject stressed his/her arms (palm downwards) to the side.
C. Data Processing
Accelerometer data were loaded from data loggers to a
computer (software CrossCut 2.01, Borland International),
and converted into gravitational units with a custom-made
Manuscript received April 2, 2007. This work was supported in part by
the Finnish Funding Agency for Technology and Innovation, grant nr.
70074/05, EU Interreg IIIA Nord grant nr. 304-13723-2005, National
Semiconductor Finland, Elektrobit Ltd., CareTech Ab, and Elektropolis
M. Kangas is with the Department of Medical Technology, University
of Oulu, Oulu, Finland (phone: +358-8-537-6008; fax: +358-8-537-6000;
A. Konttila is with the Department of Medical Technology, and the
Optoelectronics and Measurement Techniques Laboratory, University of
Oulu, Oulu, Finland (e-mail: firstname.lastname@example.org).
I. Winblad is with the FinnTelemedicum, University of Oulu, and the
Department of Medical Technology, University of Oulu, Oulu, Finland
T. Jämsä is with the Department of Medical Technology, University of
Oulu, Oulu, Finland (e-mail: email@example.com).
Proceedings of the 29th Annual International
1-4244-0788-5/07/$20.00 ©2007 IEEE1367
MATLAB program. Data processing, analyses and fall
detection simulation were done with a custom-made
LabVIEW (8.0) program. Each of the three devices was
calibrated to generate acceleration signals of 1g and -1g
when parallel to the positive and negative axis of the
The measured acceleration signal was processed by
resampling at 50 Hz and median filtering (window length 3)
to reduce the data amount and noise before any further
analyses. The processed data were low-pass (LP) or high-
pass (HP) filtered (fc= 0.25 Hz) with a digital second order
Butterworth filter when necessary. The LP data were used in
posture analyses and HP data in motion analyses.
Fig. 1. (a) SVTOT. The start of the fall can be seen as the pit before the
impact peak, and impact is detected as the highest peak. (b) SVmaxmin.
The impact is detected as the highest peak.
Total sum vector SVTOT (Fig. 1a), containing both the
dynamic and static acceleration components, was calculated
from resampled data as indicated in (1).
( )( )2
where Ax, Ay, Az = acceleration (g) in x-, y-, and z-axes.
The start of the fall was determined as the pit before the
impact, SVTOT being equal or lower than 0.6 g. Dynamic
sum vector SVD was calculated similarly from the HP
filtered data by using (1).
Fast changes in the acceleration signal were investigated
by constructing a new sum vector SVmaxmin (Fig. 1b), which
was calculated using the differences between the maximum
and minimum acceleration values in a 0.6 s sliding window
for each axis. Vertical acceleration Z2 was calculated as
indicated in (2).
G SV SV
where SVTOT = total sum vector (g), SVD= dynamic sum-
vector (g), and G = gravitational component = 1 g.
The time period between the start of a fall and the impact
was determined from the measurements at the waist, by
recognizing the minimum value of SVTOT in the pit (SVTOT <
0.6 g) and the impact-related maximum peak of SVTOT or Z2.
Velocity v0, just before the fall associated impact, was
calculated by integrating the area around the pit (Fig. 1A)
where the SVTOT was lower than 1 g. The posture was
detected 2 seconds after the impact using the LP filtered
vertical signal , .
The threshold values for different parameters were
adjusted to optimal detection of falls with minimized false
alarms from ADL samples (maximal sensitivity with 100%
specificity when possible). For fall detection, lying posture
after fall was required. The posture determination was not
used with data measured from the wrist.
For comparison with previous studies, falling index (FI)
was also calculated for the waist measurements as described
earlier by Yoshida et al.  using a time window of 0.4 s.
When measured from the waist, the value ranges of the
parameters SVTOT, SVD, SVmaxmin, and Z2 were slightly
overlapping between falls and ADL (mean and quartile
values shown in Fig. 2). The threshold values set for waist
worn application are shown in Fig. 2 and summarized in
Table I. When posture detection after the fall was included,
the specificity of fall detection was 100% for all parameters
(Table I). The velocity v0 before the impact ranged from 0.8
to 3.4 ms-1 with an average value of 1.7 ms-1, 95% of the
falls having velocities over 1.0 ms-1. On average, the impact
was detected 0.3 s after the beginning of the fall, but the
time range was up to 1 s. Falling index FI had a value range
from 2.65 to 5.89 g and from 1.10 to 3.62 g during falls and
ADL, respectively, with an overall fall detection sensitivity
The value range of parameters SVTOT, SVmaxmin, and SVD
from wrist worn device overlapped clearly between falls and
ADL (Fig. 3). Z2 values overlapped slightly, but here 75%
of falls had higher value than the maximum from ADL (Fig.
3). Threshold values are shown in Fig. 3 and summarized in
Table I. The velocity v0 before the impact varied between
0.2 and 2.9 ms-1 with an average value 1.0 ms-1.
From the head, the value ranges of SVTOT, SVmaxmin, SVD,
and Z2 had specific value ranges for falls and ADL with no
overlapping (data not shown). Threshold values are
summarized in Table I. The specificity of fall detection was
100% for all parameters (Table I). The velocity v0 before the
impact ranged from 0.8 to 2.3 ms-1 with an average value of
This study investigated the acceleration signal measured
with body attached accelerometers from intentional falls and
activities of daily living (ADL) to determine threshold
values for multiple parameters capable of discriminating
between falls and ADL.
Our results showed that even if the different parameters
measured from the waist showed typical characteristics for
fall and ADL, the value ranges had some overlapping. This
indicates that using simple thresholds alone is not optimal
for practical fall detection. This is contrary to the report of
Bourke et al.  where they were able to determine a
simple SVTOT threshold value capable of discriminating
between falls and ADL with 100% sensitivity and
specificity. However, their experimental procedure used
young test subjects for fall events and elderly for ADL,
whereas we used same subjects for both samples. When we
included the posture detection after the fall, the thresholds
obtained resulted in a sensitivity and specificity of 95-100%.
Our threshold for the acceleration sum vector (SVTOT)
from waist was 2.0 g, which was smaller than the threshold
of 3.52 g presented earlier . This difference might be
partly explained by the median filtering used here, which
changes the absolute peak value of the impact signal.
All tested parameters measured from the head were able
to totally distinguish falls and ADL. This indicates that
head-worn accelerometer would be a reasonable choice for
fall detection, using e.g. hearing-aid housing as suggested
SV maxmin (g)
Fig. 2. Quartile box plots of parameters measured from waist during
falls and ADL. Selected threshold values (th) are marked (---). SVTOT
(th 2.0 g), SVmaxmin (th 2.0 g), SVD (th 1.7 g), and Z2 (th 1.5 g).
SV TOT (g)
SV maxmin (g)
Fig. 3. Quartile box plots of parameters measured from wrist during
falls and ADL. Selected threshold values (th) are marked (---). SVTOT
(th 5.2 g), SVmaxmin (th 6.5 g), and Z2 (th 3.9 g).
, if there were no limitations in usability and acceptance.
It would also be possible to integrate a fall detector to a
wrist watch . Even though the usability of a wrist-worn
fall detector is excellent, the acceleration signal measured
from wrist varies widely. Here, the signal of ADL samples
was strongly overlapping with that of falls. In addition,
posture detection is not applicable in wrist-worn
accelerometers. Thus, it appears that wrist is not an optimal
site for fall detection, and this placement site would need the
most complicated algorithm for fall detection as previously
suggested . However, the parameter Z2 shows the highest
potential to distinguish falls and ADL at the wrist.
At the waist, most of the falls had velocities at the typical
range shown previously for intentional falls. These
velocities are capable of resulting hip fractures among
elderly . At the head level, the minimum value of
velocity was at the range that has been used as a threshold in
fall detection algorithm by Lindemann et al. .
Testing our data with the earlier published FI threshold
resulted in fall detection sensitivity being in good
accordance with the sensitivity range of 40 to 100% reported
by Yoshida et al. , supporting our test procedure.
This was a pilot study with only two test subjects. Thus,
the results presented are only suggestive and experimental
studies with a larger population, preferably including mid-
aged or elderly subjects, are needed to confirm the results.
We conclude that head and waist are relevant sites for
accelerometric detection of falls, using simple thresholds
and posture detection. On the contrary, the wrist does not
appear to be an optimal site for fall detection.
The authors thank Mr. Erkki Vihriälä, M.Sc.Eng., for his
kind technical assistance
with the accelerometric
THRESHOLD VALUES FOR FALL DETECTION ALGORITHMS
Se / Sp
2.0 g 100/100 5.2 g
1.7 g 100/100 5.1 g
2.0 g 100/100
1.5 g 95/100
Se / Sp
3.9 g 75/100
Parameter Th Th Th
Se / Sp
2.0 g 100/100
1.2 g 100/100
1.7 g 100/100
1.8 g 100/100
Th = threshold, Se / Sp = sensitivity and specificity (%), g =
acceleration of gravity. a Posture detection after fall included
 K. Doughty, R. Lewis, and A. McIntosh, "The design of a practical
and reliable fall detector for community and institutional telecare," J.
Telemed. Telecare, vol. 6, Suppl. 1, pp. S150-154, 2000.
 A. Ozcan, H. Donat, N. Gelecek, M. Ozdirenc, and D. Karadibak,
"The relationship between risk factors for falling and the quality of life
in older adults," BMC Public Health, vol. 5, pp. 90, Aug 26. 2005.
 R. W. Sattin, D. A. Lambert Huber, C. A. DeVito, J. G. Rodriguez, A.
Ros, S. Bacchelli, J. A. Stevens, and R. J. Waxweiler, "The incidence
of fall injury events among the elderly in a defined population," Am.
J. Epidemiol., vol. 131, pp. 1028-1037, Jun. 1990.
 S. J. Brownsell, D. A. Bradley, R. Bragg, P. Catlin, and J. Carlier, "Do
community alarm users want telecare?" J. Telemed. Telecare, vol. 6,
pp. 199-204, 2000.
 S. Brownsell and M. S. Hawley, "Automatic fall detectors and the fear
of falling," J. Telemed. Telecare, vol. 10, pp. 262-266, 2004.
 D. M. Karantonis, M. R. Narayanan, M. Mathie, N. H. Lovell, and B.
G. Celler, "Implementation of a real-time human movement classifier
using a triaxial accelerometer for ambulatory monitoring," IEEE
Trans. Inf. Technol. Biomed., vol. 10, pp. 156-167, Jan. 2006.
 M. Makikawa, S. Asajima, K. Shibuya, R. Tokue, and H. Shinohara,
"Portable physical activity monitoring system for evaluation of
activity of the aged in daily life," in Proc. 2nd Joint EMBS/BMES
Conf., 2002, pp. 1908.
 M. J. Mathie, B. G. Celler, N. H. Lovell, and A. C. Coster,
"Classification of basic daily movements using a triaxial accelero-
meter," Med. Biol. Eng. Comput., vol. 42, pp. 679-687, Sep. 2004.
 J. R. Boyle, M. K. Karunanithi, T. J. Wark, W. Chan, and C. Colavitti.
"An observation trial of ambulatory monitoring of elderly patient," in
IFMBE Proc., vol. 12, 2005.
 A. K. Bourke, K. M. Culhane, J. V. O'brien, and G. M. Lyons, "The
development of an accelerometer and gyroscope based sensor to
distinguish between activities of daily living and fall-events," in
IFMBE Proc., vol. 11, 2005.
 M. N. Nyan, F. E. Tay, A. W. Tan, and K. H. Seah, "Distinguishing
fall activities from normal activities by angular rate characteristics and
high-speed camera characterization," Med. Eng. Phys., vol. 28, pp.
842-849, Oct. 2006.
 M. J. Mathie, A. C. Coster, N. H. Lovell, and B. G. Celler,
"Accelerometry: providing an integrated, practical method for long-
term, ambulatory monitoring of human movement," Physiol. Meas.,
vol. 25, pp. R1-20, Apr. 2004.
 T. Degen, H. Jaeckel, M. Rufer, and S. Wyss, "SPEEDY: A fall
detector in the wrist watch," in 7th IEEE Int. Symp. on Wearable
Computers, 2003, pp. 184.
 A. K. Bourke, J. V. O'brien and G. M. Lyons, "Evaluation of a
threshold-based tri-axial accelerometer fall detection algorithm," Gait
Posture, Nov 11, 2006 (Epub ahead of print).
 U. Lindemann, A. Hock, M. Stuber, W. Keck, and C. Becker,
"Evaluation of a fall detector based on accelerometers: a pilot study,"
Med. Biol. Eng. Comput., vol. 43, pp. 548-551, Sep. 2005.
 A. Diaz, M. Prado, L. M. Roa, J. Reina-Tosina, and G. Sanchez,
"Preliminary evaluation of a full-time falling monitor for the elderly,"
Conf. Proc. IEEE Eng. Med. Biol. Soc., vol. 3, pp. 2180-2183, 2004.
 E. Vihriälä, R. Saarimaa, R. Myllylä, and T. Jämsä. "A device for long
term monitoring of impact loading on the hip," Molecul. Quantum
Acoust., vol. 24, pp. 211-224, 2003.
 K. M. Culhane, G. M. Lyons, D. Hilton, P. A. Grace, and D. Lyons,
"Long-term mobility monitoring of older adults using accelerometers
in a clinical environment," Clin. Rehabil, vol. 18, pp. 335-343, 2004.
 T. Yoshida, F. Mizuno, T. Hayasaka, K. Tsubota, S. Wada, and T.
Yamaguchi. "A wearable computer system for a detection and
prevention of elderly users from falling," in IFMBE Proc., vol. 12,
 C. Smeesters, W. C. Hayes, and T. A. McMahon, "Disturbance type
and gait speed affect fall direction and impact location," J. Biomech.,
vol. 34, pp. 309-317, Mar. 2001.