A WEARABLE GAIT ANALYSIS SYSTEM USING INERTIAL

SENSORS PART I

Evaluation of Measures of Gait Symmetry and Normality against 3D Kinematic

Data

A. Sant’Anna

, N. Wickstr

R. Z

ugner

and R. Tranberg

Intelligent Systems Lab, Halmstad University, Halmstad, Sweden

Department of Orthopedics, Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden

Keywords:

Gait Analysis, Inertial Sensors, Symmetry, Normality.

Abstract:

Gait analysis (GA) is an important tool in the assessment of several physical and cognitive conditions. The lack

of simple and economically viable quantitative GA systems has hindered the routine clinical use of GA in many

areas. As a result, patients may be receiving sub-optimal treatment. The present study introduces and evaluates

measures of gait symmetry and gait normality calculated from inertial sensor data. These indices support the

creation of mobile, cheap and easy to use quantitative GA systems. The proposed method was compared to

measures of symmetry and normality derived from 3D kinematic data. Results show that the proposed method

is well correlated to the kinematic analysis in both symmetry (r=0.84, p<0.0001) and normality (r=0.81,

p<0.0001). In addition, the proposed indices can be used to classify normal from abnormal gait.

1 INTRODUCTION

Quantitative gait analysis (GA) can improve the as-

sessment of a number of physical and cognitive con-

ditions. The importance of GA in the treatment of

children with cerebral palsy is well known and doc-

umented, e.g. (Chang et al., 2010), (DeLuca et al.,

1997). The use of GA to monitor and assess Parkin-

son’s Disease, e.g. (Salarian et al., 2004), (Frenkel-

Toledo et al., 2005), and stroke, e.g. (Cruz et al.,

2008), (Silver et al., 2000), have also been investi-

gated. Changes in gait speed, gait variability, and neu-

rologic gait abnormalities have been associated with

the risk of developing dementia and mild cognitive

impairment, e.g. (Beauchet et al., 2008), (Verghese

et al., 2002).

Although the usefulness of GA is recognized by

the medical community, e.g. (Chang et al., 2010),

routine clinical use of GA is still not a reality. This

is likely due to the costs involved in performing a full

3D GA at a gait lab. As a result of not undergoing GA,

many patients may receive sub-optimal treatment, e.g.

(Kay et al., 2000), (Lofterød and Terjesen, 2008).

The simpler alternative to in-lab 3D GA is ob-

servational GA (OGA), such as the Gillette Func-

tional Assessment Questionnaire (GFAQ) Walking

Scale (Novacheck et al., 2000) and the Edinburgh Gait

Score (Read et al., 2003). Although some OGA meth-

ods have been shown valid and reliable, it is generally

understood that they are speciﬁc to patient groups,

subjective, and sensitive to the observer’s experience

(Toro et al., 2003). In 1999, Coutts (Coutts, 1999) ar-

gued that despite its limitations, OGA would never be

totally replaced as the default GA method in the clini-

cal environment because of ease of use. Current tech-

nological advancements, however, should encourage

clinicians to re-evaluate instrumented GA.

The goal of the present study is to develop a mo-

bile, cheap, and easy to use GA system that quantita-

tively evaluates certain characteristics of gait indepen-

dent of location and/or infrastructure. Such a system

may be complementary to 3D GA, by providing con-

tinuous or frequent monitoring. Alternatively, it may

be used where 3D GA is not available such as in un-

derprivileged areas or at home. The system may also

be coupled to OGA, providing consistent and reliable

quantitative data to aid clinical evaluation.

The proposed method uses accelerometer and gy-

roscope data to derive measures of gait symmetry

and gait normality. The signal analysis is based on

the symbolic approach presented in (Sant’Anna and

Wickstr

om, 2010), and (Sant’Anna et al., 2011). 19

180

Sant’Anna A., Wickstrom N., Zügner R. and Tranberg R..

A WEARABLE GAIT ANALYSIS SYSTEM USING INERTIAL SENSORS PART I - Evaluation of Measures of Gait Symmetry and Normality against 3D

Kinematic Data.

DOI: 10.5220/0003707601800188

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2012), pages 180-188

ISBN: 978-989-8425-89-8

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

healthy subjects were measured simultaneously with

inertial sensors and a 3D motion capture (MOCAP)

system while walking in different ways. Results from

the inertial system are compared to measures of sym-

metry and normality derived from 3D kinematic data.

2 RELATED WORK

2.1 Symmetry

Symmetry refers to the similarity between the move-

ments of the right and left sides of the body. Gener-

ally speaking, gait symmetry can be computed from

discrete values, e.g. spatio-temporal parameters; or

from continuous signals, e.g. joint angles. Some au-

thors have argued that discrete values are not always

sufﬁcient to describe gait asymmetry, and that it is im-

portant to take into account continuous motion data

(Crenshaw and Richards, 2006). It is also important

to distinguish between different sources of data. In

the present study we will focus on kinematic data ex-

tracted from a MOCAP system; and accelerometer

and gyroscope data obtained via wearable sensors.

Some approaches to calculating symmetry using

continuous accelerometer data have been introduced.

(Moe-Nilssen and Helbostad, 2004), for example, in-

troduced an unbiased autocorrelation method using

trunk acceleration data. Although this may provide

a good general estimate of gait symmetry, it lacks in-

formation about each individual limb. More recently

(Gouwanda and Senanayake, 2011) used gyroscopes

on shanks and thighs to calculate symmetry using a

normalized cross correlation approach and the nor-

malized mean error between curves derived from right

and left sides. This method segments and normal-

izes the data to individual strides. As a result, only

the shape of the signal and not its relative temporal

characteristics are taken into account. Sant’Anna et

al. also suggested a symbolic method for estimating

gait symmetry using accelerometers (Sant’Anna and

Wickstr

om, 2010) or gyroscopes (Sant’Anna et al.,

2011), which takes into account not only the shape

but also the temporal characteristics of the signal.

Kinematic gait data is usually evaluated by visual

inspection of superimposed curves from right and left

sides. Few symmetry measures have been proposed

which take into account complete joint angle curves.

(Crenshaw and Richards, 2006) calculated a measure

of trend symmetry based on the variance around the

principal component of a right-side vs. left-side

plot. This trend symmetry measure is insensitive to

scaling, and an additional measure, the range ampli-

tude ratio, is required. The present study introduces

a symmetry measure based on kinematic data which

can be expressed as one index.

2.2 Normality

Normality refers to the similarity between the move-

ments of one individual compared to average move-

ments of a population that is judged healthy/normal.

The Gillette Functional Assessment Questionnaire

(GFAQ) Walking Scale is a widely accepted gait nor-

mality measure based on observation. Considerable

efforts have been put into deriving a similar measure

from kinematic data. (Schutte et al., 2000) used prin-

cipal component analysis (PCA) on 16 discrete nor-

mal gait variables to create a representation of the data

in a different space. The magnitude of the projec-

tion of an abnormal data set onto this space is used

as a normality index, known as the Gillette Gait In-

dex (GGI). (Shin et al., 2010) used the same PCA ap-

proach to create three separate indices using variables

related to ankle, knee and hip kinematics.

A very similar PCA approach, the Gait Deviation

Index (GDI), was introduced by (Schwartz and Rozu-

malski, 2008) using complete joint angle curves. This

index showed high correlation with GGI, and distin-

guished between different levels of the GFAQ. One

advantage of PCA approaches is that they transform

the possibly dependent gait variables into a new set

of independent variables. The disadvantage is that re-

sults cannot be traced back to the original gait vari-

ables.

(Barton et al., 2007) used self organizing maps

(SOM) to create a single representation from many

kinetic and kinematic curves. This representation is

then used to calculate a measure of distance between

abnormal and normal data sets. This approach was

later developed into a user friendly graphical user in-

terface that provides a deviation curve for each sub-

ject (Barton et al., 2010). The mean value of this de-

viation curve was highly correlated with the GDI and

showed signiﬁcant difference between different lev-

els of the GFAQ. The difﬁculty in using this method

derives from the fact that large amounts of normal ref-

erence data are needed to train the SOM.

A much simpler method, the Gait Proﬁle Score

(GPS) and Movement Analysis Proﬁle (MAP), was

introduced by (Baker et al., 2009). The MAP is cre-

ated by taking the root mean square error (RMS) be-

tween a reference joint angle curve and the corre-

sponding curve from a subject. This creates one nor-

mality index for each joint angle curve. A unique

index, the GPI, can be derived by concatenating all

joint angle curves end to end, and taking the RMS of

this aggregated curve. This work concluded that the

A WEARABLE GAIT ANALYSIS SYSTEM USING INERTIAL SENSORS PART I - Evaluation of Measures of Gait

Symmetry and Normality against 3D Kinematic Data

181

GPI and GDI are alternative and closely related mea-

sures. Although GDI presents some nice properties

such as normal distributions across GFAQ levels, GPS

is more easily interpreted because the original vari-

ables suffer no transformations and results are given

in degrees. (Beynon et al., 2010) also concluded that

GPS is signiﬁcantly correlated with clinical judgment.

No normality indices based on accelerometer or

gyroscope data were found in the literature.

3 METHOD

3.1 Data collection

A group of 19 healthy individuals willing to partici-

pate in the experiment were randomly selected. The

average hight of the group was 172.1 ± 7.6 cm; and

the average weight was 71.8 ± 17.2 Kg. Seven par-

ticipants were male and twelve female, averaging an

age of 34 ± 13 years.

Kinematic and kinetic data were recorded with a

3D motion capture (MOCAP) system, Qualisys MCU

240, sampling at 240Hz. A total of 15 spherical re-

ﬂective markers, of 19 mm in diameter, were attached

to the skin with double-sided tape. Markers were

placed on the sacrum, anterior superior iliac spine, lat-

eral knee-joint line, proximal to the superior border of

the patella, tibial tubercle, heel, lateral malleolus and

between the second and third metatarsals (Tranberg

et al., 2011).

Subjects were also equipped with 3 Shimmer



sensor nodes containing one 3-axis accelerometer and

one 3-axis gyroscope, sampling at 128Hz. The sen-

sor nodes were attached to the skin with double-sided

tape. One node was placed on each outer shank, about

3cm above the lateral malleolus marker, Figure 1(a).

The remaining node was placed mid-way between the

anterior superior iliac spine markers, Figure 1(b). In

addition, one reﬂective marker was also placed on

each sensor node.

(a) Shank sensor node (b) Waist sensor node

Figure 1: Placement of sensor nodes. Shank sensor node

approximately 3cm above the lateral malleolus reﬂective

marker, and waist sensor node mid-way between the ante-

rior superior iliac spine reﬂective markers.

Before starting the measurements, each sensor

node received a beacon signal from a host computer

with the host global time in milliseconds since epoch.

At this moment, each node stored in its memory card

its own local time in milliseconds since epoch to-

gether with the host global time. These records were

later used to synchronize the sensor data. The data

from the sensors is stored in the node.

The subjects were then asked to enter into the

measurement volume and a static reference record-

ing was obtained with the Mocap system while the

subjects were standing in an upright position aligned

with the x-axis of the global coordinate system. Prior

to recording, all subjects had the possibility to get fa-

miliarized with the walkway and deﬁne a comfortable

walking speed. The following instructions were then

given to the subjects: 1) walk normally at a comfort-

able speed; 2) walk with a limp, as if injured; and 3)

walk slowly, as if tired or pretending to be old. All

subjects performed three tests for each type of walk.

One test of each type was then randomly chosen for

further analysis.

This study was approved by the Regional Ethics

Board in Gothenburg, Sweden.

3.2 MOCAP Normality Measure

The normality index used for the kinematic data was

the GPS and the MAP (Baker et al., 2009). However,

the mean value was removed from all curves before

calculating the score, and foot progression was not

used because it was not available in the reference data

set. Removing the curves’ mean values makes the

normalcy measure more robust to offset errors, while

preserving the shape and range of the curves.

The reference data set is an ensemble of 34 ran-

domly selected adult subjects presenting no known

pathologies, previously acquired at the clinical gait

lab at Sahlgrenska University Hospital, Gothenburg,

Sweden. Joint angle curves were calculated for each

individual and normalized to stride time. The ensem-

ble average of the normalized curves was used as a

reference curve.

Each MAP component was calculated as the RMS

difference, MAP =

∑

n=1

sub j

(n) −C

re f

(n))

between the reference curve, C

re f

, and the subject’s

curve, C

sub j

, Figure 2, where N is the number of

points in the curve. The GPS was calculated similarly

by concatenating all joint curves end to end.

For each subject, MAP and GPS results were cal-

culated as the average between right side and left side

MAP and GPS respectively.

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

182

0 20 40 60 80 100

-20

-15

-10

-5

stride time (%)

normalized ankle df. angle (degrees)

difference

Subject's data

Reference data

Figure 2: Calculating the MAP. The MAP for this joint

angle progression is calculated as the RMS of the difference

between the subject’s curve and the reference curve, i.e. the

RMS of the shaded area.

3.3 MOCAP Symmetry Measure

Based on the GPS, a measure of symmetry was de-

rived for the kinematic data. In this case, the compo-

nents of MAP-symmetry were calculated as the RMS

error between the curves for the right and left sides,

after removing their corresponding mean values. Sim-

ilarly, GPS-symmetry was calculated by concatenat-

ing all joint curves end to end and calculating the

RMS difference between left and right sides.

3.4 Inertial Sensor Symmetry Measure

The symmetry measure used in this paper was pre-

sented in (Sant’Anna and Wickstr

om, 2010) and also

used in (Sant’Anna et al., 2011), with a different sym-

bolization technique. The sensor signal, accelerom-

eter or gyroscope, is standardized to zero mean and

unitary standard deviation, then segmented into N

symbols. Symbolization is done by quantization into

N levels. The quantization levels are chosen based on

the empirical probability distribution of the signal, so

as to produce equiprobable symbols, Figure 3.

The period between consecutive occurrences of

the same symbol are calculated and stored in a period

histogram (Sant’Anna et al., 2011). Similarly, the pe-

riod between consecutive transitions from symbol i to

symbol j are calculated and stored in a transition his-

togram. The symmetry index is a measure of the sim-

ilarity between symbol (transition) period histograms

for the right and left sides. Histograms are compared

using a relative error measure shown in Eq. 1, where

Z is the number of symbols; K is the number of bins

in the histograms; n

is the number of non-empty his-

togram bins (for either foot) for symbol i; h

(k) is

the normalized value for bin k in the period histogram

i for the right foot; and h

(k) is the normalized value

100 300 500

-1.5

-1

-0.5

0.5

1.5

2.5

time (samples)

standardized signal

A) original signal

0 0.2 0.4 0.6 0.8 1

-0.74

-0.46

-0.30

0.96

B) quantization

cumulative distr.

cut points

100 300 500

time (samples)

symbols

C) symbolized signal

Figure 3: Symbolization. In this example, the signal was

symbolized into 5 symbols. (A) shows the original signal

after standardization. (B) shows the empirical cumulative

distribution of the signal, probability levels are shown on

the horizontal axis. The quantization levels, cut points, are

the values at which the probability of the signal increases

by 1/Number

symbols

= 1/5 = 0.2. (C) illustrates the cor-

responding symbolized signal. Note that the probability of

each symbol occurring in the symbolized signal is 20%.

for bin k in the period histogram i for the left foot.

symb

∑

i=1

∑

k=1

(k) −h

(k)|

∑

i=1

∑

k=1

(k) +h

(k)|

100 (1)

3.5 Inertial Sensor Normality Measure

The normality measure for the inertial sensor data was

derived from the symmetry measure, Part 3.4. In-

stead of comparing the histograms for right and left

sides, one subject’s histograms are compared to his-

tograms derived from a reference data set. The ref-

erence data set was formed by selecting the subjects

that presented the smallest GPS based on the normal

walk kinematic data.

The normal walk inertial sensor data from these

reference subjects was standardized to zero mean and

unitary standard deviation, symbolized, and symbol

(transition) periods were calculated. The symbol

(transition) periods were normalized to stride time.

That is, a period that coincides with stride time is

represented as 1 and all other periods are scaled cor-

respondingly. This normalization is common when

dealing with kinematic data, and it ensures that the

analysis is not affected by gait speed. The symbol

(transition) periods from all reference subjects were

used to create reference histograms.

Reference histograms were compared to the his-

tograms from right and left sides of each subject,

Eq. 2, where h

re f

are the reference histograms and

sub j

are the subject histograms. Right and left results

were averaged to create the normality measure for the

A WEARABLE GAIT ANALYSIS SYSTEM USING INERTIAL SENSORS PART I - Evaluation of Measures of Gait

Symmetry and Normality against 3D Kinematic Data

183

subject.

NORM

symb

∑

i=1

∑

k=1

re f

(k) −h

sub j

(k)|

∑

i=1

∑

k=1

re f

(k) +h

sub j

(k)|

100 (2)

3.6 Stride Time Estimation

The period (transition) histograms can also be used

to estimate gait events (Sant’Anna and Wickstr

om,

2010). In the present study, stride time was calcu-

lated by ﬁrst determining which of the symbols (tran-

sitions) presented periods with standard deviations

smaller or equal to 1/5 of their mean value. This rule

identiﬁes the symbols (transitions) that occur regu-

larly in every cycle. The mean period of the high-

est symbol (transition) was used as an estimate of

the stride time. The highest symbol (transition) was

chosen based on a priori knowledge that the gyro-

scope data presents a prominent spike at mid-swing

(Salarian et al., 2004), and the accelerometer data

presents a prominent spike at heel-strike (Aminian

et al., 1999). As described in (Sant’Anna and Wick-

str

om, 2010), a priori expert knowledge can be used

to identify gait events in the signal, from which vari-

ous temporal parameters may be derived, e.g. double-

support time.

3.7 Analysis

The data acquired from the MOCAP system was pro-

cessed in Visual 3D (C-Motion Inc., Germantown,

MD) to generate kinematic join angle data and spatio-

temporal parameters such as stride time. The data

was then exported to MATLAB (MathWorks, Natick,

MA) where MAP, GPS and MOCAP symmetry were

calculated for each subject and used as reference.

The signals from the shank accelerometers and

gyroscopes were low-pass ﬁltered with a Butterworth

ﬁlter of order 6 and cut-off frequency 20Hz. The waist

sensor data was ﬁltered at 10Hz. The signals were ﬁl-

tered once, then reversed and ﬁltered again to avoid

any phase shift. The three axes of each accelerometer

were combined into a resultant signal, A

res

+ A

For each gyroscope, only pitch and roll rotations were

considered, G

res

pitch

+ G

roll

Symmetry and stride times were calculated using

right and left shank signals. Normality was calcu-

lated using both shanks and waist signals. Measures

were calculated considering both period histograms

and transition histograms, varying from 5 to 25 sym-

bols. The resulting values outside two standard de-

viations were considered outliers and removed. The

remaining values were used to calculate the Spear-

man’s rank correlation coefﬁcient with reference mea-

surements. The optimal number of symbols and the

choice of histogram were chosen so as to maximize

the correlation coefﬁcients. A two-sample t-test was

used to determine if the ﬁnal normality and symmetry

values distinguish between normal and other walks.

All hypothesis were bi-directional with conﬁdence

level, α = 0.05. All sensor data analysis was under-

taken in MATLAB.

4 RESULTS

Some of the data was excluded due marker obstruc-

tion or sensor failure. The total number of subjects

used for each analysis is shown in Table 1.

Table 1: Number of subjects available for analysis.

walk placement accelerometer gyroscope

Normal

shank 16 14

waist 16 15

Slow

shank 13 12

waist 13 12

Limp

shank 16 14

waist 16 15

An example of the shank sensor node data and cor-

responding symbolized data is shown in Figure 4.

1.5 2 2.5 3 3.5 4 4.5

time (s)

acceleration (g)

A) resultant acceleration - right shank

1.5 2 2.5 3 3.5 4 4.5

time (s)

symbols

B) symbolized acceleration

1.5 2 2.5 3 3.5 4 4.5

time (s)

angular velocity (deg/s)

C) resultant angular velocity - right shank

1.5 2 2.5 3 3.5 4 4.5

time (s)

symbols

D) symbolized angular velocity

Figure 4: Example shank inertial sensor data. (A) illus-

trates a typical resultant acceleration signal from a shank

sensor. (B) exempliﬁes the symbolization of the acceler-

ation signal using 5 symbols. (C) illustrates a typical gy-

roscope resultant signal acquired with the same sensor as

(A). (D) exempliﬁes the symbolized gyroscope signal using

5 symbols. Signals (A) and (C) come from a normal walk

trial and are synchronized in time.

Figure 5 shows the MAP and GPS results for all

subjects during normal walking and limp walking.

There is a signiﬁcant difference between the different

walks. The overall amplitude of the scores is reduced

by approximately a factor of 2 compared to results

presented (Baker et al., 2009). This is caused by the

absence of curves’ mean values. Another contribut-

ing factor is that the present work considers healthy

adults instead of children.

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

184

ankle flex knee flex hip flex hip add hip rot pelv tilt pelv obliq pelv rot GPS

RMS error (degrees)

MAP: Normal walk vs. Limp walk

limp left side

limp right side

limp both sides

normal

Figure 5: MOCAP normality. MAP and GPS results for

the limp data set are shown against the normal data set re-

sults.

The best normality results using the shank sensors

were obtained with the accelerometer data, 5 sym-

bols, and symbol period histograms. Statistically sig-

niﬁcant Spearman’s rank correlation coefﬁcients are

shown in Table 2. Note that the normality measure

derived from the shank sensor correlates better with

ankle, knee and hip ﬂexion MAP components than

with the other MAP components. The combination of

all MAP components into the GPS, however, corre-

lates better than most individual components, with the

exception of hip ﬂexion. The normality index derived

from the accelerometer placed on the waist correlated

with the MOCAP reference better than the shank sen-

sor data, Table 3. The optimal combination in this

case was 18 symbols and transition histograms.

Table 2: Correlation of normality measures - shank.

sensor accelerometer

placement shank

histogram symbol period

no. symbols 5

variable r p-value

MAP ankle ﬂex. 0.58 <0.0001

MAP knee ﬂex. 0.73 <0.0001

MAP hip ﬂex. 0.78 <0.0001

MAP pelv. obliq. 0.44 < 0.0001

MAP pelv. rot. 0.49 < 0.0001

GPS 0.74 < 0.0001

Table 3: Correlation of normality measures - waist.

sensor accelerometer

placement waist

histogram transition

no. symbols 18

variable r p-value

MAP ankle ﬂex. 0.69 <0.0001

MAP knee ﬂex. 0.77 <0.0001

MAP hip ﬂex. 0.82 <0.0001

MAP hip add. 0.47 <0.0001

MAP pelv. obliq. 0.49 <0.0001

MAP pelv. rot. 0.71 < 0.0001

GPS 0.81 < 0.0001

The best correlation of the inertial sensor symme-

try with the reference was achieved with the shank gy-

roscopes, 20 symbols, and symbol period histograms.

The best correlation coefﬁcients are shown in Table 4.

Once again, the measure derived from the shank sen-

sor correlates better with ankle, knee and hip ﬂexion

MAP components.

Table 4: Correlation of symmetry measures - shank.

sensor gyroscope

placement shank

histogram symbol period

no. symbols 20

variable r p-value

ankle ﬂex. 0.64 < 0.0001

knee ﬂex. 0.81 < 0.0001

hip ﬂex. 0.68 < 0.0001

all 0.84 < 0.0001

Stride time derived from the symbolized data were

accurate. The correlation with the reference data was

0.97, p<0.0001. After eliminating 2 outliers, the total

RMS error was 0.048 seconds. The best stride time

results were achieved using 17 symbols and transition

histograms.

Figures 6 and 7 show the symmetry and normal-

ity results respectively. Two-sample t-tests indicated

that the MOCAP symmetry data for normal walk was

signiﬁcantly different from limp, p<0.0001, and slow

walk, p=0.03. The gyroscope symmetry index was

also signiﬁcantly different between normal and limp,

p<0.0001, and normal and slow, p=0.02, data sets.

Similarly, normality indices were all signiﬁcantly dif-

ferent between data sets, p<0.0001.

normal limp slow

MOCAP symmetry

normal limp slow

shank gyroscope symmetry

Figure 6: Symmetry indices. Distributions are presented

as boxplots. (A) illustrates the MOCAP symmetry results

for normal, limp and slow data sets. Results for the normal

data set were signiﬁcantly different from limp (p<0.0001)

and slow (p=0.03) data sets. (B) shows the shank gyroscope

symmetry index, using 20 symbols and period histograms.

The symmetry of the normal data set was signiﬁcantly dif-

ferent to that of limp (p<0.0001) and slow (p=0.02) data

sets.

5 DISCUSSION

It is important to stress that the MOCAP system

and the inertial sensors measure very different things.

Nonetheless, the raw data from both system can be

A WEARABLE GAIT ANALYSIS SYSTEM USING INERTIAL SENSORS PART I - Evaluation of Measures of Gait

Symmetry and Normality against 3D Kinematic Data

185

normal limp slow

MOCAP normality - GPS

normal limp slow

100

waist accelerometer normality

normal limp slow

shank accelerometer normality

Figure 7: Normality indices. Distributions are presented

as boxplots. (A) shows the GPS for all three types of

walk. Normal and limp, and normal and slow data sets

were signiﬁcantly different (p<0.0001). (B) illustrates the

normality results of all types of walk using the waist ac-

celerometer, 18 symbols, and transition histograms. Data

sets were signiﬁcantly different (p<0.0001). (C) shows re-

sults of the normality index using shank accelerometers, 5

symbols, and symbol period histograms. Similarly, normal

walk was signiﬁcantly different from limp and slow data

sets (p<0.0001).

processed in order to determine certain properties or

characteristics of gait which are the same or com-

parable. The MAP components, for example, con-

vey the normality of very speciﬁc joint movements.

The normality measure derived from the sensors, on

the other hand, expresses an overall normality of the

shank movements, which are caused by a combination

of different joint movements. The comparison of the

sensor normality with the different MAP components,

however, may provide insight into the the factors that

inﬂuence the sensor normality measure.

Although the shank sensor normality correlates

well with hip ﬂexion and not with pelvic rotation, the

waist sensor normality correlates well with both. This

suggests that the shank sensors are mostly affected

by the abnormalities in hip ﬂexion, whereas the waist

sensor is affected by abnormalities in hip ﬂexion and

pelvic rotation. This is aligned with the fact most gait

pathologies affect the movement of the center of grav-

ity (Detrembleur et al., 2000), which is captured by

the waist sensor.

Another interesting factor is that the correlation

of sensor symmetry with the MOCAP symmetry con-

sidering all components, is greater than the correla-

tion with any individual component. This may sug-

gest that the symmetry of individual joint angles is not

representative of the symmetry of the movement as a

whole, or at least not representative of the movement

of the shanks. It is the combination of joint move-

ments that results in the overall symmetry captured

by the shank sensors.

One may argue that the lack of information about

individual joint movements or other particular kine-

matic and kinetic parameters diminishes the useful-

ness of the proposed method. However, the sym-

bolic representation, once it is properly understood,

may reveal more precise information about the move-

ment. Consider for example the signals shown in Fig-

ure 8. (A) illustrates the gyroscope resultant signal af-

ter symbolization of a subject walking normally, and

(B) shows the signal from the same sensor when the

subject was limping. Only two strides are depicted in

each plot.

Note how symbol D in the normal signal is re-

placed by signal C in the limp signal. It is known

that symbol C represents a lower angular velocity

than symbol D. As previously mentioned, it is also

known that these symbols occur shortly after heel-

strike. Therefore, this difference from normal to

limp indicates that the shank rotated more slowly af-

ter heel-strike when the subject was limping. Given

that the duration of the symbols is approximately the

same, the shank rotated to smaller angle. The extrac-

tion of such information is not trivial, but expert sys-

tems can be developed for this purpose. Another ad-

vantage of the proposed instrumented GA is that, after

a diagnosis or clinical evaluation has been made, the

recovery or progress of the patient can be easily mea-

sured according to symmetry and normality.

1.5 2 2.5 3

A) Normal walk - right shank gyroscope

time (s)

symbols

1 1.5 2 2.5

B) Limp walk - right shank gyroscope

time (s)

symbols

Figure 8: Comparing normal and limp walk. (A) illustrates

the gyroscope resultant signal after symbolization of a sub-

ject walking normally, and (B) shows the signal from the

same sensor when the subject was limping. Only two strides

are depicted in each plot. Note how symbol D in the normal

signal is replaced by signal C in the limp signal. This exem-

pliﬁes how the symbolic representation of the signals may

be used to derive particular information about the subject’s

gait pattern.

Results show that the normal walk can be distin-

guished from the other walks with respect to normal-

ity and symmetry. Although limp and slow walk were

“acted” and not real, the purpose of the instructions

was to generate abnormal gait patterns. The distribu-

tion of the GPS shows that the fake patterns ranged

over varying levels of abnormality. This variety in the

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

186

data is useful in evaluating the expressiveness of the

proposed method, and how well it covers a wide range

of abnormalities. Results suggest that the normality

and symmetry measures are gradual and can, in fact,

be used to express varying levels of abnormality.

One important observation is that the proposed

method utilizes several steps for the analysis. In con-

trast, the MOCAP system is commonly limited to

one stride per foot, because only two force plates are

available in most gait labs. The acquisition of a larger

number of steps greatly increases the amount of work

needed. Moreover, one stride may not be represen-

tative of a subject’s average gait pattern due to intra-

subject variability (Chau et al., 2005). The sensor data

used in the present study was relatively short, aver-

aging approximately 3 strides at constant speed for

normal and limp walk. Longer data recordings might

provide more accurate analysis.

When using inertial sensors for human movement

analysis, the placement of the sensors can greatly af-

fect results. In the present study, for example, the

waist sensor was attached to the front of the subject.

Most studies however, attach the sensor to the back,

e.g. (Moe-Nilssen and Helbostad, 2004), (Auvinet

et al., 1999), (Hartmann et al., 2009). The present

study is part of a larger study where hip-replacement

patients were being monitored with an inertial sensor

node for long periods of time. It was decided that a

sensor placed in the front would be more comfortable

than one in the back. The gait data collection kept the

same conﬁguration. However, for short data collec-

tions, dealing with a varied pool of subject, it might

be beneﬁcial to place the sensor at the back. This way,

the sensor may stay closer to the center of gravity re-

gardless of the weight or shape of the subject.

MOCAP systems and inertial sensor systems

should not compete, they should complement each

other. MOCAP systems measure the position of sets

of reﬂective skin markers, speciﬁc to a marker model

that deﬁnes the orientation of each body segment. Ac-

celerations, rotations and joint angles are obtained

through speciﬁc algorithms developed for that par-

ticular marker model. One may wonder if the con-

straints imposed by different models and ﬁlters distort

the original data. Gait is already a well studied area,

but new movements require new models. Inertial sen-

sors such as accelerometers and gyroscopes may con-

tribute to the study of movements that are not yet fully

understood, e.g. freezing of gait in Parkinson’s Dis-

ease patients (Plotnik et al., 2005), (Hausdorff et al.,

2003). They also complement MOCAP systems in

that they are mobile, and can be used in a wider vari-

ety of contexts.

6 CONCLUSIONS

This study presented and evaluated a method for gait

analysis using inertial sensor data. A novel method

for measuring gait normality was introduced, based

on a symbolic approach previously used for gait sym-

metry. Symmetry and normality results were com-

pared to a reference derived from kinematic data. Re-

sults support that an estimate of gait symmetry and

normality, related to that derived from MOCAP data,

can be obtained with a simple combination of inertial

sensors. This supports the development of a cheap

and easy to use system for quantitative gait analysis

that can be used both in the clinic and in other less

controlled environments.

The proposed measures of symmetry and normal-

ity correlate well with the reference GPS and sym-

metry MOCAP measures. A follow-up study on hip-

replacement patients will attempt to determine if the

proposed method is also in agreement with clinical

judgment. This expectation is supported by the fact

that GPS is, in turn, well correlated with clinical judg-

ment.

ACKNOWLEDGMENTS

This study was partially funded by the Promobilia

Foundation and the Institute of Health and Care Sci-

ences, Sahlgrenska Academy, University of Gothen-

burg, Sweden.

REFERENCES

Aminian, K., Rezakhanlou, K., De Andres, E., Fritsch, C.,

Leyvraz, P. F., and Robert, P. (1999). Temporal feature

estimation during walking using miniature accelerom-

eters: an analysis of gait improvement after hip arthro-

plasty. Medical and Biological Engineering and Com-

puting, 37:686–691.

Auvinet, B., Chaleil, D., and Barrey, E. (1999). Accelero-

metric gait analysis for use in hospital outpatients. Re-

vue du Rhumatisme: English ed., 66(7-9):389–397.

Baker, R., McGinley, J. L., Schwartz, M. H., Beynon,

S., Rozumalski, A., Graham, H. K., and Tirosh, O.

(2009). The gait proﬁle score and movement analysis

proﬁle. Gait & Posture, 30(3):265–269.

Barton, G., Lisboa, P., Lees, A., and Attﬁeld, S. (2007).

Gait quality assessment using self-organising artiﬁcial

neural networks. Gait & Posture, 25(3):374 – 379.

Barton, G. J., Hawken, M. B., Scott, M. A., and Schwartz,

M. H. (2010). Movement deviation proﬁle: A measure

of distance from normality using a self-organizing

neural network. Human Movement Science, In Press,

Corrected Proof.

A WEARABLE GAIT ANALYSIS SYSTEM USING INERTIAL SENSORS PART I - Evaluation of Measures of Gait

Symmetry and Normality against 3D Kinematic Data

187

Beauchet, O., Allali, G., Berrut, G., Hommet, C., Dubost,

V., and Assal, F. (2008). Gait analysis in demented

subjects: interests and perspectives. Neuropsychiatric

Disease and Treatment, 4(1):155–160.

Beynon, S., McGinley, J. L., Dobson, F., and Baker, R.

(2010). Correlations of the gait proﬁle score and the

movement analysis proﬁle relative to clinical judg-

ments. Gait & Posture, 32(1):129–132.

Chang, F. M., Rhodes, J. T., Flynn, K. M., and Carollo,

J. J. (2010). The role of gait analysis in treating gait

abnormalities in cerebral palsy. Orthopedic Clinics of

North America, 41(4):489 – 506.

Chau, T., Young, S., and Redekop, S. (2005). Manag-

ing variability in the summary and comparison of gait

data. Journal of NeuroEngineering and Rehabilita-

tion, 2(1):22.

Coutts, F. (1999). Gait analysis in the therapeutic environ-

ment. Manual Therapy, 4(1):2 – 10.

Crenshaw, S. J. and Richards, J. G. (2006). A method for

analyzing joint symmetry and normalcy, with an ap-

plication to analyzing gait. Gait & Posture, 24(4):515

– 521.

Cruz, H. et al. (2008). Evidence of abnormal lower-limb

torque coupling after stroke: An isometric study sup-

plemental materials and methods. Stroke, 39(1):139.

DeLuca, P. A., Davis, R. B., unpuu, S., Rose, S., and

Sirkin, R. (1997). Alterations in surgical decision

making in patients with cerebral palsy based on three-

dimensional gait analysis. Journal of Pediatric Or-

thopaedics, 17(5):608–614.

Detrembleur, C., van den Hecke, A., and Dierick, F. (2000).

Motion of the body centre of gravity as a summary

indicator of the mechanics of human pathological gait.

Gait & Posture, 12(3):243–250.

Frenkel-Toledo, S., Giladi, N., Peretz, C., Herman, T., Gru-

endlinger, L., and Hausdorff, J. (2005). Effect of gait

speed on gait rhythmicity in Parkinson’s disease: vari-

ability of stride time and swing time respond differ-

ently. Journal of NeuroEngineering and Rehabilita-

tion, 2(1):23.

Gouwanda, D. and Senanayake, A. S. M. N. (2011). Iden-

tifying gait asymmetry using gyroscopes–a cross-

correlation and normalized symmetry index approach.

Journal of Biomechanics, 44(5):972 – 978.

Hartmann, A., Murer, K., de Bie, R., and de Bruin, E. D.

(2009). Reproducibility of spatio-temporal gait pa-

rameters under different conditions in older adults us-

ing a trunk tri-axial accelerometer system. Gait &

Posture, 30(3):351–355.

Hausdorff, J. M., Schaafsma, J. D., Balash, Y., Bartels,

A. L., Gurevich, T., and Giladi, N. (2003). Impaired

regulation of stride variability in parkinson’s disease

subjects with freezing of gait. Experimental Brain Re-

search, 149:187–194.

Kay, R. M., Dennis, S., Rethlefsen, S., Reynolds, R. K.,

Skaggs, D. L., and Tolo, V. T. (2000). The effect

of preoperative gait analysis on orthopaedic decision

making. Clinical Orthopaedics and Related Research,

372:217–222.

Lofterød, B. and Terjesen, T. (2008). Results of treatment

when orthopaedic surgeons follow gait-analysis rec-

ommendations in children with cp. Developmental

Medicine & Child Neurology, 50(7):503–509.

Moe-Nilssen, R. and Helbostad, J. L. (2004). Estimation

of gait cycle characteristics by trunk accelerometry.

Journal of Biomechanics, 37(1):121 – 126.

Novacheck, T. F., Stout, J. L., and Tervo, R. (2000). Re-

liability and validity of the gillette functional assess-

ment questionnaire as an outcome measure in chil-

dren with walking disabilities. Journal of Pediatric

Orthopaedics, 20(1):75.

Plotnik, M., Giladi, N., Balash, Y., Peretz, C., and Haus-

dorff, J. M. (2005). Is freezing of gait in Parkinson’s

disease related to asymmetric motor function? Annals

of Neurology, 57(5):656–663.

Read, H. S., Hazlewood, M. E., Hillman, S. J., Prescott,

R. J., and Robb, J. E. (2003). Edinburgh visual gait

score for use in cerebral palsy. Journal of Pediatric

Orthopaedics, 23(3):296–301.

Salarian, A., Russmann, H., Vingerhoets, F., Dehollain, C.,

Blanc, Y., Burkhard, P., and Aminian, K. (2004). Gait

assessment in Parkinson’s disease: Toward an ambula-

tory system for long-term monitoring. IEEE Transac-

tions on Biomedical Engineering, 51(8):1434 –1443.

Sant’Anna, A., Salarian, A., and Wickstr

om, N. (2011). A

new measure of movement symmetry in early parkin-

son’s disease patients using symbolic processing of in-

ertial sensor data. IEEE Transaction on biomedical

Engineering. Epub ahead of print, 2011.

Sant’Anna, A. and Wickstr

om, N. (2010). A symbol-based

approach to gait analysis from acceleration signals:

Identiﬁcation and detection of gait events and a new

measure of gait symmetry. IEEE Transactions on In-

formation Technology in Biomedicine, 14(5):1180 –

1187.

Schutte, L. M., Narayanan, U., Stout, J. L., Selber, P., Gage,

J. R., and Schwartz, M. H. (2000). An index for quan-

tifying deviations from normal gait. Gait & Posture,

11(1):25–31.

Schwartz, M. H. and Rozumalski, A. (2008). The gait de-

viation index: A new comprehensive index of gait

pathology. Gait & Posture, 28(3):351–357.

Shin, K.-Y., Rim, Y., Kim, Y., Kim, H., Han, J., Choi, C.,

Lee, K., and Mun, J. (2010). A joint normalcy index

to evaluate patients with gait pathologies in the func-

tional aspects of joint mobility. Journal of Mechanical

Science and Technology, 24:1901–1909.

Silver, K., Macko, R., Forrester, L., Goldberg, A., and

Smith, G. (2000). Effects of aerobic treadmill train-

ing on gait velocity, cadence, and gait symmetry in

chronic hemiparetic stroke: A preliminary report.

Neurorehabilitation and Neural Repair, 14(1):65–71.

Toro, B., Nester, C., and Farren, P. (2003). A review of ob-

servational gait assessment in clinical practice. Phys-

iotherapy Theory and Practice, 19(3):137–149.

Tranberg, R., Saari, T., Z

ugner, R., and K

arrholm, J. (2011).

Simultaneous measurements of knee motion using an

optical tracking system and radiostereometric analysis

(RSA). Acta Orthopaedica, 82(2):171–176.

Verghese, J., Lipton, R., Hall, C., Kuslansky, G., Katz, M.,

and Buschke, H. (2002). Abnormality of gait as a

predictor of non-Alzheimer’s dementia. New England

Journal of Medicine, 347(22):1761.

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

188