Heterogeneous Inertial Measurement Units in Motion Analysis

Filip Malawski

and Natalia Paruzel-Satława

Institute of Computer Science, AGH University of Science and Technology, Krakow, Poland

Keywords:

IMU, Motion Analysis, Accelerometer, Sensor Fusion, Heterogeneous Devices, Signal Processing.

Abstract:

Inertial measurement units are commonly used in motion analysis applications, such as sports training aid, gait

analysis, medical diagnosis, or rehabilitation assistance. Linear acceleration and orientation obtained from

sensor fusion are employed for the detection and classiﬁcation of actions, as well as for measuring relevant

parameters of the motion. Typically, in multi-sensor setups, a single model of the device is used. However,

considering potential end-users, it could be beneﬁcial to allow heterogeneous setups, particularly by including

everyday-use devices with built-in inertial sensors, such as smartwatches. In this work, we perform experi-

ments with several different sensors in order to analyze agreement in their measurements. Results indicate that

devices of different models are not directly interchangeable, however, in some applications, heterogeneous se-

tups may be viable.

1 INTRODUCTION

Inertial measurement units (IMU) measure accelera-

tion and angular velocity, as well as, in the case of

most such devices, magnetic ﬁeld. Moreover, a fu-

sion of those data modalities provides an estimation

of 3D orientation, relevant for many motion analysis

applications. A number of methods for motion analy-

sis have been proposed in the literature (Lopez-Nava

and Munoz-Melendez, 2016), however it is a com-

mon practice, that the data for the experiments are

collected with homogeneous devices. From a prac-

tical point of view, a question arises - once we sub-

stitute an IMU with a different model will the pro-

posed methods work the same? Moreover, what if

we were to use a heterogeneous setup, with multi-

ple IMU models working in parallel? Recently, more

and more everyday-use devices have built-in IMUs

(e.g. smartphones, smartwatches, ﬁtness bands). Em-

ploying those as part of a multi-sensor motion anal-

ysis system would be beneﬁcial for the end user, as

it would reduce the number and therefore the cost of

additional devices needed to purchase in order to use

such a system. In this work, we perform a series of

experiments to verify the potential interchangeability

of different IMU sensors in motion analysis applica-

tions, including two dedicated IMU models and one

smartwatch.

https://orcid.org/0000-0003-0796-1253

2 RELATED WORKS

IMUs are commonly used for motion analysis in ap-

plications such as sports action recognition and eval-

uation, gait analysis, as well as rehabilitation exercise

assistance. In sports, IMUs are used in a wide range

of disciplines, both practiced indoors, such as com-

bat sports (Worsey et al., 2019a), volleyball (Wang

et al., 2018) or basketball (Ren and Wang, 2021),

and outdoors, such as rowing (Worsey et al., 2019b)

or football (Wilmes et al., 2020). Aside from event

detection and classiﬁcation, depending on the disci-

pline, a number of useful motion parameters are ex-

tracted, including velocity, acceleration, cycle length,

amplitude, angular displacement, force, and others

(Camomilla et al., 2018).

Gait analysis is often a useful tool in medical di-

agnosis, including degenerative joint diseases such as

osteoarthritis (Kobsar et al., 2020), as well as neu-

rological impairments, such as Parkinson’s disease

(Tunca et al., 2017; Petraglia et al., 2019). Mea-

sured parameters include, among others, stride and

step length, cadence, speed, as well as cycle, stance,

and swing times (Teuﬂ et al., 2018).

In rehabilitation, motion analysis with IMUs is

employed, among others, for the classiﬁcation of

shoulder activities in guided treatment (Bavan et al.,

2019), detection of correct and incorrect performance

of lower limb exercises (Giggins et al., 2014), as

well as tracking speciﬁc exercises to aid stroke pa-

Malawski, F. and Paruzel-Satława, N.

Heterogeneous Inertial Measurement Units in Motion Analysis.

DOI: 10.5220/0011799100003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 4: BIOSIGNALS, pages 351-358

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

351

tients (Zhou et al., 2006). Angles of joint movement

are most often of interest in rehabilitation scenarios

(Milosevic et al., 2020).

While some works show successful applications

of a single IMU sensor, e.g. for knee rehabilitation

monitoring (Bevilacqua et al., 2018) or swimming

motion analysis (Wang et al., 2019), most solutions

require a multi-sensor setup. Review papers indicate,

that there are at least several popular IMU devices,

dedicated to motion analysis, available on the market

(Petraglia et al., 2019; Milosevic et al., 2020). How-

ever, closer inspection reveals, that most of them are

high-cost, advanced devices targeted for research or

professional medical and sports applications. In this

work, we focus on relatively low-cost devices, that

are affordable for a home-use scenario. It is worth

noting, that in the referenced works, all multi-sensor

solutions employ homogeneous setups, with a single

model of IMU. While there are studies focused on

the validation of IMUs against optical systems (Roell

et al., 2019; Malawski, 2020; Clemente et al., 2022),

there’s a lack of works employing multiple different

IMUs together. In this study, we consider a heteroge-

neous setup, in which multiple different devices could

cooperate, including everyday-use equipment such as

smartwatches. To the best of our knowledge, there are

no previous works performing similar analyses.

3 METHODS

3.1 Sensors

Our goal is to compare low-cost devices, that could

be used in consumer-level applications. We employ

Xsens DOT

and MBientLab MeatMotionRL

sen-

sors, both priced approx. 100 USD per device. More-

over, we also include in our experiments Samsung

Galaxy Watch 4 smartwatch

, in order to verify if

everyday-use devices equipped with IMUs could be

employed in motion analysis setups as well. Em-

ployed sensors are depicted in Fig. 1.

Xsens DOT is a small (36x30x11mm) and light

(11g) device, providing measurements of orientation,

acceleration (±16g), angular velocity (±2000°/s),

and magnetic ﬁeld (± 8 Gauss), either in real-time

streaming (up to 60Hz) or by recording to internal

memory (up to 120Hz). It is equipped with a BLE 5.0

communications module. MetaMotionRL sensor is a

similar device, with comparable size (36x27x10mm)

https://www.xsens.com/xsens-dot

https://mbientlab.com/store/metamotionrl-p/

https://www.samsung.com/us/watches/galaxy-watch4/

and weight (8.5g), providing the same set of informa-

tion, however in a more conﬁgurable manner. Avail-

able accelerometer ranges span from ±2g to ±16g,

gyroscope ranges span from ±125°/s to ±2000°/s and

magnetometer is ﬁxed at ±13 Gauss for x and y axes

and ±25 Gauss for z axis. Acceleration, angular ve-

locity, and orientation can be streamed in real-time

at 100Hz, or logged with 800Hz. MetaMotionRL

is equipped with older BLE 4.0. Both sensors have

APIs available for multiple programming languages,

supporting both desktop and mobile development of

applications. As for the Samsung Galaxy Watch 4

(40x40x10mm, 26g), very little information is pro-

vided regarding the built-in IMU. Acceleration, an-

gular velocity, magnetic ﬁeld, and orientation are all

available through Android API, as the smartwatch

runs on Wear OS - a version of Android dedicated

to wearable devices. The sampling frequency is not

given explicitly, although experiments indicate that it

is possible to obtain data with as much as 100Hz fre-

quency using Android’s SENSOR DELAY FASTEST

setting. However, due to large ﬂuctuations of sam-

pling times, it seems more reasonable to use SEN-

SOR DELAY GAME setting, which effectively results

in a 50Hz sampling rate.

Figure 1: Sensors used in the experiments, left to right:

MBientLab MetaMotionRL, Xsens DOT, Samsung Galaxy

Watch 4.

3.2 Acquisition

In our experiments, we employ a total of four sen-

sors: two Xsens DOT devices, one MetaMotionRL,

and one Galaxy Watch 4. This allows us to measure

and compare the similarity of signals acquired with

sensors of the same model, and of different models.

In order to obtain comparable measurements of the

same motion, the sensors are mounted on top of each

other, along the shorter z-axis, with separation layers

made of technical foam, and strong non-elastic tape

used to hold them together. Therefore, all sensors

are moved as a single bundle. All devices are cali-

brated as indicated by their producers. For the Xsens

DOT, we use acquisition software provided with the

sensors, for the MetaMotionRL we use custom soft-

ware employing programming libraries provided with

the sensor, and for the Galaxy Watch 4 we use custom

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

352

Android-based software.

We consider two experimental scenarios. In the

ﬁrst one, we focus on acceleration in linear motion.

The sensor bundle is moved, separately, in three direc-

tions (left-right, up-down, forward-backward), with-

out rotations. The sequence is repeated three times

with different speeds, to collect data points in vari-

ous parts of the available measurement range. We

record so-called linear acceleration, which is accel-

eration with removed gravity - all employed devices

have a built-in capability of such ﬁltering.

The second scenario is focused on orientation.

The sensor bundle is rotated, in two directions, in

half-circle arcs with a radius of approx. 50cm. The

motion is repeated at 3 different speeds. We record

Euler angles, provided by each sensor’s built-in fu-

sion algorithm, as those are easier to interpret than

quaternions. Fusion makes use of all raw signals -

acceleration, angular velocity, and magnetic ﬁeld.

3.3 Synchronization

Synchronization of data is one of the key require-

ments in a multi-sensor setup. This is typically

done by synchronizing internal clocks and then pro-

viding a timestamp for each measured data point

(Khediri et al., 2012). Both Xsens DOT and Meta-

MotionRL sensors have built-in synchronization pro-

tocols, which could be employed in a homogeneous

setup. For heterogeneous setup, implementing clock

synchronization on multi-vendor devices would re-

quire low-level programming of acquisition software

for each device. Instead, for our experiments, we

employ signal-based synchronization, which handles

clock drift between sensors using the acquired data,

rather than communication between devices. Each

recording is started with the sensor bundle in a sta-

tionary position, then a short-term force is applied (a

single push), then for a few seconds the sensors are

left stationary again, then the actual movement for the

experiments starts. At the end of each recording sim-

ilar protocol is followed (a single push with no move-

ment before and after). A single push is easily iden-

tiﬁed as a single peak in the acceleration signal (see

Fig. 2 - ﬁrst and last peak of each signal). Therefore,

we can synchronize the data from all sensors, by se-

lecting only fragments between the ﬁrst and the last

peak, with a given threshold for the peak height.

This procedure ensures, that in the trimmed data,

the start and end time points are the same for each sen-

sor. However, we still need to address different sam-

pling rates and spacing of the timestamps. DOT and

MetaMotionRL provide data points evenly distributed

in time, however, Android API does not. Therefore,

data points for Galaxy Watch 4 were determined in

evenly distributed time points by taking the weighted

average of the nearest measured data points. In order

to obtain a common sampling rate, we resample the

MetaMotionRL and Galaxy Watch 4 data to 60Hz,

which is native to the DOT sensors. Finally, we re-

move leading and trailing close-to-zero signal frag-

ments, which result from maintaining a stationary po-

sition near the synchronization peaks (see Fig. 3).

Figure 2: Acceleration (single axis) acquired at the same

time with two different IMUs - before synchronization.

Figure 3: Acceleration (single axis) acquired at the same

time with two different IMUs - after synchronization.

3.4 Processing

For better comparison of the sensors, we apply pro-

cessing to the raw acquired data. The acceleration sig-

nal is very noisy, therefore it is a common approach to

apply ﬁltering methods to the raw data. We decided to

use basic running average ﬁltering with window size

7, in order to make the signals easier to compare on

the plots representing values in time. Running aver-

age results in low-pass ﬁltering, which in this case is

desirable, as high frequencies correspond mostly to

noise, and lower frequencies correspond to actual ac-

tions performed in the experiments.

In the case of orientation, Euler angles are pro-

vided in the range (-180°, 180°). Once the device is

rotated outside of this range (less than -180° or more

than 180°) the fusion algorithm returns an equivalent

value inside the range, e.g. -175° instead of 185° (see

the raw signal in Fig. 4). This makes it difﬁcult to

Heterogeneous Inertial Measurement Units in Motion Analysis

353

analyze such data, as large numeric differences be-

tween values in two signals may actually represent a

very small difference in measurement, e.g. 179° vs.

-179° is in fact just 2° of difference, not 358°. For

this reason, we apply dedicated processing in the ori-

entation experiments. In each signal, we search for

data points in which the transition between the end of

the positive and the start of the negative range occurs.

These are easy to ﬁnd as high peaks in the difference

of consecutive values. In segments located between

such data points, we add the full range (360°) to the

recorded values. Therefore, we obtain continuous sig-

nals (see. processed data in Fig. 4). We additionally

apply running average ﬁltering, with windows size 7,

to remove unwanted peaks that may still occur near

the transition points.

Figure 4: Raw and processed orientation data (single axis).

3.5 Metrics

As the ﬁrst indication of similarity between signals

from different devices, we analyze plots represent-

ing measured values in time. We compare against

the ﬁrst Xsens sensor, denoted as DOT (1), all other

sensors: the second Xsens device, denoted as DOT

(2), the MetaMotionRL sensor, and the Galaxy Watch

4. Then, we measure the numeric difference of mea-

surements in corresponding time points. For the ori-

entation data, we consider point-by-point differences

(60Hz sampling). Acceleration data, however, even

after employing running average ﬁltering, remains too

noisy to be directly compared. Instead, for numerical

analysis, we compute the mean value of the raw signal

in time windows of size 16, and overlap of half win-

dow length. This corresponds to the common tech-

nique of extracting time-domain features from the ac-

celeration data (O’Reilly et al., 2018). We then mea-

sure the difference between values computed in the

corresponding windows.

The mean and the standard deviation of difference

are computed and presented as numerical results. We

also include Bland-Altman plots, in which the dif-

ference between two paired measurements is plotted

against their mean. Bland-Altman plots are com-

monly used to depict agreement between different

measurement methods. For normally distributed data

agreement limits (dashed horizontal lines) are typi-

cally set to 1.96 of standard deviation, to include 95%

of data points. In our experiments, the differences are

not normally distributed (based on the Shapiro-Wilk

test), therefore we compute agreement limits speciﬁ-

cally to contain 95% of data points (per each signal-

to-signal comparison).

4 EXPERIMENTS

Performed experiments included the two scenarios

described in Section 3.2, focusing on measurements

of 1) acceleration in linear motion and 2) orientation.

Agreement between multiple repetitions of motion in

the same scenario was not considered in this work.

The duration of acquired signals was approx. 15 sec-

onds.

4.1 Linear Acceleration

For brevity, we present ﬁltered signal comparison in

the x-axis only, including DOT (1) vs. DOT (2) (Fig.

5), DOT (1) vs. MetaMotionRL (Fig. 6) and DOT (1)

vs. Galaxy Watch 4 (Fig. 7). Raw signals are addi-

tionally depicted with half-opacity dotted lines to in-

dicate the impact of the running average ﬁltering. The

Bland-Altman plot (Fig. 8) includes data points from

all axes. Each comparison (sensor pair) is indicated

with a different color.

Figure 5: Linear acceleration, x-axis, Xsens DOT (1) vs.

Xsens DOT (2).

Visual inspection of linear acceleration values in

time indicates that signals are generally well-aligned

in terms of peaks occurring at the same time points.

However, the magnitude of peaks varies, particularly

in data fragments with fast motion. This can be ob-

served in all comparisons - between the same device

model (Fig. 5), as well as with different models (Fig.

6 and 7). Numerical results (Tab. 1), as well as

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

354

Figure 6: Linear acceleration, x-axis, Xsens DOT (1) vs.

MetaMotionRL.

Figure 7: Linear acceleration, x-axis, Xsens DOT (1) vs.

Galaxy Watch 4.

Figure 8: Bland-Altman plot representing agreement in

measurements of linear acceleration (in time windows), all

axes, Xsens DOT (1) vs. all other sensors.

the Bland-Altman plot (Fig. 8), indicate that there

is much better agreement between measurements per-

formed with devices of the same model. Considering

the two other devices, the mean difference from val-

ues measured with the baseline sensor is better in the

case of Galaxy Watch 4 (0.65m/s

) than in the case

of MetaMotionRL (1.27m/s

). On the other hand,

the standard deviation is very similar (4.21m/s

and

4.09m/s

respectively) and so are the agreement lim-

its including 95% of data points (see Fig.8).

4.2 Orientation

Similarly, as in the case of linear acceleration, we

present signal comparison only in the x-axis to keep

Table 1: Mean and standard deviation values of differ-

ence in measuring linear acceleration (m/s

) between Xsens

DOT (1) and all other sensors, computed in time windows.

Sensor Mean Std

Xsens DOT (2) -0.05 1.28

MetaMotionRL 1.27 4.21

Galaxy Watch 4 0.65 4.09

this section concise. We include orientation data for

DOT (1) vs. DOT (2) (Fig. 9), DOT (1) vs. MetaMo-

tionRL (Fig. 10), and DOT (1) vs. Galaxy Watch

4 (Fig. 11). Raw signals are depicted with half-

opacity dotted lines. Numerical results (Tab. 2) and

the Bland-Altman plot (Fig. 12) include data from the

x-axis only. An additional close-up view of a smaller

range is depicted in Fig. 13.

Figure 9: Orientation, x-axis, Xsens DOT (1) vs. Xsens

DOT (2).

Figure 10: Orientation, x-axis, Xsens DOT (1) vs. Meta-

MotionRL.

In the plots depicting orientation value in time we

can observe, that alignment with the baseline DOT

(1) sensor data is near perfect for both DOT (2) and

Galaxy Watch 4 devices (Fig. 9 and 11). However,

it is worth noting, that the Galaxy Watch 4 requires

running average ﬁltering, as the raw data have some

undesirable peaks, while DOT (2) provides a smooth

signal by itself. MetaMotionRL sensor produces sim-

ilar peaks as the other devices, although some of them

are shifted in time (Fig. 10). This is most likely due

to different sensor fusion algorithms - there is always

Heterogeneous Inertial Measurement Units in Motion Analysis

355

Figure 11: Orientation, x-axis, Xsens DOT (1) vs. Galaxy

Watch 4.

Figure 12: Bland-Altman plot of orientation, x-axis, Xsens

DOT (1) vs. other sensors, full range.

Figure 13: Bland-Altman plot of orientation, x-axis, Xsens

DOT (1) vs. other sensors, selected range.

a trade-off between accuracy and delay in comput-

ing orientation from the noisy, raw, multi-modal sig-

nals. Please note, that in this work we do not eval-

uate which device has better consistency with a gold

standard measurement, but rather measure agreement

between the sensors, therefore we do not conclude

which data is more correct.

Results in Tab. 2, as well as the close-up of the

Bland-Altman plot (Fig. 13) also indicate very good

agreement between DOT (1) and both DOT (2) and

Galaxy Watch 4 sensors. The latter has a lower mean

difference (2.75° vs. 3.18°), but a higher standard de-

viation (5.11° vs. 3.64°). MetaMotionRL has a very

high standard deviation of difference (62.10°), as a

result of the observed time shift. However, the mean

difference (4.52°) is comparable with the other sen-

Table 2: Mean and standard deviation values of difference

in measuring orientation (degrees) between Xsens DOT (1)

and all other sensors.

Sensor Mean Std

Xsens DOT (2) 3.18 3.64

MetaMotionRL 4.52 62.10

Galaxy Watch 4 2.75 5.11

sors, due to the circular nature of the motion. The

time shift is also well visible on the full-range Bland-

Altman plot (Fig. 12).

4.3 Discussion

Based on the obtained results, we can conclude that in

the case of linear acceleration, we cannot easily sub-

stitute one model of IMU with another, as measure-

ments will differ. Depending on the application, these

differences may be less or more signiﬁcant. In partic-

ular, when relatively slow motion is considered, mea-

surements seem to be more consistent. In applications

including very dynamic motion, such as some sports

disciplines, a per-sensor adaptation of data process-

ing methods may be required, to compensate for the

difference in measurements. On the other hand, for

scenarios with less dynamic motion, such as daily ac-

tivity recognition, those differences may be insignif-

icant. While recording actual actions was out of the

scope of this work, it would be interesting to compare

differences between sensors to differences between

subjects in the context of the type of performed mo-

tion. It is also worth noting that both a dedicated IMU

(MetaMotionRL) and a smartwatch (Galaxy Watch 4)

have a very similar agreement with the baseline DOT

sensor.

Interestingly, considering orientation, measure-

ments from the smartwatch are much more consis-

tent with the baseline, than the MetaMotionRL IMU.

We expect that, in terms of orientation, Galaxy Watch

4 could be directly used with motion analysis algo-

rithms developed using the DOT sensors. On the

other hand, the time shift occurring between the DOT

and MetaMotionRL data may be difﬁcult to compen-

sate for. Using a custom fusion algorithm instead

of the provided one could be considered. Some ap-

plications may not require very dynamic measure-

ments of orientation, e.g. rehabilitation exercises are

usually performed relatively slowly, therefore precise

synchronization in time may be less relevant than the

range of motion.

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

356

5 CONCLUSIONS

In this work, we performed experiments to analyze

the agreement between measurements obtained with

the same and different models of IMUs. We in-

cluded dedicated inertial sensors as well as a common

model of a smartwatch. Results indicate, that while

the agreement is relatively good, it is not sufﬁcient

to simply substitute a device with another model or

use a heterogeneous setup without additional consid-

eration. However, in some applications, with proper

adaptation, using multiple different sensors could be

a viable solution. More importantly, from a practi-

cal point of view, employing an everyday-use device

such as a smartwatch is just as good (or even better)

as using another model of IMU. Discrepancies be-

tween measurements obtained with different devices

are more signiﬁcant in dynamic motion. Therefore, in

scenarios such as analyzing fast, sports actions it may

be necessary to calibrate motion analysis methods per

device.

In terms of future work, it would be beneﬁcial to

verify the repeatability of measurements per device,

as well as perform experiments with additional de-

vices. We also consider comparing signals from dif-

ferent sensors as inputs to machine learning meth-

ods for action detection or classiﬁcation. Such ex-

periments would include multiple subjects. Training

machine learning models on one set of sensors and

testing on another would be a good indication of the

viability of heterogeneous setups in practical scenar-

ios.

ACKNOWLEDGEMENTS

The research presented in this paper was supported

by the National Centre for Research and Develop-

ment (NCBiR) under Grant No. LIDER/37/0198/L-

12/20/NCBR/2021.

REFERENCES

Bavan, L., Surmacz, K., Beard, D., Mellon, S., and Rees, J.

(2019). Adherence monitoring of rehabilitation exer-

cise with inertial sensors: A clinical validation study.

Gait & posture, 70:211–217.

Bevilacqua, A., Huang, B., Argent, R., Caulﬁeld, B., and

Kechadi, T. (2018). Automatic classiﬁcation of knee

rehabilitation exercises using a single inertial sensor:

A case study. In 2018 IEEE 15th International Con-

ference on Wearable and Implantable Body Sensor

Networks (BSN), pages 21–24. IEEE.

Camomilla, V., Bergamini, E., Fantozzi, S., and Vannozzi,

G. (2018). Trends supporting the in-ﬁeld use of wear-

able inertial sensors for sport performance evaluation:

A systematic review. Sensors, 18(3):873.

Clemente, F., Badicu, G., Hassan, U., Akyildiz, Z., Pino-

Ortega, J., Silva, R., and Rico-Gonz

alez, M. (2022).

Validity and reliability of inertial measurement units

for jump height estimations: a systematic review. Hu-

man Movement, 23(4):1–20.

Giggins, O. M., Sweeney, K. T., and Caulﬁeld, B. (2014).

Rehabilitation exercise assessment using inertial sen-

sors: a cross-sectional analytical study. Journal of

neuroengineering and rehabilitation, 11(1):1–10.

Khediri, S. e., Nasri, N., Samet, M., Wei, A., and Kachouri,

A. (2012). Analysis study of time synchronization

protocols in wireless sensor networks. arXiv preprint

arXiv:1206.1419.

Kobsar, D., Masood, Z., Khan, H., Khalil, N., Kiwan, M. Y.,

Ridd, S., and Tobis, M. (2020). Wearable inertial sen-

sors for gait analysis in adults with osteoarthritis—a

scoping review. Sensors, 20(24):7143.

Lopez-Nava, I. H. and Munoz-Melendez, A. (2016). Wear-

able inertial sensors for human motion analysis: A re-

view. IEEE Sensors Journal, 16(22):7821–7834.

Malawski, F. (2020). Depth versus inertial sensors in real-

time sports analysis: a case study on fencing. IEEE

Sensors Journal, 21(4):5133–5142.

Milosevic, B., Leardini, A., and Farella, E. (2020). Kinect

and wearable inertial sensors for motor rehabilitation

programs at home: state of the art and an experimental

comparison. Biomedical engineering online, 19(1):1–

26.

O’Reilly, M., Caulﬁeld, B., Ward, T., Johnston, W., and Do-

herty, C. (2018). Wearable inertial sensor systems for

lower limb exercise detection and evaluation: a sys-

tematic review. Sports Medicine, 48(5):1221–1246.

Petraglia, F., Scarcella, L., Pedrazzi, G., Brancato, L.,

Puers, R., and Costantino, C. (2019). Inertial sensors

versus standard systems in gait analysis: a systematic

review and meta-analysis. European journal of physi-

cal and rehabilitation medicine, 55(2):265–280.

Ren, H. and Wang, X. (2021). Application of wearable

inertial sensor in optimization of basketball player’s

human motion tracking method. Journal of Ambient

Intelligence and Humanized Computing, pages 1–15.

Roell, M., Mahler, H., Lienhard, J., Gehring, D., Gollhofer,

A., and Roecker, K. (2019). Validation of wearable

sensors during team sport-speciﬁc movements in in-

door environments. Sensors, 19(16):3458.

Teuﬂ, W., Lorenz, M., Miezal, M., Taetz, B., Fr

ohlich,

M., and Bleser, G. (2018). Towards inertial sen-

sor based mobile gait analysis: Event-detection and

spatio-temporal parameters. Sensors, 19(1):38.

Tunca, C., Pehlivan, N., Ak, N., Arnrich, B., Salur, G.,

and Ersoy, C. (2017). Inertial sensor-based robust gait

analysis in non-hospital settings for neurological dis-

orders. Sensors, 17(4):825.

Wang, Y., Zhao, Y., Chan, R. H., and Li, W. J. (2018). Vol-

leyball skill assessment using a single wearable mi-

Heterogeneous Inertial Measurement Units in Motion Analysis

357

cro inertial measurement unit at wrist. IEEE Access,

6:13758–13765.

Wang, Z., Shi, X., Wang, J., Gao, F., Li, J., Guo, M., Zhao,

H., and Qiu, S. (2019). Swimming motion analy-

sis and posture recognition based on wearable inertial

sensors. In 2019 IEEE International Conference on

Systems, Man and Cybernetics (SMC), pages 3371–

3376. IEEE.

Wilmes, E., de Ruiter, C. J., Bastiaansen, B. J., Zon, J. F. v.,

Vegter, R. J., Brink, M. S., Goedhart, E. A., Lemmink,

K. A., and Savelsbergh, G. J. (2020). Inertial sensor-

based motion tracking in football with movement in-

tensity quantiﬁcation. Sensors, 20(9):2527.

Worsey, M. T., Espinosa, H. G., Shepherd, J. B., and Thiel,

D. V. (2019a). Inertial sensors for performance anal-

ysis in combat sports: A systematic review. Sports,

7(1):28.

Worsey, M. T., Espinosa, H. G., Shepherd, J. B., and Thiel,

D. V. (2019b). A systematic review of performance

analysis in rowing using inertial sensors. Electronics,

8(11):1304.

Zhou, H., Hu, H., and Harris, N. (2006). Application of

wearable inertial sensors in stroke rehabilitation. In

2005 IEEE Engineering in Medicine and Biology 27th

Annual Conference, pages 6825–6828. IEEE.

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

358