Algorithm for Extracting Initial and Terminal Contact Timings

during Treadmill Running using Inertial Sensors

Laura Prijot

, Cédric Schwartz

, Julien Watrin

, Alex Mendes

, Jean-Louis Croisier

Bénédicte Forthomme

, Vincent Denoël

, Olivier Brüls

and Mohamed Boutaayamou

1,4

Laboratory of Human Motion Analysis, University of Liège (ULiège), Liège, Belgium

ECAM Brussels Engineering School, Brussels, Belgium

Université de Reims Champagne-Ardenne, Reims, France

INTELSIG Laboratory, Department of Electrical Engineering and Computer Science, ULiège, Liège, Belgium

Keywords: Running, Gait, Algorithm, Concurrent Validation, Initial Contact, Terminal Contact, Temporal Events, Stride,

Stance, Swing, Gyroscope, Accelerometer, IMU.

Abstract: Inertial measurement units (IMUs) are now considered as an economical solution for long term assessment in

real conditions. However, their use in running gait analysis is relatively new and limited. Detecting the timing

at which the foot strikes the ground (initial contact, IC) and the timing at which the foot leaves the ground

(terminal contact, TC) gives access to many relevant temporal parameters such as stance, swing or stride

durations. In this paper, we present an original algorithm to extract IC and TC timings and associated

parameters from running data. These data have been measured using a newly developed IMU-based hardware

system in ten asymptotic participants who ran at three speeds (slow, normal, and fast) with different running

patterns (natural, rearfoot strike, mid-foot strike, and forefoot strike). This algorithm has been validated

against a 200 Hz video camera based on 7056 IC and TC timings and 6861 temporal parameters. This

algorithm extracted ICs and TCs with an accuracy and precision of (median [1

quartile; 3

quartile]) 5 ms [-

5 ms, 15 ms] and 0 ms [-5 ms, 5 ms], respectively. The relative errors in the extraction of stride and stance

durations are -1.56 ± 3.00% and 0.00 ± 1.32%, respectively.

1 INTRODUCTION

Quantitative analysis of running is of critical interest to

the sports science field. For example, this analysis can

give insight into aetiology or treatment and recovery of

running injuries. In the same manner, it can help sports

coaches to improve the performances of their athletes.

Initial contact (IC) and terminal contact (TC) are key

timings in running: IC occurs at landing when the foot

initiates contact with the ground while TC is when the

foot ends contact. From these two key timings, it is

possible to compute relevant temporal parameters,

such as stance, swing or stride durations.

The stance phase, also known as the ground-

contact phase, starts at the foot IC and ends at TC. The

swing phase starts at TC and ends at the next IC.

Finally, a stride phase is the duration between two

ipsilateral ICs. Temporal parameters are related to

running performances: for instance, a shorter contact

time is linked to a good running economy and a faster

speed (Weyand, 2000).

Traditionally, timings are detected by using force

platforms. Nevertheless, these systems can only be

used in controlled laboratory environments where the

capture volume could be limited to a few steps.

The rapid technological advances in micro-electro-

mechanical systems have allowed the inertial

measurement units (IMUs) to become light, small, and

relatively cheap. Due to their portability and low power

consumption, IMU-based systems allow obtaining real

condition data.

IMUs have shown to give accurate and reliable

information on walking (Boutaayamou et al., 2015 and

2016). However, running differs from walking. As the

speed increases, the double support phase (both feet

simultaneously touching the ground) of the walking

gait cycle is replaced by a double swing phase, where

both feet are in the air. Indeed, by definition, someone

is running if both feet are never simultaneously

touching the ground. Moreover, when walking, people

are usually landing on their heel first. However, during

running, there are three possible different landing

strategies: rearfoot strike (RFS), mid-foot strike

258

Prijot, L., Schwartz, C., Watrin, J., Mendes, A., Croisier, J., Forthomme, B., Denoël, V., Brüls, O. and Boutaayamou, M.

Algorithm for Extracting Initial and Terminal Contact Timings during Treadmill Running using Inertial Sensors.

DOI: 10.5220/0008983402580265

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 4: BIOSIGNALS, pages 258-265

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

(MFS), and forefoot strike (FFS). Compared to

walking, the biomechanics involved in running is also

different: a wider range of motion of all the lower limb

joints, higher impact forces, and higher eccentric

muscle contraction (Nicola et al., 2012).

The use of IMU sensors in running gait analysis is

relatively new. In the literature, different localisations

for IMU sensors are considered such as trunk

(Bergamini et al., 2012) or tibia (Purcell et al., 2006).

Among all existing studies, only a few of them

include a concurrent validation of their algorithm

using a reference system. Both Chew et al. (2017) and

Falbriard et al. (2018) used the signal of an IMU

placed on the dorsal side of the foot to compute ICs

and TCs. The first one used a threshold-based

method, while the second one compared different

algorithms. However, to the authors’ knowledge,

there is no study available using foot-worn IMU

sensors that take into account the different existing

landing strategies.

In this work, we present a newly developed

algorithm to extract IC timing and TC timing

extracted from IMU signals measured at the level of

the foot (toe and heel). From these timings, the

ipsilateral stance, swing, and stride durations are

computed. This algorithm is tested on data obtained

from ten healthy participants running at steady speeds

on a treadmill. Furthermore, we validated this

algorithm against synchronously recorded reference

data obtained from a frame-by-frame analysis of 2D

high-speed (200 Hz) videos.

2 METHOD

2.1 Participants and Treadmill

Running Setting

In total, ten asymptotic participants (7 men and 3

women), who were regularly active at the time of the

tests, were volunteered for this study. The set of

participants includes both recreational and professional

runners. They were all informed with the procedure

and they have all signed an informed consent.

Table 1 shows the anthropometric characteristics

(mean ± standard deviation (STD)) of these

participants. Among them, seven were naturally RFS

while two were MFS, and one was FFS.

Each participant was equipped with an IMU-based

hardware system (Boutaayamou et al., 2019)

integrating three-axis accelerometers (range: ±16 g)

and three-axis gyroscopes (range: 2000 deg/s). This

system includes an acquisition box (memory, micro-

controller, and battery) linked by wires

to four small

Table 1: Anthropometric characteristics of the participants

measured at the time of the test.

Mean ± STD

e [

ears] 26.1 ± 3.9

Hei

ht [cm] 179.3 ± 11.4

Bod

mass [k

] 70.0 ± 12.3

IMU sensors (2.1 × 1.0 × 0.8 cm, weight = 16 g).

Consequently, it is portable with an autonomy of 4h30.

The IMU acquisition frequency is 200 Hz. No

restrictions on the shoes were imposed, to enlarge the

range of applications of the algorithm.

The sensors were directly attached to the right shoe

at the level of the first distal phalange (toe), calcaneus

(heel), the fifth metatarsal, and dorsal side of the foot.

In this work, only the toe and the heel sensors will be

considered. The fixation procedure used has been

validated in the case of walking (Boutaayamou et al.,

2015) and shows satisfying results for running gait

analysis.

The three-dimensional linear acceleration signals

[m/s

] are denoted by 𝑎



, 𝑎



, and 𝑎



, while the three-

dimensional angular velocity signals [deg/s] are

denoted by 𝜔



, 𝜔



, and 𝜔



along sensitive axes

represented schematically in Figure 1 .

Each test began with a standardized time to warm

up and to become familiar with the treadmill and

instrumentation system (during approximately five

minutes). At the same time, a preferential running

speed (PRS) is selected with the participant, at which

he should be able to run during ten minutes without

loss of intensity. The PRS (mean ± STD) of the

volunteers is 8.3 ± 1.3 km.h

−1

. During the tests, they

were asked to run at three different speeds: slow

(computed by PRS−0.25×PRS), normal (PRS), and

fast (computed by PRS+0.25×PRS). At each speed, the

participants performed six trials (in the following

order): three with a natural foot strike pattern, one

rearfoot strike (RFS), one mid-foot strike (MFS), and

one forefoot strike (FFS). In total, the participants were

asked to perform 18 trials of 60 s. The minimum total

test duration was 69 minutes per participant, including

3 minutes of rest between each trial. All running tests

were performed at the Laboratory of Human Motion

Analysis (University of Liège, Belgium), on a

treadmill (SportsArt T650). At the same time, all the

trials were recorded using a 2D high-speed video

camera (Basler Pilot) with a sampling frequency of 200

Hz. This video camera will be used as the reference

system. Signal and data processing were carried out

using the software Matlab

(R2017a, Mathworks,

Natick, MA, USA).

Algorithm for Extracting Initial and Terminal Contact Timings during Treadmill Running using Inertial Sensors

259

2.2 Extraction Algorithm of IC and TC

Timings

The proposed algorithm first computes an estimated

IC based on the average stride duration. Then, an

exact IC timing is obtained from the different linear

accelerations. Subsequently, TCs are found between

two successive ICs.

The first step is to obtain an estimated average

stride duration, 𝑑



[

𝑠

]

, based on the Fourier Fast

Transform of the heel angular velocity signal (heel

𝜔



). The first peak, which is also the highest,

corresponds to the stride frequency [Hz]. 𝑑



is,

then, obtained from this stride frequency using the

following formula

𝑑



stride frequenc

(1)

Alternatively, 𝑑



can be obtained from the

auto-correlation of the same signal. In that case, the

positive lag corresponding to the first positive local

maxima after 0 is the average 𝑑



(available in

Matlab

using the function xcorr).

After computing 𝑑



, estimated ICs are

obtained in the filtered heel 𝜔



signal. The filter used

is a high pass Butterworth filter of order 4 with a cut-

off frequency of 15 Hz. A high pass filter allows to

remove the movement components of the signal and

to keep only the shock parts. Estimated ICs can then

be obtained by detecting a minimum in the filtered

heel 𝜔



. The distance between two successive

minima is imposed to be of at least 85% of 𝑑



allowing for small variations of 𝑑



at each stride.

Potential exact ICs are obtained by looking for

local extrema, in a time window around the estimated

IC, in different linear acceleration signals of both

sensors. Namely, the algorithm is looking for: a local

minimum in toe 𝑎



, local minimum in toe 𝑎



, local

minimum in heel 𝑎



, and local maximum in heel 𝑎



in the time window [-20 ms; 5 ms] around the

estimated IC.

Then, the exact IC corresponds to the first time

instant among all these extrema. The acceleration

signals in the transverse direction (Y-axis) are not

considered

since they are runner dependent. For

instance, they can be influenced by foot movements

like supination or pronation.

As the tip is always the last part of the foot in

contact with the treadmill, TCs will be detected using

the toe sensor. The toe total acceleration in the sagittal

plane, given by has shown the highest accuracy.

Figure 1: Schematic illustration of the position of the IMU

sensors used in the proposed algorithm, including the three

local axes (X-axis, Y-axis, and Z-axis). The two sensors are

placed on the right shoe at the level of the first distal phalange

(toe) and at the calcaneus (heel).

𝑎





𝑎



+𝑎



)

TCs are determined based on the intuitive

principle that there is always a TC between two

successive ICs. Hence, for each stride i, a TC(i) will

be searched in the time window between IC(i) and

IC(i+1). This window can be further reduced to

increase the accuracy of the event extraction method.

The upper bound of the time interval can be obtained

based on the definition of running: someone is

running if there is a double float phase, where both

legs are in the swing phase simultaneously. This is

only possible if the stance phase lasts for less than

50% of the stride duration.

Hence, the upper limit is defined as follows

𝐿𝑖𝑚



=𝐼𝐶

(

𝑖

)



(



)

()



(3)

This limits the application of the algorithm to only

running cases. However, this improves the accuracy.

In fact, in some cases, the acceleration linked to the

swing movement of the foot is higher than the shock

corresponding to the TC (see Figure 2).

Furthermore, the lower bound of the time window

(𝐿𝑖𝑚



) is obtained using the entropy of the signal.

During the stance phase, the foot has a constant

acceleration and this signal flat zone is characterized

by a low entropy. Hence, the lower limit is obtained

by computing the entropy over a sliding window. The

size of the window has been determined empirically:

on one side, it should be as small as possible to have

good local information. On the other side, it

must be

large enough to not detect the flatter zone that

appears for some runners after the toe-off peak.

This was generally a problem for FFS running

patterns. A window of 15 samples (i.e., 75 ms) has

shown good results for all participants.

𝑇𝐶(𝑖) is then determined by finding a maximum

in the toe sagittal acceleration signal over the time

window : [𝐿𝑖𝑚



(

𝑖

)

; 𝐿𝑖𝑚



(𝑖)].

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

260

Figure 2: Determination of TCs using the toe sagittal acceleration signal. TC is found between two successive ICs. An upper

limit can be obtained using the definition of running: the stance duration must be less than 50% of the stride duration. This

prevents to wrongly detect local maximum coming from the movement acceleration. The signal flat phase can be used as a lower

limit, which can be detected using the entropy of the signal over a sliding window.

Figure 3: Determination of IC and TC timings using the 2D high-speed video camera. IC (upper pictures) is the first frame

where the pixels representing one shoe are directly in contact with those representing the belt of the treadmill. TC (lower

pictures) corresponds to the last frame where the pixels of the shoe are in contact with those of the treadmill.

2.3 Concurrent Validation and

Evaluation Methods

The reference timings are obtained from a frame-by-

frame analysis of 2D high speed videos. A precise

definition is used to select the frame corresponding to

an IC and to a TC: IC is the first frame where the pixels

representing one shoe are directly in contact with those

representing the belt of the treadmill. In other words, it

is the first frame where there are no white pixel (i.e.,

background pixel) in-between the shoe and the

treadmill.

Conversely, TC corresponds to the last frame

where the pixels of the shoe are in contact with those

of the treadmill (see Figure 3).

Finally, the different temporal parameters are

computed from IC and TC timings, as follows

𝑑



(

𝑖

)

=𝑇𝐶

(

𝑖

)

𝐼𝐶

(

𝑖

)

𝑑



(

𝑖

)

=𝐼𝐶

(

𝑖+1

)

𝑇𝐶

(

𝑖

)

𝑑



(

𝑖

)

=𝐼𝐶

(

𝑖+1

)

𝐼𝐶(𝑖).

)

The reference system has a maximum achievable

resolution of 5 ms. Additionally, at some point in the

video, there are two identical frames following each

other. In that case, a 5 ms error can also occur.

These reference timings are used to concurrently

validate the events obtained using the proposed

algorithm. For each stride, the results for (1) IC, (2) TC,

Algorithm for Extracting Initial and Terminal Contact Timings during Treadmill Running using Inertial Sensors

261

(3) 𝑑



, and (4) 𝑑



are computed. The results

for 𝑑



have been computed but are not shown in

this paper

Finally, the accuracy and precision of ICs and TCs

extraction is quantified by the mean and STD or

median and inter-quartile range (IQR) values (i.e., 1

quartile (Q

); 3

quartile (Q

)) of the differences

between these timings and the reference system,

depending on the normality of data distributions. This

is done for each participant separately and for all

participants together. The normality of data

distributions is tested using Jarque-Bera test (available

in Matlab

using the function jbtest). Additionally,

relative errors are computed as the mean of the stride-

by-stride differences between the IMU temporal

parameter and the reference temporal parameter

divided by the reference temporal parameter. These

errors are only meaningful for temporal parameters and

they will not be computed for timings.

3 RESULTS

This work focuses on running trials from the acquired

data. In some trials, particularly at low speeds, some

participants exhibited a double support phase.

Consequently, as these trials are considered as

walking trials, they have been excluded from this study.

In total, 39 trials out of 183 were not considered.

Additionally, some trials have been reclassified

according to the real running pattern observed that, in

some cases, was different from the supposed running

pattern. Indeed, some participants had difficulties in

voluntarily performing MFS or FFS.

First of all, an intra-participant comparison

between the IMU results and reference results is

carried out. In this paper, the median is used as the

data are not normality distributed. However, in

general in this study, the mean and median values and

STD values IQR ones were similar. The same

conclusion can be drawn for STD values IQR ones.

Table 2 summarizes the results for each

participant, the values have been rounded to the

sample period (i.e., 5 ms) of the hardware systems.

This analysis includes all the valid trials (different

speeds and different foot strikes) and at least 30 valid

strides per trial, when available. The number of

observations depends on the number of valid events

taken into account. In the case of ICs, the mean of the

extraction accuracies is 5 ms. Consequently, the

algorithm tends to detect the ICs one frame later than

the reference system. The mean of the extraction

precisions obtained in the case of IC is 10 ms.

Table 2: Intra-participant differences between IMU

timings/temporal parameters and reference data. It includes

the median and the interquartile range values ([1

quartile

; 3

quartile Q

]) as well as the number of observations

(nbr. of obs.).

Partici-

pants

Running

timings/

arameters

Median

; Q

]

[ms]

Nbr.

of obs.

𝑑



𝑑



0 [-10; 10]

0 [-5; 5]

5 [-5; 15]

0 [-10; 10]

211

205

191

𝑑



𝑑



-10 [-20; 0]

0 [-5; 5]

10 [-5; 25]

0 [-10; 10]

306

291

𝑑



𝑑



0 [-20; 20]

0 [-5; 5]

5 [-15; 25]

0 [-10; 10]

437

403

𝑑



𝑑



5 [0; 10]

0 [0; 10]

-5 [-10; 0]

0 [-10; 10]

366

365

348

𝑑



𝑑



10 [5; 15]

0 [-5; 5]

-10 [-15; -5]

0 [-10; 10]

484

394

385

462

𝑑



𝑑



10 [5; 15]

0 [-5; 5]

-10 [-15; -5]

0 [-5; 5]

332

314

𝑑



𝑑



15 [5; 25]

0 [-5; 5]

-15 [-30; 0]

0 [-10; 10]

129

125

𝑑



𝑑



0 [-15; 15]

0 [-5; 5]

0 [-20; 20]

0 [-15; 15]

312

213

303

𝑑



𝑑



10 [5; 15]

0 [-5; 5]

-5 [-15; 5]

0 [-10; 10]

468

450

448

451

𝑑



𝑑



5 [-15; 25]

0 [-5; 5]

-5 [-25; 15]

0 [-10; 10]

608

571

567

586

The worst-case for the IC determination appears for

participants 3 and 10, with an IQR of 20 ms away

from the median value. The best case is for participant

4 with a median error of one sample (i.e., 5 ms) and

IQR of 5 ms around this median error. In that case,

the precision obtained exactly corresponds to the

maximum achievable precision. Indeed, the

maximum precision depends on the sampling rate of

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

262

both the IMU system and the high-speed video, as

well as the 5 ms error than can be explained by errors

in the reference system.

In the case of TCs, the results obtained with IMU

are similar to those obtained with the reference

system. Indeed, the mean of median errors between

the two systems is 0 ms and the mean of the IQRs is

5 ms, for all participants. Therefore, the algorithm can

detect TCs with the maximum possible accuracy.

For the stance duration, the algorithm tends to

underestimate the duration compared to the reference

values. This can be explained by the fact that ICs are

generally detected later with the algorithm. The mean

of the median values is -3 ms, which is less than one

sample of difference and less than the maximum

accuracy. The mean of the variability values is 12.5 ms,

this is slightly higher than the maximum precision

expected. Indeed, IC(i) can be determined with a

maximum precision of 5 ms and TC(i) can also be

determined with a maximum precision of 5 ms. As the

errors may cumulate, a maximum precision of 10 ms is

expected for durations. However, the variability is of 5

ms for three participants out of ten.

Table 3: Inter-participant comparison including the

extraction accuracies and precisions of 7 participants,

running with their preferential running style at speeds

ranging from 7.1 to 9 km.h

-1

Running

timings/

parameters

Mean ±

STD

[ms]

Median

; Q

]

[ms]

Median

; Q

]

[%]

IC 5 ± 9 5 [-5; 15] /

TC 1 ± 4 0 [-5; 5] /

𝑑



-5 ± 15 -5 [-15; 5]

-1.56

[-4.56; 2.56]

𝑑



0 ± 10 0 [-10;10]

0.00

[-1.32; 1.32]

The accuracy for the stride duration is 0 ms for all

participants, which is the best possible achievable

accuracy. The precision obtained is on average 10 ms,

which is the expected precision, as explained before.

Note that, one participant out of the ten has a better

precision (5 ms) and only one participant has a worst

precision (15 ms) than the one expected.

Finally, an inter-participant comparison is done,

including only the natural foot strike trials at PRS

condition (i.e., three trials per participant). We did not

include the trials performed by three participants as at

least one of the three above mentioned trials was not

valid. The speed of the trials considered was between

7.1 and 9.0 km.h

-1

Table 3 provides the mean ± STD as well as the

median and IQR values of the differences between the

extracted IMU values and the reference values. The

mean errors and the median errors are similar for all

timings and temporal parameters. The extraction

precisions expressed in terms of STD are identical to

those expressed in terms of IQR except for the stance

duration, where the STD value is influenced by some

outlier values.

The extraction accuracy in the case of ICs is 5 ms.

The algorithm tends, then, to detected ICs one sample

later than those extracted by the reference system. This

could be explained by the fact that IMUs will detect the

interaction (shock) between the shoe and the belt of the

treadmill while, in the video, the selected frame is the

one when the shoe and the treadmill touch each other

but have not yet interacted. The precision obtained for

the ICs is 10 ms, which is one frame higher than the

maximum achievable precision. On the other side, TCs

are extracted with both the maximum achievable

accuracy (i.e., 0 ms) and precision (i.e., 5 ms).

The stance durations tend to be underestimated by

the algorithm. On average, they are 5 ms shorter than

those obtained with the reference system. Again, this is

explained by the fact that ICs have a tendency to be

detected 5 ms later with the IMUs. Finally, the stride

duration, which only depends on successive ICs, are

extracted with the best possible accuracy (i.e., 0 ms)

and a precision equal to the maximum expected

precision due to the accumulation of errors. Indeed,

there could be 5 ms of error in the first IC (IC(i)) and 5

ms of error for the successive IC (IC(i+1)). All in all,

it can be seen that the inter-participants and intra-

participant comparison give similar results.

It is also interesting to express the errors in both

stance and stride duration estimates as a percentage of

the total duration. The 𝑑



relative error is (median

, Q

]) -1.56 % [-4.56 %; 2.56 %] and the maximum

relative error is -9.52 %. The 𝑑



relative error is

(median [Q

, Q

]): 0.00 % [-1.32 %; 1.32 %] and the

maximum computed error is 4.49%.

4 DISCUSSION

This article presents an original algorithm to extract

the two main timings (ICs and TCs) at different

running speeds (slow, normal, and fast) and with

different running styles (natural, RFS, MFS, and

FFS). The data collected for this work are obtained

using two IMU sensors placed on regular shoes at the

level of the heel (calcaneus) and toe (first distal

phalange).

Algorithm for Extracting Initial and Terminal Contact Timings during Treadmill Running using Inertial Sensors

263

Only the right shoe has been used in this work.

However, the algorithm is supposed to work in the

same way for the left foot. Additionally, the IMU

hardware system used here can record the data of up

to four sensors at the same time. It is thus possible to

record the data of both legs simultaneously.

Therefore, it would be possible to obtain other

parameters, such as the step duration. Besides, it would

be possible to make a comparison between the two

legs, which has a wide range of applications, including

monitoring recovery after injury or surgery.

The performance of the algorithm is determined

by a concurrent validation with 2D high-speed

videos, recorded simultaneously. The algorithm

presented here has been concurrently validated using

a total of 7056 timings and 6861 temporal parameters.

This comparison has shown a good agreement

between timings obtained using the IMU signals and

timings detected on the 2D videos. The measures

include running speeds ranging from 6.0 to 11.3 km.h

. The obtained global extraction accuracy and

precision (median [Q

; Q

]) is 5 ms [-5 ms; 15 ms] and

0 ms [-5 ms; 5 ms] for, respectively, ICs and TCs.

Besides, the accuracy and precision for the stance

durations and stride durations (median [Q

; Q

]) are -5

ms [-15 ms; 5 ms] and 0 ms [-10 ms; 10 ms],

respectively. This corresponds to a relative error of

respectively -1.56 ± 3.00% and 0.00 ± 1.32%.

The stride duration average error obtained here

(i.e., 0 ms) is consistent with the one measured by

Chew et al. (2018), which is between -0.44 ms and

0.33 ms. However, Chew et al. (2018) used an

algorithm based on a thresholding-method that relies

on experimental values needed to determine the

threshold. This is not the case for the algorithm

presented here. Similarly, the stance duration errors

are similar to those found by Purcell et al. (2006).

They found an error (mean ± STD) of 0 ± 12 ms and

−2 ± 3 ms, depending on the running speed, using a

tibial accelerometer. However, they used a force

platform with higher accuracy than the 2D video

system used here. Falbriard et al. (2018) found better

accuracy and precision (median [Q

, Q

]): 2 [1 ms, 3

ms]

for IC and only a better precision for TC (4 ms [2

ms, 6 ms]). Nevertheless, this precision cannot be

achieved here with the 200 Hz reference system used.

The algorithm presented here is only valid for

steady state running over a treadmill. Walking cases

cannot be analysed using the present method,

however, there exist algorithms to detect the type of

activity (walking, running, and rest). Once the

activity is appropriately determined, either a walking

or a running algorithm can be selected to extract

temporal events.

5 CONCLUSION

In this article, we presented an original algorithm to

extract timings (IC and TC) in the case of steady-state

running over a treadmill, using IMU sensors. From

these two timings, three temporal parameters can also

be computed: stance, swing, and stride durations. The

method developed here has the following advantages:

- The sensors are placed on the shoes and not

directly on the feet, which allows running in

many different conditions.

- The algorithm only uses two IMU sensors per

foot: one at the level of the heel and the other at

the level of the first distal phalange (toe).

Additionally, only one sensor (i.e., toe sensor) is

used to determine TCs with the maximum

achievable precision and accuracy.

- This method has been concurrently validated

using a 2D high-speed video camera as the

reference system.

- The analysis is done over a large number of

strikes including a wide range of running speeds

(from 6 km.h

-1

to 11.3 km.h

-1

) and different

running styles (natural, RFS, MFS, and FFS).

The results showed that it is possible to achieve

acceptable accuracy and precision using a foot-worn

IMU-based system. These results are encouraging for

the use of IMU for daily and out-of-lab monitoring.

They can be seen as a good trade-off between

expensive and laboratory-limited measurement

instruments like force platforms that show high

accuracy and wearable systems that can be found in

smartwatches or in smartphones.

Future researches may focus on the use of a single

IMU sensor to extract the timings and associated

temporal parameters or on the detection of spatial

parameters like the stride length. Further work could

also focus on extracting the durations of the stride

sub-phases.

ACKNOWLEDGEMENTS

The authors would like to thank all the participants who

were volunteers for this study. We would also like to

thank the reviewers for their relevant comments, which

will be taken into consideration in future articles.

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

264

REFERENCES

Amini, N., Vahdatpour, A., Wenyao Xu, and Sarrafzadeh.,

M., 2011. Accelerometer-based on-body sensor

localization for health and medical monitoring

applications. Pervasive Mobile Computing, 7:746–760

Bergamini, E., Picerno, P., Pillet, H., Natta, F., Thoreux, P.,

and Camomilla, V., 2012. Estimation of temporal

parameters during sprint running using a trunk-

mounted inertial measurement unit. Journal of

Biomechanics, 45:1123–1126.

Benson, L., Clermont, C., Bošnjak, E., and Ferber, R., 2018.

The use of wearable devices for walking and running

gait analysis outside of the lab: A systematic review.

Gait & Posture, 63.

Boutaayamou, M., Denoël, V., Brüls, O.,Demonceau, M.,

Maquet, D., Forthomme, B., Croisier, J.-L., Schwartz,

C., Verly, J., and Garraux. G., 2016. Extraction of

temporal gait parameters using a reduced number of

wearable accelerometers.

Boutaayamou, M., Denoël, V., Brüls, O., Maquet, D.,

Forthomme, B., Croisier, J.-L., Schwartz, C., Verly, J.,

Macq, B., Garraux, G., and Stamatakis, J., 2015.

Development and validation of an accelerometer-based

method for quantifying gait events. Medical

Engineering and Physics, 37:226–232

Boutaayamou, M., Schwartz, C., Joris, L., Forthomme, B.,

Denoël, V., Croisier, J.-L., Verly, J. & Garraux, G., and

Brüls, O., 2019. Adaptive Method for Detecting Zero-

Velocity Regions to Quantify Stride-to-Stride Spatial

Gait Parameters using Inertial Sensors. 12th

International Conference on Bio-inspired Systems and

Signal Processing, 229-236.

Chew, D.-K., Ngoh, K., Gouwanda, D., and Gopalai., A.,

2017. Estimating running spatial and temporal

parameters using an inertial sensor. Sports

Engineering, 21(2):115–122.

Falbriard, F., Meyer, B., Mariani, G., Millet, P., and

Aminian, K., 2018. Accurate estimation of running

temporal parameters using foot-worn inertial sensors.

Frontiers in Physiology.

Hanley, B., Bissas, A., and Drake, A., 2015. The

contribution of the flight phase in elite race walking.

33rd International Conference on Biomechanics in

Sports, At Poitiers, France

Koska, D., Gaudel, J., Hein, T., and Maiwald, C., 2018.

Validation of an inertial measurement unit for the

quantification of rearfoot kinematics during running.

Gait & Posture, 64:135–140.

Maiwald, C., Sterzing, T., Mayer, T.A., and Milani, T.L.,

2009. Detecting foot-to-ground contact from kinematic

data in running. Footwear Science, 1(2):111–118.

Mariani, B., Rouhani, H., Crevoisier, X., and Aminian, K.,

2013. Quantitative estimation of foot-flat and stance

phase of gait using foot-worn inertial sensors. Gait and

Posture, 37(2):229 – 234.

Nicola, T. and Jewison, D., 2012. The anatomy and

biomechanics of running. Clinics in Sports Medicine,

31:187-201.

Nicolai, L., Mifsud, N., Kristensen, H., Villumsen, M.,

Hansen, J., and Kersting, U., 2014. Portable inertial

motion unit for continuous assessment of in-shoe foot

movement. Procedia Engineering, 72:208–213.

Purcell, B., Channells, J., James, D., and Barrett, R., 2006

Use of accelerometers for detecting foot-ground

contact time during running. Proc. SPIE

Weyand, P.G., Sternlight,, D. B., Bellizzi, M. J., and

Wright, S., 2000. Faster top running speeds are

achieved with greater ground forces not more rapid leg

movements. Journal of Applied Physiology,

89(5):1991–1999.

Wixted, A., Thiel, D., Hahn, A., Gore, C., Pyne, D., and

James, D., 2007. Measurement of energy expenditure in

elite athletes using mems-based triaxial

accelerometers. Sensors Journal, IEEE, 7:481–488.

Yang, S., Mohr, C., and Li, Q., 2011. Ambulatory running

speed estimation using an inertial sensor. Gait and

Posture, 34(4):462 – 466.

Algorithm for Extracting Initial and Terminal Contact Timings during Treadmill Running using Inertial Sensors

265