Estimation of Gait Parameters based on Motion Sensor Data

Kaitai Li

and Cong-Cong Zhou

2,*

Fuzhanghuanzhou BioTech Company, Guoxia Road.258, Yangpu District, Shanghai, P.R. China

College of Biomedical Engineering and Instrument Science, Biosensor National Special Laboratory, Zhejiang University,

Hangzhou 310027, P.R. China

Keywords: IMU, Stride Parameters, Gait Analysis.

Abstract: Recently, the spreading application of intelligent mobile devices with integrated sensors such as inertial

measurement units (IMU) has attracted the interest of the researchers for designing gait analysis methods

based on the captured sensor data. This paper focuses on designing a system which can evaluate the walking

ability and the physical agility level of normal people and people with Parkinson’s disease or stroke. The

motion signal is collected by three wearable MPU9250 sensors located on both ankles and the center of the

waist. Three test scenarios, including 10 meters walking test (10MWT), Time up and go test (TUGT) and

Dual-task walking (DTW), are designed in this paper. The results, which concluded time parameters such as

standing up time and turning back time as well as walking parameters such as stride length and stride

frequency, showed good consistency and high accuracy with Vicon device.

1 INTRODUCTION

Wireless motion sensors can be placed on different

positions of people to evaluate walking ability and

physical agility. The analyzed features can be

delivered to the scale assessment module to quantify

the physical condition (Hanson et al., 2009).

Currently, the unified Parkinson's disease rating scale

(UPDRS) and the Glasgow coma scale (GCS) are

widely used methods to do the assessment to

Parkinson’s and stroke patients (Mov, 2003).

Normally, evaluators can assess the body

condition of patients by experience. However, this

method is not objective enough and it is the reason for

the extensive attention of using data collected from

sensors to quantify the body condition. Moreover,

according to the normal algorithm designed to

process the motion signal, there is still improvement

should be done to get better performance. Different

division methods for movement periods and different

data fusion methods for transformation of axes were

discussed and developed. For example, Madgwick

(Madgwick, 2010) developed an attitude and heading

reference system (AHRS) algorithm. Besides, classic

data fusion modules, such as Kalman filter and

 

 Corresponding author

Complementary filter, are also extensively used and

developed. Generally speaking, data fusion modules

need to achieve different requirements to obtain the

ideal result, such as Kalman filter needs to establish a

complicated prediction model, in which the

covariance matrix is difficult to establish and the

signal processed must be linear.

In this paper, we aim at designing a system to

quantify some parameters used in the scale

assessment module which are hard to quantify by

observation methods. New methods used to divide

motion periods and fuse the 6 axis data are discussed

and tested. Specifically, three test scenarios, 10 MWT,

TUGT and DTW, are designed. During the test

process, three data acquisition modules with IMU

units are placed on two ankles and the center of the

waist. The collected motion data is pre-processed by

a digital signal processor integrated in the data

acquisition module and then transmitted wirelessly to

the stationary computer. Three data processing

functions are designed on stationary computer

corresponding to the three test scenarios mentioned

above. The designed motion periods division

methods used in the three functions should be able to

eliminate the effect of slight shaking or movement of

sensors (select profit axis). Firstly, the up and down

Li, K. and Zhou, C.

Estimation of Gait Parameters based on Motion Sensor Data.

DOI: 10.5220/0008963901290135

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 1: BIODEVICES, pages 129-135

ISBN: 978-989-758-398-8; ISSN: 2184-4305

129

function divides standing up and sitting down periods

and outputs standing up time and sitting down time.

Secondly, the turn back function divides turn back

periods during the walking process and outputs

turning back time with turning back angle. Thirdly,

the walking calculation function divides gait periods

and outputs stride length as well as relevant

parameters.

To validate the efficiency of functions, reference

outputs calculated by the Vicon device are recorded

and compared with the results calculated by the

proposed functions. Vicon is an optical motion

capture system produced by OML Company. The

accuracy rates are shown by tables and the advantages

and disadvantages of the designed method is

discussed in this work.

2 EXPERIMENTS AND

METHODS

The system can be divided into two main modules:

the data acquisition module and the data processing

module. According to the data acquisition module, a

hardware system is designed to collect 9 degree of

freedom motion signals of individuals. Two data

acquisition modules are placed on both ankle joints

and one data acquisition module is placed on the

center of the waist. As for the data processing module,

a reliable and promoted algorithm which is used to

obtain features during the gait process of participants

is designed. The following figure shows the simple

system flow,

Figure 1: System flow chart.

2.1 Data Acquisition Module

The multi-chip module, MPU9250, which concludes

a 3-axis gyroscope, a 3-axis accelerometer and a 3-

axis magnetometer, is applied in the data acquisition

module. The initial motion signals are amplified and

digitalized in the module and transmitted to the

stationary computer wirelessly.

2.2 Sensors Wearing

As is shown in Figure 1, the data acquisition modules

are adhered to the elastic bandages with adhesive

tapes and then the bandages are wore on two ankle

joints and the center of the waist of the user.

Considering there might be trembles or waggles

during the process of experimental paradigms,

positions of the first two data acquisition modules are

not required to be precisely placed on the center of

feet. The third module is required to be placed on the

center of the waist as much as possible. The impacts

of trembles and waggles are designed to be tackled in

algorithms.

2.3 Test Scenarios

The tests aim at evaluating the walking ability and the

physical agility level of normal people, Parkinson’s

disease patients and stroke patients. Three designed

test scenarios are shown as below:

 10 Meters Walking Test (10MWT): The

participant is wore with three data acquisition

modules and asked to walk straightly on a 14

meters’ walkway with normal speed. The

walkway is marked on two meters and 12 meters.

The data collected from the in-between 10 meters

test will be used.

 Time Up and Go Test (TUGT): The participant

is wore with three data acquisition modules and

asked to stand up from a chair, walk 3 meters in a

straight line, turn around, and then walk 3 meters

back to sit down.

 Dual-Task Walking (DTW): The test aims at

Parkinson’s disease and stroke patients. Based on

the 10WMT, the patient is required to calculate

100 keeps subtracting 2 as the walking process.

Times of correct answers and wrong answers

should be noted down.

2.4 Data Processing Module

The signal collected from the data acquisition module

is wirelessly transmitted to the stationary computer.

The accelerometer signal unit is converted to









the

gyroscope signal unit is converted to









and the

magnetometer signal unit is converted to µT based on

BIODEVICES 2020 - 13th International Conference on Biomedical Electronics and Devices

130

the range of data collection. Afterwards, the signals

are filtered by a Butterworth band pass filter.

The algorithms are developed and tested in

Matlab 2015b (www.mathworks.com), the data

obtained and figures used in this paper are produced

in this software as well. According to the engineering

landing of the algorithms, the software which

integrates all algorithms is developed by C# in

Microsoft Visual Studio 2017

(visualstudio.microsoft.com). Specifically, there are

three algorithm modules that correspond to the three

test scenarios. Figure 2 describes the flow diagram for

the developed gait analysis algorithm.

Figure 2: Flow diagram of the algorithm.

2.4.1 Up and Down Algorithm

The algorithm in this function corresponds with the

up and go test scenario. Output features contain

standing up time and sitting down time. The core part

of the algorithm is to find out the suitable axis that

contains enough information to figure out the

standing up and sitting down periods. As participators

are not required to wear data acquisition modules

with a standard direction, which means the sensors

might slide the location or skew with an angle,

selecting the applicable axis can help us eliminate the

effect caused by these issues. According to the up and

down function, the algorithm is designed with the

following steps,

 Find the Applicable Axis:

According to the physical movement situation of

sitting down and standing up, the sensors should have

vertical shifts. On the other hand, along with the

sitting down and standing up process, body leans back

and forward, thereby the angular velocity of forward

axis changes. As a result, the changes of vertical

accelerated speed and the forward angular velocity

can be considered as features used to define the

suitable axis.

When the variation of angular velocity is used as

the feature, the suitable axis should be the axis with

the maximum variance, which means the axis has

maximum changes during the test process. If a

window is added to the signal of the suitable axis, it

can be found that the variance is higher during

standing up and sitting down process. Considering

that the algorithm should be applicable to Parkinson’s

disease and stroke patients who might have tremor on

the forward direction, the axis with the lowest

accelerate speed variance is selected to use.

 Interval Divisions:

The standing up and sitting down intervals are half

sine waves at the beginning and the end of the test.

Peaks on the axis are found by using a developed

peaks finding algorithm. The threshold height value

of peaks is set as E (the expectation of signal) +

(weight value) x



2 ∗Vthevarianceofsignal

;

thus giving:

T1E 0.6∗

√

2∗

(1)

The formula (1) comes from 3 sigma principle

(Marilyn et al., 2007) and be modified according to

actual signal condition. The maximum distance

between two peaks is set as,

Maxdistancemeandistance/2



(2)

It is also necessary to confirm that the peak has

the largest value within certain range, which is set as

[-Max distance, Max distance]. The zero crossing

points nearby the peaks are considered as the border

of getting up and sitting down periods. The getting up

and sitting down time can be obtained by calculating

the length of two intervals.

2.4.2 Turn Back Algorithm

Output features of this function are turning time and

Estimation of Gait Parameters based on Motion Sensor Data

131

turning back angle. The applicable axis of the

gyroscope signal is found and output features are

calculated based on this. The algorithm is designed

with the following steps,

 Find the Suitable Axis:

The turning mainly happens around the vertical axis

of the body. The angle around the vertical axis

changes apparently thereby the variation of angular

velocity is considered to be the feature used to define

the axis. Actually, the axis we use here should be as

same as the axis we use in the standing up and sitting

down function. However, considering about the

physical significance, the axis with the maximum

angular rate variance is selected.

 Turning Back Interval Division:

The turning back interval might be a half sine wave,

a half cosine wave, a whole sine wave or a whole

cosine wave. This depends on how an individual turns.

For example, if an individual turns back with an

almost 360 degree angle, the interval should be a

whole sine wave. To figure out the intervals turning

back period during test process, firstly the threshold

height value of peaks is set with the formula below,

T2E

√

2∗∗1.5

(3)

The maximum distance between two peaks and

the certain range that contains one largest peak value

are set as same as the standing up and sitting down

function. The number of zero crossing points between

two peaks should be two. The turning back angle is

obtained by integrating the angular rate in the turning

back period and the turning back time is the length of

the turning period.

As a result, if there is a turning during the test

process, and the turning angle is larger than 90 degree,

the turning is considered as a real turning and counted.

2.4.3 Walking Algorithm

Walking algorithm function is called in all three test

scenarios, which makes it the core part of the project.

In the walking algorithm function, the output objects

are stride length and relevant parameters such as

stride speed and stride frequency. The gait time

periods should be classified firstly and then output

parameters can be calculated. The function can be

divided into the following two sections,

 Gait Time Periods Classification:

Gait time periods can be divided into swing phase and

stance phase roughly. It is common to use the variance

of accelerometer values of the forward axis as the

feature to classify gait periods. On account of there

should be no significant change of the acceleration

during the stance phase, the variance value in stance

phase should be almost zero. On the other hand,

during the swing phase, the variance value should

fluctuate with a non-zero value. However, the

classification method is limited in a series of

conditions. For example, the threshold of the

acceleration variance is nearly impossible to be

applicable to all gait periods even though an adaptive

threshold is used.

Considering about the disadvantages of previous

methods, a developed method is designed to classify

gait time periods precisely. Specifically, the

developed method contains two parts,

Selection of Reference Axis:

The gyroscope data can be used to divide the gait time

periods as stated by Koichi (Koichi et al., 2000). In

the developed method, the appropriate gyroscope axis

is selected with a new combined feature T3, which

comes from a combination of features on time domain.

The axis of gyroscope that we need is the one has the

most obvious and periodic signal feature on the time

domain. Specifically, the selected axis should have

three typical time domain characteristics: the extent

of the fluctuation of peak values is smaller, the

number of peaks after filtered (with the method

introduced in 2.4.1) is larger and the data of the axis

is with good symmetry.

T3 is obtained by adding weights to above three

features, which is shown below,

T3





difference



peaks







peaks



∗Number









difference



valleys







valleys



∗Number







∗ABS



Number





Number









(4)

The axis with the smallest value of T3 is the axis

of gyroscope used to classify gait time periods.

Gait Time Periods Classification:

Gait time periods are divided on the selected

gyroscope axis by finding out appropriate peaks and

zero crossing points. Peaks are found with the

methods introduced in 2.4.1. During the stance period,

the gyroscope value should pass through zero

crossing points, therefore, the zero crossing points

nearby peak values can be considered as borders of

gait periods. With the developed method, gait time

periods is divided as the figure below,

 Parameters Calculation:

Data

on the geodetic coordinate (GC) can be

BIODEVICES 2020 - 13th International Conference on Biomedical Electronics and Devices

132

Figure 3: Divisions of gait time periods.

transformed from data on sensor coordinate (SC) via

the transformation matrix 





represented by

quaternion. To obtain the updated and accurate

quaternion matrix 



,



,



,



, the data fusion

algorithm is necessary. Normally speaking, there are

several types of data fusion algorithms, such as

Kalman filter, complementary filter and Mahony &

Madgwick Filter. While there are limitations for the

algorithms mentioned above, take Kalman filter as

example, it needs significant modeling assumptions

to reach its theoretical results (Reza, 2009). On the

other hand, the generalized complementary filter is

not as accurate as the Kalman Filter. Thus a

developed complementary filter is designed.

As the aimed outputs do not require knowing the

yaw angle, a six axes data fusion method is applied.

What is confirmed is that in the GC, the accelerate

data on two axes of horizontal plane should be zero

and the value on vertical axis should be close to 1

(after normalization). According to the basic rationale

of complementary filter, the updated gyroscope

values can be shown as a fusion of previous

gyroscope values and parameters calculated from the

errors of acceleration values. A new error

compensation model is designed in the developed

algorithm and the steps are shown as below,

Calculation the Error Parameter:

The reference acceleration vector is obtained by

multiply the acceleration vector [0, 0, 1] by the

transformation matrix represented by quaternion. By

doing the difference calculation of the reference

acceleration vector and the acceleration vector

collected from the sensor, the error parameter is

obtained. It can be estimated that the error parameter

can be used to support long time periods of time.

Data Fusion:

Considering wavelet transform can effectively

describe the time domain characteristics on high

frequency periods and the frequency domain

characteristics on low frequency periods, it is selected

to combine the gyroscope signal and the accelerate

signal, instead of the traditional proportional-plus-

integral (PI) control method. The selection of

appropriate PI control parameters is difficult. On

contrast, there are already a series of mature threshold

setting methods of wavelet transform.

A window whose length is 50 is added to

accelerate error signal and gyroscope signal to

calculate appropriate filtered values. The updated

angular rate values (ω

x,y,z

) are represented by the

combination of low pass filtered angular rate values

(ω

x,y,zlf

) and the weighted error parameters come

from high pass filtered acceleration values

(error

x,y,zhf

). The formula is shown as,



,,



,,



,,



(5)

Quaternion Update:

The updated quaternion (q



) can be represented by the

combination of previous quaternion ( q



) and

updated gyroscope values (ω



), which is shown in the

following formula (James, 2006),















∗



∗∆

(6)

The corresponding accelerate signal on GC can be

obtained by multiplying the accelerate signal on SC

by the transformation matrix which is composed by

quaternion.

The transformed accelerate signal on x axis is

shown as below,

Figure 4: Original and fused accelerate signals on X axis.

0 500 1000 1500 2000 2500

Sampling numbers

-100

-80

-60

-40

-20

100

Periods Divisions

Gyroscope Signal

Fused Accelerate Signal

Accelerate Signal(m/s

)

Estimation of Gait Parameters based on Motion Sensor Data

133

3 RESULTS AND DISCUSSIONS

3.1 Performance of Up and Down

Function

10 participants who signed the Informed Consent

Form were randomly selected to do the TUGT test,

which means the external factors such as height and

weight were not considered. Each of them was asked

to do the TUGT test for 5 times. The calculated

outputs were recorded as the mean values of the 5

times’ tests of each participant. The reference

outputs were mean values recorded by a second

chronograph. The comparison was shown in the

following Bland-Altman figure,

Figure 5: Bland-Altman figure of the standing up time.

The up and down time obtained by the two

methods mentioned above showed a good consistency.

However, the up and down time was not an absolutely

objective value, the standard of border of up and

down periods was without an authoritative standard.

Moreover, there might be errors during recording

process.

3.2 Performance of Turn Back

Function

During the 10 groups of up and down tests introduced

above, the turn back process was also filmed. The

videos were replayed and the turn back time was

recorded as reference. The comparison was shown in

the following Bland-Altman figure, which

demonstrated the consistency is good.

Figure 6: Bland-Altman figure of the Turning back time.

3.3 Performance of Walking Function

According to walking function, the outputs were

based on the step length factor. A contrast test was

designed to figure out the difference between the

proposed method and gold standard. A Vicon device

with the pressure trail was used to obtain relatively

precise reference stride length values as gold standard.

10 Participants were asked to do the 10MWT test and

the accurate rate was defined as the absolute value of

1- calculated values/recorded values. The comparison

was shown as below:

Table 1: The comparison between calculated and recorded

values of stride length.

Left / Right

Leg

Mean Value of Stride

Length (m)

Accurate

Rate

Calculated Recorded

1 1.55/1.86 1.5/1.73 89.6%

2 1.59/1.66 1.5/1.57 91.4%

3 1.51/1.41 1.53/1.53 95.4%

4 1.82/1.71 1.55/1.41 83.8%

5 1.04/1.65 1.00/1.48 92.9%

6 1.22/1.42 1.26/1.32 94.9%

7 1.16/1.48 1.25/1.35 92.0%

8 0.78/1.53 0.92/1.38 87.5%

9 1.43/1.29 1.29/1.32 94.0%

10 1.16/1.33 1.20/1.20 93.5%

Calculated time-Reference time(s)

BIODEVICES 2020 - 13th International Conference on Biomedical Electronics and Devices

134

The results showed the mean accuracy rate is

around 91.5% and thus the error rate is around 8.5%.

It is interesting to make a comparison with the result

reported in (Benoît et al., 2015), obtained on a similar

database collected from inertial sensors during

walking process. On one hand, the error rate was

around 5.5%, which got better results than the

promoted algorithm. On the other hand, the promoted

algorithm did not have a strict sensors wearing

requirement.

4 CONCLUSIONS

This paper represented a system which could assess

the body condition of normal people as well as

Parkinson’s and Stroke’s patients based on the data

collected from IMU placed on two ankles and the

center of waist. New motion periods division method

and data fusion method were discussed. The

validation was conducted by comparing the

calculated values with the reference values which

were collected from camera videos and Vicon device.

The results showed that the three functions processed

the motion signal and calculated the output values

precisely. At present, the system was tested by normal

people, however, in the near future, the system can be

applied to patients and help them recover from

diseases.

REFERENCES

Hanson M, Powell H, Barth A, et al., 2009, Body area

sensor networks: Challenges and opportunities, IEEE

Computer, 42(1):58－65.

Mov Disord, 2003, Movement disorder society task force

on rating scales for Parkinson's disease, The unified

parkinson's disease rating scale (UPDRS): status and

recommendations, 18:738．

Sebastian Madgwick, 2010, An efficient orientation filter

for inertial and inertial/magnetic sensor arrays (internal

report on his thesis and MARG).

Marilyn K. Hart, Ph.D. & Robert F. Hart, Ph.D, 2007,

Statistical process control Techniques, Statit Software,

Inc., 1128 NE 2nd Street, Corvallis, Oregon 97330.

Koichi S, Hikaru I, Yut Aka S, Non-restricted measurement

of walking distance, Proceedings of 2000 IEEE

International Conference on System, man, and

Cybernetics. Nashville: IEEE, 2000: 1847-1852.

Reza Olfati-Saber, 2009, Kalman-Consensus Filter:

Optimality, stability, and performance. Proceedings of

the 48h IEEE Conference on Decision and Control

(CDC) held jointly with 2009 28th Chinese Control

Conference.

James Diebel, 2006, Representing Attitude: Euler Angles,

Unit Quaternions, and Rotation Vectors, Stanford

University, Stanford, California 94301–9010.

Benoît S, Mourad B, Jennifer D, Roger P-G, Christian G, et

al., Implementation and Validation of a Stride Length

Estimation Algorithm, Using a Single Basic Inertial

Sensor on Healthy Subjects and Patients Suffering from

Parkinson's Disease, Electronic Healthcare, 2015,

704-714.

Estimation of Gait Parameters based on Motion Sensor Data

135