Inertial Measurement Units in Gait Analysis Applications

Questions, Suggestions and Answers

Harald Loose

, Katja Orlowski

and Robert Amann

Fachhochschule Brandenburg, Magdeburger Str. 50, 14770 Brandenburg, Germany

Fachhochschule Vorarlberg, Hochschulstr.1, Dornbirn, Austria

Keywords: IMU, Sensors, Systems, Gait Analysis, Gait Events, Gait Feature, Positions, Angles.

Abstract: The paper deals with inertial measurement units (IMU) and their application in gait analysis in the wide

range from movement monitoring through rehabilitation feedback to sports improvement. An IMU sensor

incorporates three microelectromechanical sensors - triple-axis gyroscope, accelerometer, magnetometer –

and, optionally, a barometer. The outputs of all sensors are processed by an on-board microprocessor and

sent over a serial interface using wired or wireless communication channels. The on-board processing may

include sensor conditioning, compensations, strap-down integration as well as determination of orientation.

The sensor output is sent to applications working on standard PC, tablets or smart phones using different

sampling rates. The output data of one IMU sensor allow motion analysis of the sensor unit itself as well as

the motion of the limb where the sensor is mounted to. Using a combination of two or more sensors the

movement of limbs/legs can be compared; their relative motion can be investigated; angles can be

calculated.

In general, in motion and gait analysis, we like to get primary information about the position of all

interesting points, the orientation of the limbs and the joint angles at each moment of time as well as derived

averaged and summarized characteristics about the motion and the gait. Based on our own investigations the

paper discusses how much information is really necessary to determine gait events and gait features for

different purposes.

1 INTRODUCTION

Inertial measurement units (IMU) are freely

available at the market: from low cost boards or

sticks to relatively expensive sensors assembled in

small and light weight packages. An IMU sensor

incorporates three microelectromechanical sensors -

triple-axis gyroscope, accelerometer, magnetometer

– and, optionally, a barometer. The outputs of all

sensors are processed by an on-board

microprocessor and sent over a serial interface using

wired or wireless communication channels. The on-

board processing includes sensor conditioning,

compensations, strap down integration (SDI) as well

as determination of orientation. Otherwise post-

processing tools are provided to perform these

calculations, e.g. to calculate orientation. 9 DOF-

IMU sensors including data acquisition software and

software development kits (SDK) are provided, e.g.,

by Shimmer (www.shimmersensing.com), Xsens

Technologies (www.xsens.de), stt engineering and

systems (www.sst-systems.com), life performance

research (www.lp-research.com), Kionix (www.

kionix.com), Noraxon (www.noraxon.com), Analog

Devices (www.analog.com). IMU sensors are

applied for motion capture, measurement,

processing, navigation and control. In this paper we

consider only applications related to the gait analysis

and healthcare. Many companies providing IMU

sensors developed software products for research,

clinical needs, rehabilitation and sports (Xsens MVN

Biomech, sst clima, noraxon clinical). Small

companies develope purpose oriented, low cost tools

like RehaWatch (www.hasomed.de). Based on the

data of a number of IMU sensors (17 in MVN

Biomech) and the kinematic model of the human

body various derived values and features are

determined:

 position of characteristic points, orientation of

limbs, angles of joints at each moment,

 cadence, distance and velocity of motion, index

of symmetry,

 characteristics of each or averaged stride like

initial and terminal point, length, height,

300

Loose H., Orlowski K. and Amann R..

Inertial Measurement Units in Gait Analysis Applications - Questions, Suggestions and Answers.

DOI: 10.5220/0005278703000305

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2015), pages 300-305

ISBN: 978-989-758-069-7

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

circumduction, relationship between stance and

swing.

IMU based motion capture systems are usually

compared to the accuracy standard of conventional

optical motion capture systems such as Vicon

(www.vicon.com). Sometimes they are considered

to have the same accuracy. Alternatively the output

of IMU based calculations can be proofed using high

accurate position measurement systems like API

Radian (www.apisensor.com) or comparing with

well-known results from gait observations and

analysis (Perry, 2010, Murray, 1964).

Since we have been working for a long period on

this topic, mostly in student projects, the aim of this

position paper is to ask questions and to try to give

answers.

2 SYSTEMS AND EXPERIMENTS

Since about five years we have been using sensor

systems for the acquisition of various bio-signals

like ECG, EEG, EMG or motion data. Sensors

applied directly to limbs/body were tested as well as

position measurement systems which in-motion data

acquisition and are used for comparison. A-priori

knowledge about gait patterns and kinematic models

of the human skeleton are involved in algorithms as

well as in plausibility tests.

In the field of motion analysis we have been focused

on human gait with respect to health applications

e.g. in orthopaedics, physiotherapies and rehabili-

tation. The motion of the patient is relatively slow

(~1-2 m/s) with moderate changes of the linear and

angular velocity.

In clinical practice experts observe the movement of

patients going straight forward about five to ten

strides. Assistant measurement systems and

applications will be able to quantify those

observations, to make them comparable and

traceable over time. In this paper we discuss the last

experimental setup where 9DOF Xsens sensors were

placed on the pelvis and all lower limbs, forming

together the kinematic gait chain.

2.1 Systems

2.1.1 IMU Sensors

In the current experiments we use up to seven 9DOF

Xsens MTw sensor units connected via Bluetooth to

one Awinda station and data acquisition software

“MT Manager”. On-board the data of the primary

sensors are sampled with 1800 Hz, strapped down

by integration (SDI) incorporating the estimate of

orientation to the transfer rate of 100 Hz for two or

60 Hz for seven MTw. Finally the “MT Manager”

provides synchronized data from all involved MTw

(< 10 s accuracy), i.e., linear acceleration a,

angular velocity ω, magnetic field m and quaternion

q (orientation estimated on-board < 1° of static and

2° RMS of dynamic accuracy (www.xsens.de ).

Before starting measurements sensors need calm or

slow motion to “warm up the filters”, to calculate

the initial orientation of the sensor with respect to

the world coordinate system. The implemented

Xsens-Kalman-Filter is based on the assumptions

that on the average the acceleration due to the

movement is zero and that the magnetic field is

homogenous or steady state.

2.1.2 Gait Pattern

In the middle of the last century Perry (Perry, 2010)

and Murray (Murray, 1964) observed, measured and

analysed the normal human gait. The gait pattern

covers one stride, the full period of movement of

one leg, one stance and one swing phase. The given

pattern includes average trajectories of joint angles

(hip, knee and ankle), the angle between thigh and

vertical (in sagittal plane) as well as average

trajectories of the center of hip (pelvis). They

discovered several gait events, e.g. initial and

terminal foot contact to the floor, heel strike, flat

foot, heel off, toe off. Events disjoin the stride into

stance and swing phase as well as into eight more

detailed sub-phases.

2.1.3 Kinematic Model

On the base of a planar model of the kinematic chain

of lower limbs average trajectories of hip, knee,

ankle, middle foot and toes are derived from a given

gait pattern (position, linear and angular velocity and

acceleration). These patterns allow the identification

of correspondences between gait events and

characteristic points of acquired or derived data

(minima, maxima, zero crossings).

2.1.4 Position Measurement

The API Radian laser tracker was used to measure

the movement of the foot and the ankle with 1 kHz

sampling rate and accuracy of 50 m. To process the

measurement a relatively heavy controller ball is to

be mounted to foot or ankle. The ball dynamics may

not be neglected (impact of heel strike in vertical

direction).

InertialMeasurementUnitsinGaitAnalysisApplications-Questions,SuggestionsandAnswers

301

2.1.5 Motion Capture

Canon EOS 5D was used to record the movement

with 60 fps - the sampling rate of each Xsens Mtw

sensors (for seven IMU). Two strides were captured

with a resolution of 1280 x 720 px, so that it is

possible to reference gait events to single frames.

2.1.6 Evaluation Software

To answer various questions and to evaluate several

approaches we have developed an open MATLAB

script which is organized to process experimental

data automatically step by step. After each step the

intermediate results are saved. Optionally, figures

can be created and written to hard disc. In

dependence of the task steps can be skipped or

repeated. The following steps are included:

 reading and reorganizing sensor by sensor the

acquired data, given in the sensor related

coordinate system (SCS),

 estimation of orientation (quaternion), if

necessary, using the Madgwick algorithm

(Madgwick, 2011),

 transformation of sensor data into world

coordinate system (WCS),

 calculation of orientation relative to the initial

one,

 calculation of angles between z-axes of a sensor

and the vertical or the horizontal plane,

 calculation of various features as candidates for

gait events,

 detection of inner strides,

 determination of direction of movement,

transformation of the sensor data into motion

coordinate system (MCS),

 integration of acceleration to calculate velocity

and position data stride by stride,

 calculation of stride related and average

features,

 determination of average stride,

 calculation of joint angles, if relationship

between the sensors is given,

 calculation of symmetry ratios, if couples of

sensors (left and right) are given,

 extraction of gait characteristics.

2.2 Experiments

The described experiments were designed this

summer, first, to acquire input data for the

evaluation software and, second, to prepare a

“standard” experiment addressed to a large number

of healthy subjects for statistical analysis. To have

the chance to attract later on experts the movement

was partly recorded (> 2 strides per leg).

2.2.1 Experimental Setup

Xsens MTw sensors are attached laterally pairwise

on left and right lower limbs at shoes below knees,

at thighs, and one on back (pelvis) as shown in

ﬁgure 1. The camera was placed about 15 cm above

the floor, 5 m from and orthogonal to sagittal plane

of the subject.

Figure 1: Alignment of the sensors.

2.2.2 Evaluation

The movement of two healthy subjects, one female

and one male, mid-twenties, was observed. They

were asked to choose the speed for normal, slow and

fast walking themselves passing a distance of about

10 m twice from left to right and back. After some

time the same experiment was repeated. The

experiments were executed indoors where magnetic

field was neither homogenous nor steady state as

learned during the post-processing of data.

3 DISCUSSION AND RESULTS

This contribution is a position paper allowing to

outline questions followed by our suggestions and

current answers. Both can be treated as a source of

discussions.

Most of the results stated in the discussion

belong to sensors mounted on left or right shoe

(foot) acquired with 60 Hz; otherwise it will be

mentioned.

BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

302

Figure 2: Acquired acceleration and angular velocity with

respect to SCS (black – length of the vector).

Do the acquired data figure out the periodicity of

movement stride by stride?

Repeating patterns obviously appear periodically,

varying little from stride to stride (see figure 1).

During stance phase acceleration is close to zero,

during swing phase large changes are typical.

Can the number of strides simply be counted?

Repeating patterns are clearly separated. Each

pattern can be assigned to one stride, so that the

number of patterns is equal to the number of strides.

The first and last step is incomplete (see last pattern

in figure 2). Both feet are side by side before and

after movement.

Is acceleration due to movement dominant in

relation to g? Is its average zero?

Yes, peaks of acceleration are up to 6 g. Walking is

characterized by change of stance and swing phases,

so that average of acceleration is zero periodically

after every second.

Is there any predominance of components of

acceleration or angular velocity?

Forward component of acceleration (direction of

movement) dominates vertical one. Lateral

component of angular velocity is dominant.

Does it seem to be possible to determine, e.g.,

length of stride or passed distance, from sensor

data without transformation into WCS?

There are several software systems like RehaWatch

and papers (e.g., Orlowski, 2013) showing that it

seems to be possible. Sensors need to be mounted

with high accuracy such that SCS coincides with

WCS at calm (stance phase). Small and short

rotations of sensors are neglected.

What can the quaternion of orientation be used

for?

The orientation is not measured, but estimated from

angular velocity, acceleration and magnetic field. If

orientation is related to WCS components of all

vectors are transformed into that inertial coordinate

system. Gravitation vector g can be eliminated.

Various angles, e.g., between sensors, vertical and

floor as well as changes of angles between sensors,

can be easily calculated.

Are the algorithms of Xsens and Madgwick

comparable?

The Xsens-Kalman-Filter (Roetenberg, 2009) is

implemented on-board the Xsens MTw sensors.

The

algorithm of Madgwick is implemented in post-

processing of acquired sensor data. Because of

heterogeneity of magnetic field orientation was

estimated only from angular velocity and

acceleration. Gait features were calculated using the

same algorithm. Differences of about 5% are

noticeable, e.g., average length of stride is mostly

larger for the Madgwick algorithm.

Figure 3: Acceleration and angular velocity with respect to

WCS (black – length of the vector).

Which definition of gait cycles is goal-oriented

with respect to the integration of the

acceleration?

There are various possibilities to determine the gait

cycle (GC). Movement starts and finishes at calm.

Any significant moment of the stance phase, where

the velocity of foot is zero, can be used as transition

point from one stride to the next one. In principle

any other significant moment, e.g., maximum of

angular velocity vector can be considered as the

transition point too. Perry (Perry, 2010) proposed the

initial contact point (IC) as the beginning of GC.

Stride-by-stride integration should start at a moment

of calm (velocity equal zero) or at least at minimum

of motion.

InertialMeasurementUnitsinGaitAnalysisApplications-Questions,SuggestionsandAnswers

303

Are there any features of data indicating

transition from stride to stride?

To find indicators of the transition point all

measured and derived data can be considered.

Manifold ideas are given in literature (e.g. Green,

2010) focussing on points nearby IC, maximum of

lateral component of angular velocity, forward

component of acceleration or angle between foot and

ground. In our algorithm the minimum of weighted

sum of acceleration and angular velocity (maximal

calm) is used. This indicator is significant for whole

stance phase, so that additional plausibility checks

should be added.

How can the direction of movement be

determined in WCS?

While z-axis of WCS coincides with vertical the

other axis are in horizontal plane. Integrating

horizontal components of acceleration velocity

vector is calculated. It defines the current direction

of movement. WCS is rotated about vertical so that

x-axis of resulting MCS coincides with direction of

movement.

Figure 4: Acceleration and angular velocity with respect to

MCS (black–begin of integration).

Does measured stride coincide with normalized

gait pattern?

Comparing figures 4 and 5 the similarity between

gait pattern given by Perry (Perry, 2010) and our

data is obvious.

Is the integration stride-by-stride preferable in

relation to integration over the whole movement?

Choosing integration intervals, so that acceleration

average as well as initial and end velocity are equal

to zero, algorithms work very well (foot sensors). In

our algorithms we considered only inner strides,

excluding first and last half-strides (figures 6 and 7).

Figure 5: Pattern of acceleration and angular velocity

(Perry, 2010).

Figure 6: Measured acceleration, calculated velocity and

displacements – forward, lateral and vertical.

Are the algorithms applicable considering the

movement to other sensors?

The assumption, that initial and end velocity equals

to zero, is not fulfilled for sensors mounted above

the ankle. Their unknown minimum velocity (during

inner strides) increases with the distance from ankle.

The detection of transition points should be

improved because of absence of calm.

Is there any chance to estimate the minimum

velocity for other sensor locations?

Unknown minimum velocity can be estimated, first,

BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

304

based on of a model or, second, through including

first and last stride into the integration interval.

During those strides acceleration average is not zero.

Figure 7: Measured displacement, calculated acceleration,

and velocity– forward, lateral and vertical – using API

Radian.

Which and how many GC have to be involved in

the calculation of average and standard

deviation?

First and last strides are used to accelerate

(decelerate) before establishing steady movement.

To determine average only inner strides, not less

than 5, should be considered. The exact selection of

all GC and scaling to 100 % seems to be very

important.

Is there any advantage to include two or more

sensors into algorithms?

Including two or more sensors into algorithms offers

a lot of chances: to calculate angles between two

sensors (joint angles), to compare their behaviour

(symmetry) or to include improvements based on

kinematic relationships.

Does model-based calculation promise large

improvement of accuracy?

The Xsens MVN Biomech studio is based on 17

sensors and biomechanical skeleton model. It is used

in motion capture as a substitute of optical systems.

Following model-based algorithms promise

improvement of accuracy even in small systems.

4 CONCLUSIONS

The evaluation algorithms work well regarding the

data of foot sensor. Sometimes there are problems

with the accuracy of transition points between swing

and stance, i.e. starting points of the stride-by-stride

integration. Improvement of their accuracy can be

achieved including more features as well as more

sensors. At the same time calculation of average

stride will be improved and, following, the

symmetry analysis based on average GC.

Algorithms incorporating only one sensor and

processing integration stride by stride maybe applied

to other sensor location. If the absolute velocity

should be determined, the question of the steady part

of velocity has to be solved. Steady part increases

with the distance of sensors from floor. Model-based

calculation may be goal-oriented in this context, as

well as for the determination of joint angles and

symmetry indexes.

REFERENCES

Perry, J. (2010). Gait Analysis - Normal and Pathological

Function. Slack Inc.

Murray, M. (1964). Walking patterns of normal men. J

Bone Joint Surg Am, 46A:335 – 360.

Roetenberg, D., Luinge, H., and Slycke, P. (04/2009).

Xsens mvn: Full 6 dof human motion tracking using

miniature inertial sensors. XSENS TECHNOLOGY,

White Paper.

Madgwick, S., Harrison, A. J. L., and Vaidyanathan, R.

(2011). Estimation of IMU and MARG orientation

using a gradient descent algorithm. In Rehabilitation

Robotics (ICORR), 2011 IEEE International

Conference on, pages 1–7.

Orlowski, K. and Loose, H. (2014). Simple Algorithms for

the Determination of the Walking Distance based on

the Acceleration Sensor. Proceedings of

BIOSIGNALS 2014, Int. Conference on Bio-Inspired

Systems and Signal Processing, Angers, France, 03-06

March 2014, pp. 264–269.

Greene, B. (2010). An adaptive gyroscope-based

algorithm for temporal gait analysis. Med Biol Eng

Comput, 48:1251–1260.

InertialMeasurementUnitsinGaitAnalysisApplications-Questions,SuggestionsandAnswers

305