Gait & Posture 30 (2009) 523–527

Contents lists available at ScienceDirect

Gait & Posture

journal homepage: www.elsevier.com/locate/gaitpost

Real-time gait event detection using wearable sensors

Michael Hanlon a,*, Ross Anderson b
a
Sport and Exercise Sciences Research Institute, University of Ulster, School of Sports Studies, Jordanstown, Co. Antrim, Northern Ireland
b
Biomechanics Research Unit, University of Limerick, Limerick, Ireland

A R T I C L E I N F O A B S T R A C T

Article history: Real-time gait event detection is a requirement for functional electrical stimulation and gait biofeedback.
Received 19 January 2009 This gait event detection should ideally be achieved using an ambulatory system of durable, lightweight,
Received in revised form 26 May 2009 low-cost sensors. Previous research has reported issues with durability in footswitch systems. Therefore,
Accepted 31 July 2009
this study describes the development and assessment of novel detection algorithms using footswitch
and accelerometer sensors on 12 healthy individuals. Subjects were equipped with one force sensitive
Keywords: resistor on the heel, one accelerometer at the foot, and one accelerometer at the knee. Subjects
Gait event
performed 10, 8-m walking trials in each of three conditions: normal, slow, and altered (reduced knee
Heel strike
ROM) walking. Data from a subset of four subjects were used to develop prediction algorithms for initial
Accelerometer
Footswitch contact (IC). Subsequently, these algorithms were tested on the remaining eight subjects against
Biofeedback standard forceplate IC data (threshold of 5 N on a rising edge). The footswitch force threshold algorithm
was most accurate for IC detection (mean absolute error of 2.4  2.1 ms) and was significantly more
accurate (p < 0.001) than the optimal accelerometer algorithm (mean absolute error of 9.5  9.0 ms). The
optimal accelerometer algorithm used data from both accelerometers, with IC determined from the second
derivative of foot fore-aft acceleration. The error results for footswitch and accelerometer algorithms are
lower (60%) than in previous research on ambulatory real-time gait event detection systems. Currently,
footswitch systems must be recommended over accelerometer systems for accurate detection of IC, however,
further research into accelerometer algorithms is merited due to its advantages as a durable, low-cost sensor.
ß 2009 Elsevier B.V. All rights reserved.

1. Introduction
Identification of temporal events in clinical gait analysis is
typically carried out using forceplates [1]. The instant when the
ground reaction force (GRF) value rises above a set threshold is
classified as initial foot contact (IC). At least two forceplates are
required to gather timing information for a complete stride, with at
least three required for bilateral stride information. Gait labs with
an insufficient number of forceplates to provide the required foot
contact data will usually supplement the forceplate data with
kinematic data. A large variety of kinematic algorithms exist which
use marker position data (recorded from optoelectronic motion
analysis systems) and derivatives of this data to identify gait events
(e.g. [1,2]).
Neither of these conventional systems meet the portable gait
event detection requirements of applications such as functional
electrical stimulation (FES [3]), ambulatory gait monitoring [4] and
gait biofeedback [5]. This has led to research into the use of several
wearable sensors (e.g. accelerometers, footswitches, gyroscopes,
* Corresponding author. Tel.: +44 28 90366388.
E-mail address: M.hanlon@ulster.ac.uk (M. Hanlon).
etc.) for detecting gait events [6–15]. These sensors offer
advantages as low-cost, low-power sensors that can provide data
in real-time on large numbers of strides in unrestricted environ-
ments.
The footswitch is the most commonly used wearable sensor for
gait event detection and the sensor of choice for most commercial
FES devices. Footswitches consist of a pressure/force sensor placed
on the plantar surface of the foot that produces a voltage relative to
the amount of force on the sensor. A common choice of footswitch
sensor is the force sensing resistor (FSR) due to its very slim build,
commercial availability, ease of electronic interfacing and fast
dynamic response [7]. While not suitable for precise measure-
ments, the selection of an appropriate threshold allows FSRs to give
simple on–off contact information. Validation studies of foots-
witches versus forceplates for event detection have produced
varying results [3,7,11–15]. Hausdorff et al. [7] reported mean
errors in IC detection of 0  3 ms (range 10 to +8 ms) using two
FSRs and a three-step automated detection approach which cannot
function in real-time. Using a similar system, Mills et al. [13] reported
errors of 2.7  2.7 ms for IC. In contrast to these relatively low errors,
Wall and Crosbie [14] reported mean errors of 45 and 10 ms for heel
strike in shod and barefoot conditions, respectively, while Smith et al.
[3] reported mean errors for IC detection of 30 ms using a two FSR

0966-6362/$ – see front matter ß 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.gaitpost.2009.07.128
524 M. Hanlon, R. Anderson / Gait & Posture 30 (2009) 523–527
system. In addition, Mansfield and Lyons [15] reported spurious heel
strike detection (e.g. heel strike during pre-swing) in 4% of trials when
using footswitches. Therefore, despite the impressive results shown
by Hausdorff et al. [7], many subsequent studies have shown reduced
accuracy and precision from similar footswitch systems. This reduced
temporal accuracy aside, several authors have also commented
negatively on the structural reliability of footswitch systems
[8,15,16]. The specific placement of footswitches also rules them
out for implantation, which is of benefit in FES systems.
Consequently, recent research has focussed on alternatives to
footswitches, such as accelerometers and gyroscopes [17–19]. In
the validation studies of these sensors that provide timing errors,
none of the systems can detect gait events in real-time and the
timing errors are still notably larger than those reported for
footswitch systems. The timing errors reported for off-line IC
detection include 15  16 ms using a single tri-axial accelerometer
[10], 34  25 ms for two uniaxial accelerometers [20], 11  23 ms for
a combined accelerometer and gyroscope system [8], and 10 ms (95%
confidence interval of 7–13 ms) for a three goniometer system [6]. It
should be noted that gyroscopes do have certain weaknesses as gait
detection sensors, in comparison to accelerometers. These include the
fact that piezoelectric gyroscopes are less structurally robust, are
more sensitive to temperature and shock, and require powerful
filtering to cancel drift and artefacts in the signal [6].
In this study, an investigation is conducted to develop an
accurate ambulatory sensor system capable of real-time gait event
detection. Where previous studies have used algorithms which
incorporate data post-IC in the determination of IC, the algorithms
in this study must make predictions of IC based solely on data pre-
IC. This real-time gait event detection is necessary for applications
such as FES and gait biofeedback training. Footswitches and
accelerometers will be investigated as, based on the literature,
these are the two sensors which currently offer the highest
accuracy and sensor integrity, respectively. The flexibility of these
sensors will be assessed by testing their accuracy under three
different walking conditions (slow, normal and altered). Absolute
error values (in addition to mean relative errors) will also be
provided to comprehensively describe the magnitude and bias of
timing errors in the developed systems. Both systems will be
compared against criterion measurements from a forceplate as the
recognised gold standard for gait event detection [2].
2. Methodology
2.1. Subjects
Twelve healthy adult volunteers, eight males and four females (age: 25  3
years), were familiarised with the experimental procedure and all possible risks before
providing written consent to participate as approved by the University’s Research
Ethics Committee. Subjects were barefoot and wore a Donjoy eXtreme armor action
knee brace (dj Orthopaedics, Vista, USA) on their right knee to simulate an external
device worn for FES/biofeedback. Subjects were required to perform 10 successful
(right foot fully within forceplate) walks along an 8-m walkway in each of three
conditions: (a) normal walking; (b) slow walking; (c) altered walking—mechanical
stops were used on the knee brace to limit right knee sagittal plane range of motion to
208 (from approximately 158 to 358 of flexion) in this condition. This limitation was
used in an attempt to simulate a form of hemiplegic gait.
Subjects took the first step of each walk with their left foot and subsequently
struck the platform with the second step of their right foot (fourth step overall). This
ensured that the relevant step data was not from the gait initiation stage. To avoid
potential forceplate targeting [21] subjects were told to focus on a point (at head
height) on the wall in front of them.
2.2. Equipment
Subjects were equipped with two dual-axis accelerometers (10 g, ADXL210,
Analog Devices) during testing. The first accelerometer was placed on the lower medial
border of the knee brace. The second accelerometer was placed just below the lateral
malleolus of the right foot (see supplementary material for accelerometer images). In
the standing position, the y-axis of both accelerometers was aligned vertically (by eye),
while the x-axis was aligned in the fore-aft direction. Both accelerometers were taped
securely in position and hardwired to a connector unit at the hip.
One FSR (Interlink Electronics, Camarillo, USA) was attached to the underside of
the subject’s right heel. The FSR has a sensing circle with radius of 12.5 mm and a
thickness of 0.46 mm. A protective plastic sheath was placed around the FSR,
followed by a layer of insulating tape for extra protection, thus increasing the
thickness of the FSR device to 1.2 mm. The FSR was wired in series with a 10 kV
resistor and hard-wired back to the hip connector unit (tethered to data acquisition
hardware).
A forceplate (AMTI, Watertown, USA) provided the criterion measure of IC. A 6-
camera Hawk motion analysis system (Motion Analysis Corporation, Santa Rosa,
USA) captured kinematics of a marker placed over each subject’s hip connector unit
(used to calculate gait speed). All marker (200 Hz), forceplate and sensor data (both
1000 Hz) were then acquired using EVaRT 4.4 software (Motion Analysis
Corporation).
2.3. IC detection
Forceplate thresholds of 2.5 N [14], 10 N [22], 20 N [20] have been used in
previous studies. Selection of a lower threshold should give a closer approximation
of the actual IC, once false early detections are avoided. Using this theoretical
foundation, and based on pilot test data, a forceplate threshold of 5 N on a rising
edge was selected as the criterion measure for identifying IC. Additional sensor
algorithms were developed using an evolutionary process based on an initial subset
of four subjects. Data from this subset were analysed to create algorithms that could
then be tested on the remaining eight subjects. This method is in agreement with
that of Willemsen et al. [16]. Data from the four-subject subset were graphed, and
descriptive statistics were calculated relating to gait speeds, vertical GRF values, FSR
values and accelerometer values. Filters, with a range of cut-off values, were then
applied to the accelerometer data, thus providing more time-series datasets. In
addition, first, second and third derivatives of acceleration were also calculated.
Each newly created accelerometer time-series dataset was analysed to identify
repeatable quantifiable events which were time-matched with the initial rise in
GRF.
2.3.1. FSR algorithms
Fig. 1 displays a typical scatter plot of FSR and GRF data during the initial contact
phase. FSR values saturate at a maximum value during each early stance phase at a
GRF value approximating 120 N. Following analysis of FSR and GRF values for the
first four healthy subjects, an FSR threshold value was selected balancing the
requirements that the threshold is crossed close to IC, yet not crossed during swing.
A threshold value representing 39% of the maximum FSR value (on a rising edge)
was selected as the FSR threshold. Due to the steep FSR curve response observed at
IC (see Fig. 1), this threshold was typically crossed within 3 ms of IC. This threshold
was also selected as it represented twice the highest FSR value seen during the
swing phase (of four-subject subset), and thus was unlikely to be crossed prior to IC.
In addition to the FSR force algorithm, an additional FSR algorithm was designed
based on the first derivative of FSR values (threshold = 3900 N/s). This algorithm
was utilised as, theoretically, the first derivative of FSR data shows large values at IC
(when there is a rapid rise in force), while having relatively low values throughout
the rest of the stride.
2.3.2. Accelerometer algorithms
Based on results from the first four subjects, 41 accelerometer algorithms were
deemed suitable for full analysis on the test cohort of eight subjects. The key
adjustable components of accelerometer algorithms are as follows (see Table 1):
Fig. 1. Exemplar scatter plot of FSR values versus synchronous force platform values
during initial foot contact (seven data points pre-IC and seven data points post-IC).
FSR force threshold is passed 1 ms (1 sample) after the criterion force platform
threshold of 5 N.
 M. Hanlon, R. Anderson / Gait & Posture 30 (2009) 523–527 525

Table 1 Table 2
Options for adjustable components in accelerometer algorithms. Jerk and snap are Error results for FSR force, FSR derivative and AccA algorithms in eight test subjects
the first and second derivatives of acceleration, respectively. with all walk data combined. All values are in milliseconds.

Threshold value Data type Filter type Accelerometer direction Algorithms Mean relative Minimum Maximum Mean absolute
error (SD) error error error (SD)
Set Acceleration None Ankle fore-aft
Time-based Jerk 1 Hz Ankle vertical FSR Force 1.5 (2.8) 12 13 2.4 (2.1)
None Snap 5 Hz Knee vertical FSR derivative 2.7 (3.3) 12 4 3.3 (2.7)
30 Hz AccA 6.7 (11.2) 53 39 9.5 (9.0)

(a) Threshold value (set, time-based, or none). A set threshold is one that is
independent of the subject’s individual walking characteristics. It depends
solely on the speed set in which the subject is walking. A time-based threshold
is one that is trial specific. It uses accelerometer values from a portion of the
stride cycle to create related threshold values. No threshold means that a
specific local maximum or local minimum turning point is used instead of any
threshold value.
(b) Data type (acceleration, jerk or snap). Acceleration data is the directly
measured accelerometer signal. Jerk and snap are the first and second
derivatives of this acceleration data, respectively. The use of acceleration
values has an advantage over the use of jerk and snap data in terms of signal–
noise ratio [23]. However, jerk and snap values allow a numerical differentia-
tion between instances when the rate of acceleration change is high or low,
even when the acceleration values are similar.
Filter type (none, 1 Hz, 5 Hz or 30 Hz). When filtering, a Butterworth fourth
(c)
order low-pass filter was used. The 1 Hz cut-off was used to produce data
with a more consistent presentation of local minimum/maximum values.
The 30 Hz cut-off was used to produce data that maintained most of the
high frequency data relating to impact, while eliminating the highest
frequency noise. The 5 Hz cut-off was used as a compromise between
extreme low and high cut-offs. This filtering is particularly important in
improving the signal:noise ratio in jerk and snap data [23,24]. Raw data is
also used as this data contains all the high frequency impact data relating to
IC.
(d) Accelerometer direction (ankle fore-aft, ankle vertical or knee vertical).
2.3.3. Sensor assessment
Custom-built software was designed using LabVIEW 7.1 (National Instruments,
Austin, USA) to allow efficient algorithm development and the subsequent off-line
testing of selected algorithms on the remaining eight subjects. Sensor results were
compared using Mann–Whitney U-tests at a 95% confidence level due to non-
normal data distribution.
3. Results
The results for IC using a 5 N forceplate threshold were used as
the reference criterion against which all other algorithms were
assessed. Deviations from the reference criterion value are termed
errors and measured in milliseconds. Negative error values
indicate early algorithm prediction errors, while positive values
indicate late algorithm predictions. The most accurate acceler-
ometer algorithm of the 41 tested was entitled AccA. Table 2
contains the error results for sensors based on aggregate scores for
all walking protocols. Table 3 displays the error results for sensors
in each of the three walking protocols.
The FSR force threshold algorithm was most accurate for IC
detection (average absolute error of 2.4 ms based on all walking
conditions). The absolute errors using this algorithm were
significantly less (p < 0.001) than both the FSR derivative and
optimal accelerometer algorithms.
The error results for each sensor’s optimal algorithm are
illustrated in Fig. 2 using Bland and Altman style plots [25,26].
These graphs illustrate the spread of error values, any skewness
towards early or late estimation and any relationship between
error and gait speed. Both graphs illustrate a mean positive error
and a reduction in average error range as gait speed increases.
The optimal accelerometer algorithm (AccA) used data from
both accelerometers to detect IC. The functionality of AccA is
presented in Eq. (1). AccA uses the local minimum in 1 Hz filtered
knee vertical acceleration data to approximate mid swing. It then

Table 3
Error results for FSR force, FSR derivative and AccA algorithms in eight test subjects for each individual walk type. All values are in milliseconds.

Algorithms Normal walking Slow walking Altered walking

Mean relative Mean absolute Mean relative Mean absolute Mean relative Mean absolute
error (SD) error (SD) error (SD) error (SD) error (SD) error (SD)

FSR force 2.0 (2.0) 2.2 (1.7) 1.4 (2.9) 2.5 (1.9) 1.0 (3.4) 2.5 (2.5)
FSR derivative 1.7 (2.8) 2.4 (2.2) 3.1 (3.8) 3.9 (2.9) 3.4 (3.1) 3.7 (2.6)
AccA 2.9 (8.4) 5.7 (6.8) 10.2 (11.0) 12.8 (7.8) 7.1 (12.7) 10.1 (10.5)

Fig. 2. Bland and Altman style plots of (A) error in predicted IC time between FSR algorithm and the reference force platform criterion and (B) error in predicted IC time
between AccA algorithm and the reference force platform criterion. Negative errors indicate early IC prediction by the specific algorithm.
526 M. Hanlon, R. Anderson / Gait & Posture 30 (2009) 523–527
searches after this mid swing event for the first 30 Hz filtered ankle
fore-aft snap value above/below a set positive/negative threshold.
These high magnitude positive and negative spikes in ankle fore-
aft snap at IC are indicative of the rapid changes in ankle fore-aft
jerk and acceleration values at IC.
Tc ¼ Tn
(1)
when ðSn > PT _ Sn < NTÞ ^ T n > T 1
where Tc: time of contact; Tn: time at frame n; Sn: 30 Hz foot fore-
aft snap at frame n; PT: positive snap threshold; NT: negative snap
threshold; T1: time of 1st local minimum in 1 Hz filtered knee
accelerometer data.
4. Discussion
The goal of this study was to investigate the use of potential
mobile systems for real-time IC detection. A footswitch-based
(FSR) system and accelerometer-based system were tested. The
results for the most accurate FSR algorithm were very good (mean
relative error of 1.5  2.8 ms) and are just 1.5 ms worse than the
results of Hausdorff et al. [7] and 1.2 ms better than the results of
Mills et al. [13], who both used more complex algorithms that cannot
function for real-time IC detection. In addition, neither the Hausdorff
or Mills study used patients or altered walking styles to test their
prediction algorithms, while conversely, one third of predictions in
this present study were based on altered walking styles. The FSR
prediction errors are also distinctly better than those shown by
several other studies using similar hardware [3,14,15]. In contrast to
reports in previous studies [8,16], no problems with FSR reliability or
durability were recorded in this study.
The low error results seen with altered walking demonstrate
some robustness of the algorithms against shifts in the walking
pattern, however adjustments to the algorithms may be required
for use with gait pathologies causing greater alterations in the
walking pattern. Similarly, the FSR placement in this study was
designed to identify heel contact as the indicator of IC, and such
detection systems would need to be adjusted in pathologies where
no heel strike occurs.
The results for the most accurate accelerometer algorithm were
good (mean relative error of 6.7  11.2 ms). The average error
magnitude is similar to that reported by Aminian et al. [6] for an
offline three gyroscope system. The errors reported in the present
study for AccA are also less than other previously reported
ambulatory systems using either accelerometers or gyroscopes [8–
10,15,20,27]. The ideal accelerometer detection system is one that
employs just one accelerometer (based on cost, instrumentation time,
flexibility, etc.). The proposed accelerometer system does not meet
this requirement as AccA integrates data from two accelerometers to
detect IC. This increases the time required for sensor placement and
limits the potential for implantability. As an alternative, the most
accurate single accelerometer algorithm had a mean absolute timing
error of 39.4 ms. These results are specific to the accelerometer
attachments used in this study and adjustments may be required to
optimise accelerometer algorithms when different surface types/
attachment sites are used. Despite less accurate IC detection, the
advantage of the accelerometer as a more durable sensor may make it
a more suitable sensor for certain applications where the precision of
gait event detection is less important (e.g. long-term activity
monitoring).
The most accurate FSR algorithm showed significantly lower
timing errors than the most accurate accelerometer algorithm
(AccA). The practical benefits of this improved accuracy are
dependant on the proposed application for gait event detection.
With typical gait analysis studies sampling kinematics at
frequencies between 50 and 100 Hz, the majority of errors shown
by the accelerometer system would be within only 1–2 samples of
the actual event. Ghoussanyi et al. [22] concluded that algorithms
with errors in this range were sufficiently accurate for gait event
detection. However, reports on the effect of gait speed on
kinematic data [28] indicate increased errors in kinematic data
for each added millisecond of event timing error. With peak knee
angular velocities of approximately 3008/s for gait speeds above
1 m s1, errors of 1 ms in IC detection would have the effect of
producing 0.38 errors in knee angle values during these periods of
high angular velocity (1.28 angular error for just 4 ms timing error).
With angular errors already a well documented issue due to
marker placement errors and skin movement artefact [29], these
findings emphasise the importance of minimising the error in IC
detection in order to avoid further magnification of angular errors.
The use of snap data (derived from an accelerometer) in this
study appears novel. Algorithms using snap and jerk data showed
lower timing error values than those using acceleration. Post hoc
analysis revealed that this may be related to snap and jerk data
showing greater proportional change at impact events than
acceleration data, or because snap and jerk data are less affected
than acceleration data by accelerometer alignment errors.
5. Conclusion
Real-time algorithms were developed, using data from either
foot and knee accelerometers or an FSR, and shown to accurately
detect the IC event in slow, normal and altered walking. The FSR
system showed significantly lower errors than the accelerometer
system, and both systems displayed improved accuracy compared
to previously reported real-time ambulatory systems.
Acknowledgements
The authors acknowledge funding from the Irish Research
Council for Science, Engineering and Technology, funded by the
National Development Plan.
Conflict of interest statement
There is no conflict of interest related to the preparation of this
manuscript or the research work discussed herein.
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in
the online version, at doi:10.1016/j.gaitpost.2009.07.128.
References
[1] Hreljac A, Marshall RN. Algorithms to determine event timing during normal
walking using kinematic data. J Biomech 2000;33:783–6.
[2] O Connor M, Thorpe SK, O Malley MJ, Vaughan CL. Automatic detection of gait
events using kinematic data. Gait Posture 2007;25:469–74.
[3] Smith BT, Coiro DJ, Finson R, Betz RR, McCarthy J. Evaluation of force-sensing
resistors for gait event detection to trigger electrical stimulation to improve
walking in the child with cerebral palsy. IEEE Trans Neural Syst Rehabil Eng
2002;10:22–9.
[4] Zijlstra W. Assessment of spatio-temporal parameters during unconstrained
walking. Eur J Appl Physiol 2004;92:39–44.
[5] Hanlon M, Anderson R. Learning gait patterns—kinematic biofeedback versus
conventional training. In: Proceedings of the North Eastern American Society
of Biomechanics; 2007.
[6] Aminian K, Najafi B, Bula C, Leyvraz P, Roberts P. Spatio-temporal parameters
of gait measured by an ambulatory system using miniature gyroscopes. J
Biomech 2002;35:689–99.
[7] Hausdorff JM, Ladin Z, Wei JY. Footswitch system for measurement of the
temporal parameters of gait. J Biomech 1995;28:347–51.
[8] Jasiewicz JM, Allum JHJ, Middleton JW, Barriskill A, Condie P, Purcell B, et al.
Gait event detection using linear accelerometers or angular velocity transdu-
cers in able-bodied and spinal-cord injured individuals. Gait Posture
2006;24:502–9.
 M. Hanlon, R. Anderson / Gait & Posture 30 (2009) 523–527
[9] Pappas IPI, Keller T, Popovic MR. Validation of a new gait phase detection
system. Gait Posture 2001;13:301–2.
[10] Zijlstra W, Hof AL. Assessment of spatio-temporal gait parameters from trunk
accelerations during human walking. Gait Posture 2003;18:1–10.
[11] Bamberg SJM, Benbasat AY, Scarborough DM, Krebs DE, Paradiso JA. Gait
analysis using a shoe-integrated wireless sensor system. IEEE Trans Inf Tech-
nol Biomed 2008;12:413–23.
[12] Skelly MM, Chizeck HJ. Real-time gait event detection for paraplegic FES
walking. IEEE Trans Neural Syst Rehabil Eng 2001;9:59–68.
[13] Mills PM, Barrett RS, Morrison S. Agreement between footswitch and ground
reaction force techniques for identifying gait events: inter-session repeat-
ability and the effect of gait speed. Gait Posture 2007;26:323–6.
[14] Wall JC, Crosbie J. Accuracy and reliability of temporal gait measurement. Gait
Posture 1996;4:293–6.
[15] Mansfield A, Lyons GM. The use of accelerometry to detect heel contact events
for use as a sensor in FES assisted walking. Med Eng Phys 2003;25:879–85.
[16] Willemsen ATM, Bloemhof F, Boom HBK. Automatic stance-swing phase
detection from accelerometer data for peroneal nerve stimulation. IEEE Trans
Biomed Eng 1990;37:1201–7.
[17] Aminian K, Najafi B. Capturing human motion using body-fixed sensors:
outdoor measurement and clinical applications. Comp Animat Virt W
2004;15:79–94.
[18] Lau H, Tong K. The reliability of using accelerometer and gyroscope for gait
event identification on persons with dropped foot. Gait Posture 2008;27:248–
57.
[19] Maffiuletti NA, Gorelick M, Kramers-de Quervain I, Bizzini M, Petrich-Mun-
zinger J, Tomasetti S, et al. Concurrent validity and intrasession reliability of
527
the IDEEA accelerometry system for the quantification of spatiotemporal gait
parameters. Gait Posture 2008;27:160–3.
[20] Selles RW, Formanoy MAG, Bussmann HBJ, Janssens PJ, Stam HJ. Automated
estimation of initial and terminal contact timing using accelerometers: devel-
opment and validation in transtibial amputees and controls. IEEE Trans Neural
Syst Rehabil Eng 2005;13:81–8.
[21] Kirtley C. Clinical gait analysis: theory and practice. New York: Churchill
Livingstone; 2005.
[22] Ghoussayni S, Stevens C, Durham S, Ewins D. Assessment and validation of a
simple automated method for the detection of gait events and intervals. Gait
Posture 2004;20:266–72.
[23] Winter DA. Biomechanics and motor control of human movement, 2nd ed.,
New York: John Wiley & Sons; 1990.
[24] Giakas G. Power spectrum analysis and filtering. In: Stergiou N, editor.
Innovative analysis of human movement. Champaign, IL: Human kinetics;
2004.
[25] Bland JM, Altman DG. Comparing methods of measurement: why plotting
difference against standard method is misleading. Lancet 1995;346:1085–7.
[26] Bland JM, Altman DG. Statistical methods for assessing agreement between
two methods of clinical assessment. Lancet 1986;I:307–10.
[27] Williamson R, Andrews BJ. Gait event detection for FES using accelerometers
and supervised machine learning. IEEE Trans Rehabil Eng 2000;8:312–9.
[28] Hanlon M, Anderson R. Prediction methods to account for the effect of gait
speed in lower limb angular kinematics. Gait Posture 2006;24:280–7.
[29] Lucchetti L, Cappozzo A, Capello A, Della Croce U. Skin movement artefact
assessment and compensation in the estimation of knee-joint kinematics. J
Biomech 1998;31:977–84.