Journal of Biomechanics 35 (2002) 689–699

Spatio-temporal parameters of gait measured by an ambulatory

system using miniature gyroscopes
. b, P.-F. Leyvrazc, Ph. Roberta
K. Aminiana,*, B. Najafia, C. Bula
a
Metrology Laboratory, Swiss Federal Institute of Technology (EPFL), CH-1015 Lausanne, Switzerland
b
Service of Geriatric Medicine, CHUV & CUTR Sylvana, Epalinges, Switzerland
c
#
Hopital Orthop!edique de la Suisse Romande, Lausanne, Switzerland
Accepted 28 December 2001

Abstract

In this study we describe an ambulatory system for estimation of spatio-temporal parameters during long periods of walking. This
original method based on wavelet analysis is proposed to compute the values of temporal gait parameters from the angular velocity
of lower limbs. Based on a mechanical model, the medio-lateral rotation of the lower limbs during stance and swing, the stride length
and velocity are estimated by integration of the angular velocity. Measurement’s accuracy was assessed using as a criterion standard
the information provided by foot pressure sensors. To assess the accuracy of the method on a broad range of performance for each
gait parameter, we gathered data from young and elderly subjects. No significant error was observed for toe-off detection, while a
slight systematic delay (10 ms on average) existed between heelstrike obtained from gyroscopes and footswitch. There was no
significant difference between actual spatial parameters (stride length and velocity) and their estimated values. Errors for velocity
and stride length estimations were 0.06 m/s and 0.07 m, respectively. This system is light, portable, inexpensive and does not provoke
any discomfort to subjects. It can be carried for long periods of time, thus providing new longitudinal information such as stride-to-
stride variability of gait. Several clinical applications can be proposed such as outcome evaluation after total knee or hip
replacement, external prosthesis adjustment for amputees, monitoring of rehabilitation progress, gait analysis in neurological
diseases, and fall risk estimation in elderly. r 2002 Published by Elsevier Science Ltd.

Keywords: Gait analysis; Gyroscope; Wavelet analysis; Ambulatory system; Stride velocity

1. Introduction A gait laboratory based on camera, walkway with

implanted sensors (Hirokawa and Matsumara, 1987) or
Walking is one of the most common human physical force-plates, and electromyography allows a complete
activities. Evaluation of time and distance parameters gait analysis but require a dedicated laboratory.
during walking is helpful in assessing abnormal gait, to Although these techniques have been widely used for
quantify improvement resulting from interventions, or research purposes, their sophistication, time required for
to predict subsequent events such as falls. This evalua- setting up the instrumentation and analyze the data, as
tion is useful in several clinical situations such as: well as cost, have hindered their use in clinical practice.
functional performance assessment after treatment or In addition, these techniques require the subjects to walk
surgery such as hip and knee arthroplasty (Wall et al., in a pre-defined specific path, and assume that data
1981), refining proper alignment and fit of external measured from only a few steps are representative of
prosthesis or orthesis (Czerniecki and Gitter, 1996), fall usual gait performance. To avoid these limitations,
risk assessment in elderly persons (Maki, 1997), or ambulatory recording systems carried by the subject and
selecting the appropriate assistive device. allowing outdoor measurement have been developed,
using new technology, such as a powerful microcon-
troller, miniature sensors, high capacity memory and
*Corresponding author. Tel.: +41-21-693-2617; fax: +41-21-693-
small batteries.
2614. Ambulatory systems using foot switches or pressure
E-mail address: kamiar.aminian@epfl.ch (K. Aminian). sensors attached to the sole are used to monitor

0021-9290/02/$ - see front matter r 2002 Published by Elsevier Science Ltd.

PII: S 0 0 2 1 - 9 2 9 0 ( 0 2 ) 0 0 0 0 8 - 8
690 K. Aminian et al. / Journal of Biomechanics 35 (2002) 689–699

temporal parameters (Zhu et al., 1991; Abu-Faraj 2. Method

et al., 1997). These techniques provide satisfactory
results for normal walking, but sensor attachment 2.1. Gyroscopes
proves difficult when assessing subjects with abnormal
gait. Even when correctly done, several problems, such Lower limbs movement during walking was measured
as shuffling gait, mechanical failure, or patient accep- using miniature gyroscopes. The gyroscope consists of a
tance limit their applicability. Another system using vibrating element coupled to a sensing element, acting as
insole with a matrix of multitude pressure sensors has a a Coriolis sensor. Coriolis effect is an apparent force
high reliability to evaluate the time of each phase of the that arises in a rotating reference frame and is
gait cycle. Although this system provides additional proportional to the angular rate of rotation. The
information on foot pressure distribution during gait, it piezoelectric gyroscope measures the Coriolis accelera-
cannot be used for a long period of measurement, and tion generated when a rotational angular velocity is
sensors do not provide information during the swing applied to the oscillating piezoelectric bimorph. Three
phase. miniature piezoelectric gyroscopes (Murata, ENC-03J)
Recently, technical progress has made possible the were used. Because of their low current consumption,
realization of miniature kinematic sensors such as these sensors can be battery powered and are therefore
accelerometers and angular rate sensors with integrated appropriate for ambulatory monitoring. The signal from
conditioning and calibrating module (Sparks et al., each gyroscope was amplified and low-pass filtered to
1998). In addition, due to their very low consump- remove the electronic noise of the gyroscope. Each
tion, these sensors can be battery powered and are gyroscope and the conditioning electronic were placed in
promising tools for long-term ambulatory monitoring. a small circular box (D ¼ 25 mm, H ¼ 10 mm) (Fig. 1).
The possibility of detecting simple parameters such as A swiveling electromechanical system was designed to
steps and cycle time from trunk or heel acceleration has calibrate and characterize each sensor. Various angular
already been demonstrated using miniature acceler- velocities were applied to the gyroscope. An optical
ometers (Evans et al., 1991; Aminian et al., 1995). More sensor measured the time corresponding to perform a
recently the efficiency of accelerometers to detect rotation of 3601 and allowed calibration of the gyro-
temporal features of walking has been shown (Aminian scope. The amplifier gain of each gyroscope was set to
et al., 1999). However, the acceleration signal is often obtain a sensitivity of 4 mV/deg/s. The gyroscopes were
biased by the gravitational acceleration and depends on attached with a rubber band to each shank and on the
the attachment site along the segment (e.g. limb). Since right thigh. Each sensor measured the angular rate
the actual nature of walking consists of lower limb rotation parallel to the mediolateral axis (perpendicular
rotation around joint articulation, the use of miniature to the sagittal plane). The signals were digitized (12 bit)
angular rate sensors (gyroscopes) has proven to at a sampling rate of 200 Hz by a portable data logger
be an alternative technique for gait analysis (Tong and (Physilog, BioAGM, CH) and stored on a memory card.
Granat, 1999; Nene et al., 1999). However, no At the end of the recording the data were transferred to
clear correspondence has been established between the computer for analysis.
gait events and the patterns of angular velocity.
This is probably because gait events (e.g. heelstrike 2.2. Footswitches
and toe-off) are transitory signals that cannot be
properly enhanced by traditional signal processing. Footswitches have been used as criterion standard to
We decided to use wavelet transform to overcome detect the gait events and validate the gyroscope
this problem because this technique has proven its measuring method. The footswitches consisted of two
superiority over traditional filtering techniques in pairs of force sensing resistors, FSR (Interlink, LU)
simple and complex kinematic data (Ismail and Asfour,
1999).
In this study, we describe an ambulatory system for
estimation of spatio-temporal parameters during long
periods of walking. An original method, based on
wavelet transform is proposed to compute the values of
gait parameters from the angular velocity of lower
limbs. The accuracy of measurements was assessed,
using as a criterion standard the information provided
by foot pressure sensors. In order to estimate the
accuracy of the method on a broad range of perfor-
mance for each gait parameter, data were gathered from Fig. 1. Gyroscope module with embedded electronic. The gyroscopes
young and elderly subjects. are attached on shank segments and right thigh.
 K. Aminian et al. / Journal of Biomechanics 35 (2002) 689–699 691

Shank
Angular rate
deg/s
300

200

100 toe-off
heelstrikes
peaks
peaks
0

-100

-200

0 0.2 0.4 0.6 0.8 Time,s

Fig. 3. Shank angular velocity showing the presence of peaks during

the toe-off and heelstrike.

Fig. 2. Footswitches (FSR) attachment beneath the heel (two sensors) foot first touches the floor) and toe-off (when it takes
and beneath the big toe (two sensors). off) during that cycle. These events have distinctive
signal features of shank angular velocity appearing as
rather sharp negative peaks involving some medium and
placed under the right foot (Fig. 2): FSRheel beneath the relative high frequencies (Fig. 3). Although the ampli-
heel (two sensors) and FSRtoe beneath the big toe (two tude of these peaks vary according to various para-
sensors). For each pair, the sensors were connected in meters such as subject’s velocity or weight, or the
series and the outputs were simultaneously digitized and presence of limping due to a painful articulation, they
recorded by Physilog with those from the gyroscopes at can always be localized as long as one knows where to
a sampling rate of 200 Hz. During heelstrike and toe-off, look for it in time and frequency domains. To
FSRheel and FSRtoe signals are subjected to abrupt successfully achieve this task, we have decomposed the
changes. By thresholding the derivative of these two shank angular velocity into wavelet packages. Wavelet
signals the exact time of heelstrike and toe-off was transformation is well adapted for gait events identifica-
obtained. tion because it allows detection of a specified frequency
at a specified time (e.g. toe-off and heelstrike). In other
2.3. Subjects words, toe-off and heelstrikes events consist of com-
bined features that can be well located in the time and
Measurement using gyroscopes and FSR were taken frequency range. The methods of analysis need to be
from a group of 9 young (2172 yr) and 11 elderly versatile enough to handle gait events that can be at
(7978 yr) subjects. Each young subject performed four opposite extremes in terms of their time–frequency
trials including at least 20 gait cycles. The first three location. Moreover, wavelets transformation allows the
trials were performed on a treadmill at the preferred use of a suitable wavelet function (Coiflet wavelet),
speed, under and over the preferred speed, respectively. which is more similar to the pattern of heelstrike and
The fourth trial was performed with a comfortable toe-off.
velocity over ground on a 30 m long walkway. The A multi-resolution wavelet decomposition (Mallat,
elderly performed only two trials on the walkway at 1989) was used to enhance the toe-off and heel-strike
their comfortable velocity. The study protocol was event during walking. This technique consists in splitting
accepted by the institutional review board (Faculty of a signal into high-scale (low-frequency components)
Medicine, University of Lausanne). Written informed called approximation and low-scale (high-frequency
consent was obtained from each subject. components) called detail. The decomposition is iter-
ated, with successive approximations (or details) being
2.4. Wavelet analysis: gait temporal parameters decomposed in turn, so the original signal is broken
estimation down into many lower-resolution components. At each
decomposition scale the number of samples in the time
In order to compute the temporal parameters such as domain is decreased by throwing away every second
the duration of swing, single and double stances during sample (down sampling with a factor of ‘2’). By
a gait cycle, it is necessary and sufficient to determine for considering the original signal sðnÞ (shank angular
each leg the precise moments of heelstrike (when the velocity), the approximation of the signal at scale j ¼ 0
692 K. Aminian et al. / Journal of Biomechanics 35 (2002) 689–699

is A2o s which corresponds to the original signal sðnÞ: At s

each new decomposition the approximation and detail 9 Wavelet Approximation 9
sa=∑D2js sa=∑ D2j s
decrease in resolution since at each scale 2 samples will 1 1
be replaced by one. At scale j the A2j s represents Wavelet Analysis Wavelet Analysis
approximation of sðnÞ with a resolution of one sample A21sa-A29sa A23sa-A29sa
for every 2j samples of the original signal. We have
global maximum global maximum
s ¼ A21 s þ D21 s ¼ A22 s þ D22 s þ D21 s ¼ ? peak detection
peak detection
¼ A2j s þ D2j s þ D2j1 s? þ D21 s: ð1Þ ms

local minima detection

Thus inside [ms-2s, ms-0.05s]
ms
A2j s ¼ A2jþ1 s þ D2jþ1 s: ð2Þ
local minima detection hs to local minima detection of s
By using a suitable low-pass filter h; and a high-pass inside [ms+0.25s, ms+2s] inside [to, to+0.075s]
filter g; the approximate signal A2jþ1 s and detail signal
D2jþ1 s can be further written as following: k=1, 1<l<N
X
þN
A2jþ1 s ¼ hð2n  kÞ A2j s; ð3Þ
k¼N
yes no
0.1s<hs(k)-to(l)<2.5s k=k+1
X
þN
D2jþ1 s ¼ gð2n  kÞA2j s: ð4Þ
k¼N save hs(k), to(l) delete hs(k)
Furthermore the signal can be reconstructed from
approximate and detail signal using the following
yes no
transform: k=N
X
þN
A2j s ¼ 2 *  2kÞA2jþ1 s
hðn Temporal parameters (stance, double support)
k¼N
Fig. 4. Algorithm flowchart for temporal parameters estimation from
X
þN
þ2 *  2kÞD2jþ1 s;
gðn ð5Þ shank angular rate signal (s).
k¼N

where They will only be used as reference to select the intervals

in which negative peaks reminiscent of heelstrikes and
* ¼ hðnÞ;
hðnÞ ð6Þ toe-offs are to be found. So we next looked for local
minima inside intervals ½ms þ 0:25s; ms þ 2s and
* ¼ gðnÞ:
gðnÞ ð7Þ
½ms22s; ms20:05s; assuming that a gait cycle is always
The coefficients of h and g filter are associated with the less than 2 s. The result was two different series of
shape of wavelet considered for the analysis. In this samples hs (heelstrikes) and to (toe-offs). The toe-off
study, a decomposition into 10 scales with the Coiflet detection was refined by finding the local minima inside
wavelet of order 5 has been used. of the original signal s in the intervals [to; to þ 0:075s].
The flowchart for the estimation of toe-off and heel- Finally, at each cycle, the actual heelstrike and toe-off
strikes are shown in Fig. 4. For clarity, let us consider must be so that their difference is considered reasonable
only one shank. It is obvious that the algorithm will be for walking. Thus, for each hsðkÞ only the nearest toðlÞ
applied in the same way for both shanks. First an verifying the condition 0:1sohsðkÞ  toðlÞo2:5s was
approximation
P9 of sðnÞ corresponding to the sa ¼ saved. If no toðlÞ met this condition, hðkÞ was deleted
D
j¼1 2 j s was obtained. With this approximation the and the next heelstrike was considered. Once we are in
original signal was only considered in the range of 0.14– possession of hsL (left heelstrike), toL (left toe-off), hsR
36 Hz. Drift and high frequency movement artifacts (right heelstrike) and toR (right toe-off), every temporal
were canceled in this way. Then two new approxima- parameter of a gait cycle can be easily computed as
tions were obtained. The first one enhanced the percentage of gait cycle. These parameters are:
heelstrike component ðA21 sa  A29 sa Þ and the second
one enhanced the toe-off component ðA23 sa  A29 sa Þ:
* Duration of each gait cycle (measured at right or left
For each of these approximations, the time correspond- leg)
ing to the global maximum values of the signal ðmsÞ RGCTðkÞ ¼ toR ðk þ 1Þ  toR ðkÞ or
were detected. These values corresponded roughly to the LGCTðkÞ ¼ toL ðk þ 1Þ  toL ðkÞ; 1pkpN: ð8Þ
time of midswing during a gait cycle (see Fig. 5),
however their exact significance is not of interest to us. N: number of gait cycles
 K. Aminian et al. / Journal of Biomechanics 35 (2002) 689–699 693

400
tR
deg/s
200

-200 mid-swing (ms)

400
sR
deg/s
200

-200

FSRheel

FSRtoe

0 1 2 3 4 5 Time,s
Fig. 5. FSR signal from heel (FSRheel ) and toe (FSRtoe ), angular rate signal at right shank ðsR Þ and right thigh ðtR Þ: Heelstrike (J) and toe-off (&)
detected by FSR and sR are shown on each signal during 5 typical gait cycles. ms correspond approximately to the time of mid-swing.

* Left stance (time between left heelstrike and left toe- A

d3 A' d3
off)
stance
LSðkÞ ¼ toL ðkÞ  hsL ðkÞ: ð9Þ right left
right
TOR swing swing
right right
d1 d2
* Right stance (time between right heelstrike and right HSR
stride length
toe-off)
(a)
RSðkÞ ¼ toR ðkÞ  hsR ðkÞ: ð10Þ Swing Stance
d = d3
* Initial double support (time between right heelstrike α
and left toe-off), known also as left double thrust l1 l1 M1
β l1 l1
support time (Wall et al., 1981) M2
ϕ d' γ
d'
IDSðkÞ ¼ toL ðkÞ  hsR ðkÞ: ð11Þ ϕ γ
l2 l2 l2 l2
* Terminal double support (time between left heelstrike M2 β M1
d = d1+d2
and right toe-off), known also as right double thrust α
support time (Wall et al., 1981)
(b)
TDSðkÞ ¼ toR ðk þ 1Þ  hsL ðkÞ: ð12Þ
Fig. 6. (a) Body segments model during a gait cycle started by right
toe-off (toR ) and ended by right heelstrike (hsR ). (b) Shank and thigh
* Double support configuration during stance and swing phase. Thigh and shank length
(l1 and l2 ; respectively) are measured on each subject. a and b obtained
DSðkÞ ¼ IDSðkÞ þ TDSðkÞ: ð13Þ from angular rate signals (tR and sR ) allow to estimate the distance d
during stance and swing phases.
In order to reach a good precision for the temporal
parameters, at least 20 gait cycles have been considered
to analyze the average difference between estimated data
for analysis ðN > 20Þ: For each trial, the value of
by gyroscope and actual data obtained by FSR.
temporal parameters was averaged over the N cycles.
Statistical analysis is then performed to find the
significance and accuracy of the parameters obtained sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P 2
by gyroscopes in comparison with the FSR-based number of samples ðestimated  actualÞ
RMSE ¼ : ð14Þ
system. The root mean square error (RMSE) was used number of samples
694 K. Aminian et al. / Journal of Biomechanics 35 (2002) 689–699

2.5. Gait model: spatial parameters estimation right shank ðsR Þ as follows:
Z hsR ðkÞ
We propose a double segment gait model involving aðkÞ ¼ tR ðtÞ dt;
both shank and thigh. In this model, the swing phase is toR ðkÞ
considered as a double pendulum model, while the Z hsR ðkÞ
stance phase is considered as an inverse double bðkÞ ¼ sR ðtÞ dt: ð20Þ
pendulum model. Consequently, knowledge of stance toR ðkÞ
and swing phases obtained from the method described During the stance phase we have

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
dðkÞ ¼ d3 ðkÞ ¼ ðl1 þ M2 ðkÞÞ2 þ ðl1 þ M1 ðkÞÞ2  ðl1 þ M2 ðkÞÞðl1 þ M1 ðkÞÞ cos aðkÞ; ð21Þ

in Section 2.4 is necessary. Fig. 6(a) illustrates the

rotation of the shanks and thighs during a complete where:
gait cycle. The distance d1 þ d2 corresponds to the part sin jðkÞ 0
of stride length performing by the swing of right shank M1 ðkÞ ¼ d ðkÞ; ð22Þ
sin bðkÞ
and right thigh. In addition, during this right swing
phase, the body has moved forward by the distance d3 sin gðkÞ 0
M2 ðkÞ ¼ d ðkÞ ð23Þ
involved by the rotation of left shank and left thigh sin bðkÞ
throughout the contralateral stance phase. Assuming the
same length for left and right steps, one can consider with :
that the rotation during left stance phase is equal to p  bðkÞ
gðkÞ ¼
rotation during right stance phase. Stride length (SL) 2
can then be measured as follows: and
ð24Þ
SL ¼ d1 þ d2 þ d3 : ð15Þ p  2aðkÞ þ bðkÞ
jðkÞ ¼
2
By considering a and b to be the angular rotation of
and
the right thigh and right shank, respectively, one can pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
calculate the distance d ¼ d1 þ d2 from the trigono- d 0 ðkÞ ¼ l2 ðkÞ 2ð1  cos bðkÞÞ
metric relations (Fig. 6b).
During the swing phase we have for each gait cycle k:

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
dðkÞ ¼ d1 ðkÞ þ d2 ðkÞ ¼ ðl2 þ M1 ðkÞÞ2 þ ðl2 þ M2 ðkÞÞ2  ðl2 þ M1 ðkÞÞðl2 þ M2 ðkÞÞ cos bðkÞ; ð16Þ

where, l1 is the thigh length, l2 the shank length and

R toR ðkÞ
sin jðkÞ 0 aðkÞ ¼ hsR ðkÞ tR ðtÞ dt;
M1 ðkÞ ¼ d ðkÞ; ð17Þ R toR ðkÞ ð25Þ
sin bðkÞ
bðkÞ ¼ hsR ðkÞ sR ðtÞ dt:
sin gðkÞ 0 For each gait cycle k; the stride length (SL) and stride
M2 ðkÞ ¼ d ðkÞ ð18Þ
sin bðkÞ velocity (SV) were obtained as follows:
SLðkÞ ¼ d1 ðkÞ þ d2 ðkÞ þ d3 ðkÞ; ð26Þ
with :
p  aðkÞ SLðkÞ
gðkÞ ¼ SVðkÞ ¼ ; ð27Þ
2 RGCTðkÞ
and where RGCTðkÞ was obtained from (8). In order to
p  2bðkÞ þ aðkÞ evaluate the accuracy of stride length and velocity given
jðkÞ ¼
2 by (26) and (27), the following measurements were
and carried out. For treadmill walking, the value of treadmill
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi speed has been considered as the actual mean walking
d 0 ðkÞ ¼ l1 2ð1  cos aðkÞÞ: ð19Þ
velocity ðVa Þ: A preliminary experiment in which various
The angles a and b were estimated by integration of treadmill speeds were compared to the speed of an
the angular rate rotations of the right thigh ðtR Þ and odometer placed on the treadmill belt confirmed that the
 K. Aminian et al. / Journal of Biomechanics 35 (2002) 689–699 695

treadmill speed is accurate. During ground (or ‘‘non- p > 0:97). The RMSE between RGCTm and GCTa was
treadmill’’) walking, the mean velocity was evaluated by 8 ms (0.5%). The 95% confidence interval for the
considering the time needed to cover 20 m using a difference between heelstrike detection by FSR and
stopwatch. The actual mean stride length ðLa Þ was gyroscope was [7 ms, 13 ms]. Heelstrike detection based
calculated as on the gyroscopes occurred later (10 ms in average) than
La ¼ Va GCTa ; ð28Þ FSR (underestimation). The 95% confidence interval for
the difference between toe-off detection by FSR and the
where GCTa is the mean gait cycle time obtained from gyroscopes was [5 ms, 4 ms]. Therefore there was no
footswitches. The accuracy of stride length and velocity significant difference in toe-off detection. The RMSE
was evaluated by comparing La and Va with the mean of between LGCTm and RGCTm was 15 ms (r > 0:99;
SL and SV (SLm and SVm ) over 20 m. 1.3%), which shows the validity of the method in
providing the same time for both left and right gait
cycles.
3. Results The mean values of right and left stance (RSm and
LSm ) as well as double supports (TDSm ; IDSm ; DSm )
3.1. Temporal parameters estimated by the gyroscopes are shown in Table 1. In
general, higher values are observed for elderly subjects.
The correspondence between temporal events de-
tected by FSR and the gyroscope pattern is shown in 3.2. Spatial parameters
Fig. 5 for five consecutive gait cycles. Using the
algorithm of Fig. 4, left and right heelstrikes (respec- For each trial the mean value of stride length and
tively toe-off) were detected from shank gyroscopes velocity (SLm and SVm ) were estimated by finding the
(Fig. 7). Then based on relation (8) to (13), the mean average of stride length (SL) and stride velocity SV
value (over the N cycles) of left and right gait cycle time (relation 26 and 27) over the 20 m of walking. Fig. 9
(LGCTm ; RGCTm ), left and right stance (LSm and compares actual La and Va to SLm and SVm : No
RSm ), initial and terminal double stance (IDSm ; TDSm ; significant change was found between La and SLm
DSm ) were computed for each trial on treadmill and (N=56, p > 0:52) and neither between Va and SVm ðp >
ground. Simultaneously, the periods of the actual gait 0:66Þ: The coefficient of correlation is high in both cases
cycle (GCTa ) and right stance (RSa ) were detected from ðr > 0:96Þ: Actual velocity and stride length can be
FSRheel and FSRtoe : estimated from the measured values as follows:
Fig. 8 compares stance period (RSm and RSa ) and gait
cycle times values (RGCTm and GCTa ) obtained from Va ¼ 1:02SVm  0:04; ð29Þ
the gyroscopes to those of FSR from a trial performed
by the young subjects (treadmill and ground and elderly La ¼ 0:97SLm : ð30Þ
during overground walking. A good agreement ðr >
0:99Þ was found. The difference between RSm and RSa The RMSE for velocity estimation is 0.06 m/s (6.7%)
was not significant (N ¼ 58; p > 0:80) with a mean and that of stride length estimation is 0.07 m (7.2%).
standard error RMSE=23 ms. No significant change Stride length and velocity had lower averaged values
was observed between RGCTm and GCTa (N ¼ 58; for the elderly compared to young subjects.

Angular Rate
deg/s

200
Shank Left 0
-200 LS

IDS TDS
RS

200
Shank Right 0

-200

0 1 2 3 4 5 Time, s

Fig. 7. Temporal parameters estimation from heelstrikes (J) and toe-offs (&) of each leg based on relations (8) to (13) in the text.
696 K. Aminian et al. / Journal of Biomechanics 35 (2002) 689–699

RSm RGCTm
s s
1.6
1.0

0.8 1.2

0.6
0.8
(a) 0.6 0.8 1.0 RSa , s (b) 0.8 1.2 1.6 GCT , s
a

Fig. 8. (a) Right stance duration, in s, obtained from gyroscope (RSm ) compared with that obtained from pressure sensors (RSa ), r > 0:99: (b) Gait
cycle time from gyroscope (RGCTm ) compared to actual gait cycle time obtained from FRS (GCTa ), r > 0:99: ( ) young, (J) elderly subjects.

Table 1
The averaged values of right stance (RSm ), left stance (LSm ), initial double stance (IDSm ), terminal double stance (TDSm ) and double stance (DSm ) in
the percentage of gait cycle obtained for young and elderly subjects

Subjects RSm (%) LSm (%) IDSm (%) TDSm (%) DSm (%)

Young 6172 6272 1272 1272 2373

Elderly 6473 6572 1473 1572 2974

Fig. 10 shows typical results obtained during 2 min of double support (Rose and Gamble, 1994; Maki, 1997).
treadmill walking where the velocity was fixed at 1.11 m/ Few studies have been reported about temporal gait
s (4 km/h). The subject was asked to walk at his normal parameters’ detection based on gyroscope. Pappas et al.
pace, to decrease his pace then increase his pace. All (1999) used a gyroscope on the heel in conjunction with
parameters were obtained at each gait cycle from FSR to detect the precise time of heel-off. Tong and
gyroscopes. As it is illustrated, the estimated velocity Granat (1999) used a gyroscope on the shank and thigh.
(SV) is almost constant (SVm =1.14 m/s) with a coeffi- They found that the pattern on the shank has two
cient of variation of 4%. The change of gait cycle minima, one occurs at foot flat and the other occurs at
(RGCT) provides a corresponding change of stride toe-off. This last result is consistent with ours. In
length (SL) in order to maintain the velocity constant. contrast to their results, we found that the first peak
The corresponding coefficient of variation of SL and after wavelet enhancement corresponds to heelstrike and
RGCT is 15%. During the same period there is no not foot flat. It must be noted that only 2 subjects were
significant variation of left and right stance (coefficient involved in their study and the sampling rate was 50 Hz.
of variation less than 3%) while the double support time Toe-off and heelstrike detection were based on the
(DS) has a coefficient of variation of 16%. choice of a threshold on the FSR signal. It is clear that
the value of the threshold can modify the results.
Moreover, only a few steps were considered for analysis
4. Discussion which cannot be considered as statistically significant.
The method presented in this study compares
We have proposed a new algorithm based on wavelet favorably with other methods used to identify gait
transformation to detect toe-off and heelstrike from events that require more complex instrumentation and a
shank angular velocity. A high correlation with actual laboratory setting. Hreljac and Marshall (2000) used
gait events detected by foot pressure sensors was found. four motion analysis cameras (60 Hz), involving 12 trials
In addition, no significant error was observed for toe-off of two healthy subjects and found an absolute average
detection, while a slight systematic delay (10 ms in error of 4.7 and 5.6 ms for heelstrike and toe-off,
average) existed between heelstrike detected by gyro- respectively, considering information from a forceplate
scopes compared to FSR. Considering a sampling as the ‘‘gold standard’’. Stanhope et al. (1990) using a
period of 5 ms (200 Hz), this error is acceptable. The 50 Hz camera collection system, reported an error
obtained values correspond to the established values for greater than 20 ms in over 20% of the cases. Using
young and elderly populations which is around 60% for accelerometers on the thigh and considering FSR as
the duration of stance and 20% for the duration of reference, Aminian et al. (1999) found the 95%
 K. Aminian et al. / Journal of Biomechanics 35 (2002) 689–699 697

La Va
m m/s
La=0.97.SLm 1.6 Va=1.02.SVm-0.04
r=0.96 r=0.98
1.6

1.2
1.2

0.8 0.8

0.8 1.2 1.6 0.8 1.2 1.6

SLm, m SVm, m/s

Fig. 9. Correspondence between actual stride length ðLa Þ and velocity ðVa Þ and estimated values from shank and thigh gyroscopes (SLm ; SVm ) for
young ( ) and elderly subjects (,).

1.5
m/s
SV 1
normal increase cadence decrease cadence

2
m
SL 1.5

0.5

2
s
1.5
RGCT

0.5

60
%
RS, LS
20
0

40
%

DS 20

0
0 20 40 60 80 100 120
Gait Cycles

Fig. 10. Stride velocity (SV), stride length (SL), gait cycle time (RGCT), right and left stance (RS and LS) and double support (DS) in the percent of
gait cycle obtained from a young subject during 2 min of treadmill walking using the proposed method. The velocity of treadmill was kept at 1.11 m/s
while the subject changed his cadence. All parameters are estimated at each gait cycle allowing to analyse the stride to stride variability.

confidence interval of stance period equal to [20 ms, Gait events can also be explained by looking at the
20 ms]. In the present study, which involve young and particular pattern of the shank and thigh angular
elderly subjects the error of toe-off and heelstrike velocities (Fig. 5). Shank angular velocity (signal sR ) is
detection is less that 5 ms (95%), with a systematic a bipolar signal. Considering as positive the counter-
delay for heelstrike time of 10 ms. clockwise rotation, one can observe that during the
698 K. Aminian et al. / Journal of Biomechanics 35 (2002) 689–699

swing period sR is rather positive while in stance phase gyroscope on the left thigh. However, few of our elderly
sR is almost negative. Just prior to heelstrike, sR subjects had slightly asymmetric gait and we did not
becomes negative. At heelstrike, there is a flexion observe significant variation in the performance of the
(Vaughan et al., 1992) of the knee shown by a negative system. There are many advantages to use gyroscopes
sR and positive tR : The time of heelstrike is also instead of other kinematic sensors such as acceler-
characterized by a change of slope sign of sR : This is ometers. First, unlike the accelerometer, the gyroscope
expected since before heelstrike, the shank decelerates can be attached to anywhere of any body segment as
(negative slope) to stop and after heelstrike there is an long as its axis is parallel to the mediolateral axis: the
acceleration (positive slope) due to rotation around the angular rotation is still the same along this segment.
ankle. After heelstrike, there is a clockwise rotation of Tong and Granat (1999) have shown that the signals
the shank around the ankle involving a negative sR : In from different gyroscopes at different attachment site
mid-stance, the knee becomes straight leading to a slight are almost identical. Second, the angular rate signal is
angular velocity, which could become positive. Then the less noisy than acceleration since acceleration is the
shank continues its clockwise rotation around the ankle derivative of velocity and involves higher frequency
and accelerates its rotation at the toe-off (positive slope components. Third, accelerometry provides essentially
of sR ). The rotation is still clockwise (sR negative) at the temporal gait parameters, using singular peak enhance-
beginning of the swing phase until the knee reaches its ment. In contrast, rotation angles can be estimated from
maximum flexion (tR is positive and sR negative), then angular velocity by simple integration, and provides
the knee comes into extension (sR positive and tR spatial gait parameters in addition to temporal para-
positive) moving the shank from backward to forward. meters. Recently, the possibility to estimate knee angle
In this way, shank and thigh angular velocities also give (Williamson and Andrews, 2001) and trunk tilt (Najafi
remarkable insight into the recognition of the periods of et al., 2001a) has also been shown. Finally, there is no
knee flexion and extension during gait cycle. influence of gravity acceleration on the measured signal.
In order to measure spatial parameters such as stride Gyroscopes however, do have some weaknesses. First,
length, we have used the change of lower limb angles. the piezoelectric gyroscope is more delicate to use than
Miyazaki (1997) has shown that for the estimation of accelerometer. It is more sensitive to temperature and
the stride length from the integral of angular velocity, shock due to the mechanical fastening of vibrating piezo
the relative change of angular rotation is sufficient and beam inside the sensor’s case. In addition, powerful
therefore there is no need of initial angular position. signal processing and filtering are necessary to cancel
However this method consider thigh and shank as a drift and artefact in the signal. Nevertheless, this
unique segment and the error reached 15%. We propose filtering was done in this study by using wavelet
a double segment gait model involving both shank and transformation. The use of wavelet was particularly
thigh, which consider a double pendulum model during important to ensure a good time resolution in finding
the swing phase and the inverse double pendulum model gait events such as toe-off and heelstrike.
during the stance phase. The identification of swing and Overall, the measuring device is light (300 g), portable,
stance phases was therefore necessary, and was deter- low cost and involves no discomfort for the patient who
mined by toe-off and heelstrike estimation. The estima- carries it for a long period of time. The datalogger is
tion error for velocity and stride length is around 7%. battery powered with a lifetime of 20 h. It can collect
This error is principally due to the inaccuracy of the data up to 8 Mbytes corresponding to a period of 2.5 h
gyroscopes which induce some error on a and b in (20) (with three sensors at 200 Hz sampling rate). Additional
and (25). This error could increase if the period of memory cards or lower sampling rate can be used to
integration increases considerably (very slow speed). In increase the period of recording. The off-line calculation
this case a new calibration of (29) and (30) is necessary. of temporal and spatial parameters is very fast: for a
Other factors such as the difficulty to measure the exact walking trial of 2 min, it takes from few seconds to less
shank and thigh length, the misalignment of the than a second, depending on computer’s performance.
gyroscope axis regarding the mediolateral axis and the
error due to gait phase detection are also to be
considered. The low values of the intercept (close to 0) 5. Conclusion
and the slope close to 1 in (29) and (30) confirm the
appropriateness of the proposed gait model. The proposed method appears a promising monitor-
In this study, the stride length and velocity were ing tool for several purposes. First, it allows measure-
estimated based on the right leg’s angular velocity and ments of gait features during a long period of walking
assuming that in straight walking, right and left stride and thus supplies the stride-to-stride variability of gait
length are equal. If this assumption is no longer valid as illustrated in Fig. 10. In addition, because of its
(e.g. curled trajectory or asymmetric gait) right and left portability, this system can be used in other settings than
leg rotation can be estimated separately by adding a 4th a gait laboratory and therefore provides information
 K. Aminian et al. / Journal of Biomechanics 35 (2002) 689–699 699

that is more likely to reflect the actual performance of Maki, B.E., 1997. Gait changes in older adults: predictors of falls or
the subjects. indicators of fear. Journal of the American Geriatrics Society 453,
It can be used in many clinical applications such as 313–320.
Mallat, S.G., 1989. A theory for multi-resolution signal decomposition
outcome evaluation after knee and hip replacement, or (the wavelet representation). IEEE Transactions on Pattern
external prosthesis adjustment for amputees. In elderly Analysis and Machine Intelligence 11, 674–693.
subjects, this system can also be proposed as a Miyazaki, S., 1997. Long-term unrestrained measurement of stride
diagnostic tool for abnormal gait analysis, as a predictor length and walking velocity utilizing a piezoelectric gyroscope.
tool for fall risk estimation (Najafi et al., 2001b), or as a IEEE Transactions on Biomedical Engineering 44, 8.
Najafi, B., Aminian, K., Paraschiv-Ionescu, A., Loew, F., Blanc, Y.,
monitoring tool to assess progress through rehabilita- Robert, Ph., 2001a. Body postures and walking period estimation
tion. During this study the system proved to have a high using a kinematic sensor: application for long term monitoring of
acceptability by elderly subjects. physical activity in elderly, ESMAC-2001. Gait and Posture 14,
119–120.
Najafi, B., Aminian, K., Bula, C., Ruggieri, G., Robert, Ph., 2001b.
Falling risk evaluation in elderly using miniature gyroscope:
Acknowledgements relation between gait and risk of falling. In: Duysens, J., Smits-
Engelsman, B.C.M., Kingma, H. (Eds.), Control of Posture and
The authors wish to acknowledge the contribution of Gait. Symposium of the International Society for Postural and
M.J. Gramiger and M.P. Morel through the design of Gait Research (ISPG 2001), Maastricht, pp. 135–139.
Physilog system and sensors. Nene, A., Mayagoitia, R., Veltink, P., 1999. Assessment of rectus
femoris function during initial swing phase. Gait and Posture 9, 1–
9.
Pappas, I.P., Keller, T., Popovic, M.R., 1999. A novel gait phase
References detection system. In: Proceeding of Automed’99, 2. Workshop,
‘‘Automatisierungstechnische Verfahren der Medizin’’, TU Darm-
Abu-Faraj, Z.O., Harris, G.F., Abler, J.H., Wertsch, J.J., 1997. A stadt, pp. 69–70.
holter-type, microprocessor-based, rehabilitation instrument for Rose, J., Gamble, J., 1994, Human Walking. Williams & Wilkins,
acquisition and storage of plantar pressure data. Journal of Baltimore, pp. 29, 143.
Rehabilitation Research and Development 34, 187–194. Sparks, D.R., Huang, X., Higdon, W., Johnson, J.D., 1998. Angular
Aminian, K., Robert, Ph., Jequier, E., Schutz, Y., 1995. Incline, speed, rate sensor and accelerometer combined on the same microma-
and distance assessment during unconstrained walking. Medicine chined CMOS chip. In: Microsystem Technologies 4. Springer,
and Science in Sports and Exercise 27, 226–234. Berlin, pp. 139–142.
Aminian, K., Rezakhanlou, K., De Andres, E., Fritsch, C., Leyvraz, Stanhope, S.J., Kepple, T.M., McGuire, D.A., Roman, N.L., 1990.
P.-F., Robert, Ph., 1999. Temporal features estimation during Kinematic-based technique for event time determination during
walking using miniature accelerometers: an analysis of gait gait. Medical and Biological Engineering and Computing 28,
improvement after hip arthroplasty. Medical and Biological 355–360.
Engineering and Computing 37, 686–691. Tong, K., Granat, M.H., 1999. A practical gait analysis system using
Czerniecki, J.M, Gitter, A.J., 1996. Gait analysis in the amputee: has it gyroscopes. Medical Engineering and Physics 21, 87–94.
helped the amputee or contributed to the development of improved Vaughan, C.L., Davis, B.L., O’Coonor, J.C., 1992. Dynamics of
prosthetic components? Gait and Posture 4, 258–268. Human Gait. Human Kinetics Publishers, Champaign, IL, pp.
Evans, A.L., Duncan, G., Gilchrist, W., 1991. Recording accelerations 33–35.
in body movements. Medical and Biological Engineering and Williamson, R., Andrews, B.J., 2001. Detecting absolute human knee
Computing 29, 102–104. angle and angular velocity using accelerometers and rate gyro-
Hirokawa, S., Matsumara, K., 1987. Gait analysis using a measuring scopes. Medical and Biological Engineering and Computing 39,
walkway for temporal and distance factors. Medical and Biological 294–302.
Engineering and Computing 25, 577–582. Wall, J.C., Ashburn, A., Klenerman, L., 1981. Gait analysis in the
Hreljac, A., Marshall, R.N, 2000. Algorithms to determine event assessment of functional performance before and after total hip
timing during normal walking using kinematic data. Journal of replacement. Journal of Biomedical Engineering 3, 121–127.
Biomechanics 33, 783–786. Zhu, H.S., Wertsch, J.J., Harris, G.F., Loftsgaarden, J.D., Price, M.B.,
Ismail, A.R., Asfour, S.S., 1999. Discrete wavelet transform: a tool in 1991. Foot pressure distribution during walking and shuffling.
smoothing kinematic data. Journal of Biomechanics 32, 317–321. Archives of Physical Medicine and Rehabilitation 72, 390–397.