Location Determination of On-body Inertial Sensors

Hisham Madcor

1 a

, Osama Adel

1 b

and Walid Gomaa

1,2 c

Cyber-Physical Systems Lab, Department of Computer Science and Engineering,

Egypt-Japan University of Science and Technology, Borg El-Arab, Alexandria, Egypt

Faculty of Engineering Alexandria University, Alexandria, Egypt

Keywords:

Sensor Position, Multi-sensory, Intertial Sensors, Human-centered Computing.

Abstract:

Human Activity Recognition has gained tremendous drive in recent years. This is due to the increasing ubiq-

uity of all types of sensors in commodity devices such as smartphones, smart watches, tablets, etc. This has

made available to the normal user a continuous stream of data including visual data, inertial motion data, au-

dio, etc. In this paper we focus on data streamed from inertial motion units (IMUs). Such units are currently

embedded on almost all wearable devices including smart watches, wrist bands, etc. In many research works,

as well as in many real applications, different specialized IMU units are mounted on different body parts. In

the current work, we try to answer the following question: given the streamed inertial signals of a gait pattern,

as well as some other activities, determine which sensor location on the subject’s body generated this signal.

We validate our work on several datasets that contain multi-dimensional measurements from a multitude of

sensors mounted on different body parts. The main sensors used are the accelerometer and gyroscope. We use

the Random Forest Classiﬁer over the raw data without any prior feature extraction. This has proven yet very

effective as evidenced by the results using different metrics including accuracy, precision, recall, F1-score,

etc. An important application of such research can be in data augmentation of timeseries inertial data. This

can be used as well for healthcare applications, for example, in treatment assessment for people with motion

disabilities.

1 INTRODUCTION

Human Activity Recognition (HAR) is a ﬁeld of re-

search that boomed in popularity in recent years with

the ubiquity of wearable smart devices such as smart

watches and smartphones. Using the inertial sen-

sors (such as accelerometers, gyroscopes and mag-

netometers) embedded inside these devices, the mo-

tion dynamics of different activities can be recorded,

streamed to a processing unit, and be analyzed to rec-

ognize the action the current user is doing. Similarly,

recent research showed also that the identity of the

user can be identiﬁed using these sensors during one

of the most common activities done throughout the

day, i.e. walking (Adel et al., 2020).

Nevertheless, since these devices can be used in

so many different contexts (e.g. sports, medical pur-

poses, research purposes), they can be placed on dif-

ferent locations on human body (wrists, legs, arms,

https://orcid.org/0000-0002-8727-7242

https://orcid.org/0000-0003-3471-5787

https://orcid.org/0000-0002-8518-8908

etc.). The variety of places that these sensors can be

placed on can greatly affect the accuracy of human

activity recognition or person identiﬁcation. Mean-

while, detecting sensor placement has not attracted

enough attention in research and has been done im-

plicitly in few works on few common places (such as

right and left pockets and right and left hands only)

(Primo et al., 2014).

In this paper we focus on the detection of loca-

tion of on-body inertial sensors using Random Forest

Classiﬁer during various activities using four differ-

ent publicly available datasets that use inertial sensors

( accelerometer and gyroscope included). The contri-

butions of this work are as follows:

• We present a full detection of inertial wearable

sensors on different parts of human body during

various activities on four different datasets.

• We work on raw inertial data without any feature

extraction algorithms.

• We achieved an average accuracy of 92.33% on

all the four datasets.

Madcor, H., Adel, O. and Gomaa, W.

Location Determination of On-body Inertial Sensors.

DOI: 10.5220/0010582806930700

In Proceedings of the 18th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2021), pages 693-700

ISBN: 978-989-758-522-7

693

Model Evaluation Data Processing Model Training Results

Data

Unfolding

Segmentation

Labeling

Load data as

Numpy

Random

Forest

Classifier

Dataset

Confusion

Matrix

Classification

Report

Accuracy

Figure 1: Experiments pipeline.

The rest of the paper is organized as follows. Sec-

tion 1 as introduction. The related work is presented

in Section 2. Then, we explain our proposed method-

ology in Section 3, which we explain in it the experi-

ments pipeline shown in Figure 1. After that, we show

our experiments in Section 4. Our results and discus-

sion are presented in Section 5. Finally, Section 6 is

the conclusion of our work and the future plans.

2 RELATED WORK

(Kunze et al., 2005) managed to derive the location

of the acceleration sensor on the user’s body using

the sensor’s signals only. Their algorithm in detecting

the location of the sensor is not affected by the sen-

sor’s orientation. Using this algorithm, they managed

to identify the time periods where the user is walking

and then by using the unique characteristics of walk-

ing they could identify the location of the sensor on

the user’s body. The fact that the location of the sensor

is an interesting context was what motivated them to

do this work. In addition to that, They denoted that the

locations of the sensors was chosen according to reg-

ularly used devices and sensors. They used four clas-

siﬁers: C 4.5, Naive Bayes, Naive Bayes simple and

Nearest Neighbor. The experiments were on 6 sub-

jects. They conducted 3 runs each run was between

12 and 15 minutes with 8 activities performed. The

senors were placed on 6 different body parts: wrist,

right side of the head, left trouser’s pocket and left

breast pocket. The C 4.5 classiﬁer got the highest ac-

curacy among the four used classiﬁers by accuracy

89.81%.

On the other hand, (Vahdatpour et al., 2011) used

accelerometers to capture the motion data of subjects’

actions. This data allowed them to detect the loca-

tion of the sensor on the subject’s body by using a

mixed supervised and unsupervised time series analy-

sis model. They used Support Vector Machine (SVM)

algorithm to identify the sensors’ locations. Further-

more, they used their own dataset which consisted

of 25 subjects with sensors mounted on 10 different

places on the body. Those 10 places were classi-

ﬁed into 6 regions: forearm, upper arm, head, thigh,

shin and waist. In their ﬁrst conducted experiment,

they trained and tested the SVM on each subject sep-

arately with training and testing ratio 2:8. The results

were 88%, 98% and 100% for the minimum, mean

and maximum precision, respectively. In the second

conducted experiment, they trained the model on ran-

domly chosen segment from different subjects and the

average accuracy for this experiment was 89%.

In (Sztyler and Stuckenschmidt, 2016), the au-

thors presented a dataset with 17 subjects with 8 per-

formed activities, each subject wore 7 sensors on the

head, chest, upper arm, waist, forearm, thigh and shin.

They used acceleration sensor’s data like the previ-

ous two works to identify the sensor’s location by us-

ing a Random Forest Classiﬁer. They introduced their

method for the location identiﬁcation by treating the

position as a multi class classiﬁcation problem that

made the sensors’ locations the targeted classes. They

conducted their experiments on each subject individu-

ally, as they reasoned, due to the difference in individ-

ual behaviour and ages. After they had conducted the

experiments, they achieved an average performance

accuracy of 89% across all positions and subjects.

The authors in (Weenk et al., 2013) introduced an

automatic identiﬁcation method for inertial sensors

on different body parts during walking. This intro-

duced method allows the user to place inertial sen-

sors in a full body or lower body plus trunk conﬁgu-

rations. The number of sensors that implemented by

the user ranges from 17 to 8 inertial sensor in the men-

tioned places respectively. Based on the acceleration

and angular velocity data extracted from the user’s

walking for a few seconds, the identiﬁcation process

is automatically done. Their dataset is composed of

11 healthy subjects performed 35 walking trials, and

then tested on 7 patients after a reconstruction surgery

in their knee. The authors extracted RMS, variance,

correlation and inter-axis correlation co-efﬁcients fea-

tures from magnitudes and the 3D components of the

acceleration, angular acceleration and angular veloc-

ity. In their experiments, the authors used J4.8 deci-

sion tree algorithm as their classiﬁer. J4.8 is an im-

plementation for C 4.5 algorithm which is the same

classiﬁer used in (Kunze et al., 2005). Their process

gets 100% for lower body sensors plus trunk conﬁgu-

rations and 97% for full body sensors.

In this work, however, we conduct our experi-

ments on four publicly available datasets with vary-

ing number of subjects and activities using a Random

ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics

694

Forest Classiﬁer trained on the raw inertial data of

accelerometer and gyroscope only. We also achieve

state-of-the-art performance in comparison to the pre-

vious works.

3 PROPOSED METHODOLOGY

In this section, we will explain brieﬂy the different

stages of our proposed methodology. As can be seen

in Figure 1, we start by loading the datasets before

performing any processing on them in the preprocess-

ing stage. Next, the data is fed to a machine learn-

ing classiﬁer model for training. After the model is

trained, we evaluate our model using different metrics

on a test set and report our results.

3.1 Data Preprocessing

The data preprocessing stage consists of three steps:

loading the data, data segmentation and labeling, then

data unfolding.

In the ﬁrst stage, we make sure that all dataset ﬁles

are converted to Numpy arrays. This is because each

dataset we use has a different ﬁle structure and a dif-

ferent ﬁle format, it is an important step to unify all

these datasets in one common format.

Each IMU data ﬁle consists of R×C matrix, where

R is the number of timesteps, and C is the number

of axes or channels (i.e. 6 channels for a 3-axis ac-

celerometer and 3-axis gyroscope). We iterate over

each data ﬁle for each IMU device and segment the

data into 3-second ﬁxed-length segments. For exam-

ple, if the sampling rate is in our dataset is equal to

50 Hz, our ﬁnal segments will be 150 × 6 matrices.

For each data segment, we construct a label y corre-

sponding to the IMU device from which the data was

recorded. Hence, if we have M segments, we end up

with M labels as well. To be able to use the data seg-

ments in our machine learning model, we unfold each

data segment into a 1-D vector x ∈ R

. Therefore,

the output of the preprocessing stage is a M × RC data

matrix X, and a M-dimensional labels vector y.

3.2 Model Training

After the data preprocessing stage, the data is ready to

be fed to the model. We train a Random Forest Classi-

ﬁer to predict the location of the IMUs in a number of

datasets. Random Forest Classiﬁer is a ensemble clas-

siﬁer that uses a large number of decision trees classi-

ﬁers on features selected randomly from the training

data, and then use averaging to improve the prediction

accuracy and control over ﬁtting (Breiman, 2001). In

our model, We set the number of decision trees to 500,

use the gini index which is the function that measures

the quality of a split. We set the minimum number of

samples to split to 2 samples and used bootstrapping.

3.3 Model Evaluation

After training the model, we then evaluate it on a test

set using four different metrics: the accuracy, preci-

sion, recall and F1-score.

We calculate the test accuracy of the model as fol-

lows:

accuracy = [

N−1

∑

i=0

1( ˆy

= y

)] × 100% (1)

where y

is the test label number i, ˆy

is the pre-

dicted label number i and N is the total number of

samples in the test set. The expression 1(x

= y

) is

called indicator function which indicates the member-

ship of ˆy in y.

We calculate the precision and recall using the fol-

lowing formulae:

precision =

T P

T P + FP

(2)

recall =

T P

T P + FN

(3)

where T P is the number of true positives, FP is

the number of false positives and FN is the number

of false negatives for each class. The precision ratio

shows the classiﬁers ability to avoid labeling a pos-

itive sample as negative while the recall ratio shows

the ability of the classiﬁer to correctly ﬁnd all the pos-

itive samples.

On the other hand, The F1-Score is the harmonic

mean of precision and recall. The F1-Score is calcu-

lated as:

= 2 ×

precision × recall

precision + recall

(4)

The F1-score is a single number that incorporates

both the prevision and recall. Its maximum value is 1

when the model has perfect precision and recall. The

minimum value is 0 when either the precision or recall

is zero.

4 EXPERIMENTS

In our experiments we apply the steps mentioned in

Section 3 on four publicly available datasets: EJUST-

GINR-1, RealDisp, HuGaDB and MMUSID. Each

Location Determination of On-body Inertial Sensors

695

Table 1: Datasets summary.

Datasets No. of samples No. of sensors No. of activities

Body parts Other

Back Waist Legs Arms BP HCB

EJUST-GINR-1 46236 8 1 1 1 2 4 - -

RealDisp 148698 9 33 1 - 4 4 - -

HuGaDB 70980 6 12 - - 6 - - -

MMUSID 26924 6 1 - - 2 2 1 1

(b)

RealDisp

(a)

EJUST-GINER-1

Back

LUA

RUA

LCRC

Waist

Back

RUA

RLA

LUA

LLA

RC LC

LTRT

BackPack

HCB

(c)

HuGaDB

(d)

MMUSID

Figure 2: Datasets Sensors’ Locations.

dataset has a slightly different number of devices

mounted on different locations on human body as

shown in Figure 2. We load each dataset then feed

75% of the raw data (without any feature extraction)

to our Random Forest Classiﬁer to train on and ﬁnally,

we evaluate our model on 25% of the data reporting

the four measure dicussed in Section 3.3.

4.1 Datasets

In this section we give a brief overview of the datasets

used in our experimentation. The highlights of the

four datasets are summarized in Table 1.

4.1.1 EJUST-GINR-1

EJUST-GINR-1 dataset consists of 8 IMUs each unit

streams tri-axis accelerometer and tri-axis gyroscope

data with a sampling rate of 50 Hz. Gait data is

collected from 20 subjects, 10 males and 10 fe-

males (Adel et al., 2020) and the sensory units are

located on 8 different body parts as show in Figure 2a.

4.1.2 RealDisp

The RealDisp dataset consists of 9 IMUs each streams

tri-axis accelerometer, tri-axis gyroscope and tri-axis

magnetic ﬁeld readings, in addition to orientation es-

timates in 4D format with sampling rate of 50 Hz.

The data is collected from 17 subjects with 33 activ-

ities. There are three scenarios for placing the sen-

sors: Ideal, Self and Mutual. In the ideal scenario, the

instructor places the sensors in the predeﬁned places

on the subject’s body. In the self scenario, the user

is asked by the instructor to position 3 sensors on

the speciﬁed locations. In the mutual scenario, the

instructor introduces an intentional disposition of a

number of sensors (4 to 7 sensors) using rotations and

translations with respect to the ideal case (Banos and

oth, 2014). In our experiment we only use the ideal

scenario ﬁles. The sensors are located on 9 different

body parts as shown in Figure 2b.

4.1.3 HuGaDB

HuGaDB dataset consists of 8 units, 6 units are iner-

tial sensor units streaming tri-axial accelerometer and

tri-axial gyroscope readings while the other two are

EMG sensors. The data is collected from 18 subjects

with 12 different activities (Chereshnev and Kert

esz-

Farkas, 2017). In our experiment we used only the

inertial sensors readings which are located mainly on

6 different parts of the leg as shown in Figure 2c

ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics

696

4.1.4 MMUSID

MMUSID consists of 6 sensor units each streaming

tri-axial accelerometer and tri-axial gyroscope read-

ings for 120 subjects with sampling rate of 50 Hz. The

dataset includes only the walking activity but with

different speeds: slow, normal and fast walking (Per-

matasari et al., 2020). The sensors in this dataset are

placed in places that mimic the real life such as in a

backpack, hand-carried bag, right and left pockets and

right and left hands as shown in Figure 2d.

5 RESULTS AND DISCUSSION

5.1 Results

In this section we report the results of our experiments

on each dataset separately. We constructed a confu-

sion matrix for our model on each dataset showing

the cross-accuracy between the classes while the the

vertical axis is the ground truth and the the horizon-

tal axis is the prediction. The confusion matrices for

all datasets are shown in Figure 3a, Figure 3b, Fig-

ure 3c and Figure 3d. It is clear from the diagonal

of the four confusion matrices that the model could

observe all the sensor places with high accuracy in

all the datasets. Moreover, as we see in Table 2,

the model has a test accuracy over all classes for

EJUST-GINR-1, RealDisp, HuGaDB and MMUSID

of 92.86%, 84%, 93.98% and 98.49%, respectively.

5.1.1 EJUST-GINR-1

The EJUST-GINR-1 dataset has 8 sensors that are

placed in 8 different body parts. From the confusion

matrix shown in Figure 3a we could observe that the

Left Calf (LC) and Right Calf (RC) have the highest

detection accuracy among the 8 sensors and the con-

fusion between their place detection is nearly negligi-

ble. The Right Hand (RH) and Left Hand (LH) sen-

sors have some confusion in their place prediction,

but does not have any conﬂict with any other sen-

sor. The Right Upper Arm (RUA) and the Left Upper

Arm (LUA) have confusion between them only by ac-

curacy of 12.195% and 13.800%, respectively. The

Waist and RUA have a noticeable conﬂict between

each other by accuracy of 3.579% and 2.826%, re-

spectively.

5.1.2 RealDisp

The RealDisp dataset has 9 sensors in 9 different

places. From the confusion matrix in Figure 3c,

we observe some confusion between sensors loca-

tions. The RUA and Right Leg (RL) both have notice-

able miss-classiﬁed cross predictions between each

other by about 6.674% and 6.993% accuracy, respec-

tively. Similarly, we have the Left arm sides LLA and

LUA miss-classiﬁed cross-predictions by 6.510% and

6.211%, respectively. The LUA and RUA have miss-

classiﬁed cross predictions by 4.197% and 5.628%,

respectively. Nevertheless, the other miss predictions

are surprisingly almost negligible. On the other hand,

for the lower part of the body, we have the Right

Thigh (RT) and the Left Thigh (LT) having predic-

tion confusion by about 6.950% and 7.868%, respec-

tively, while we have the RC and LC confusion by

about 8.038% for RC and 7.249% for LC.

5.1.3 HuGaDB

As we saw in Section 4.1.3 this dataset is mush differ-

ent in the placement of their sensor units, as all of the

6 sensors are located on different places along the two

legs, so we are inspecting how the model could differ-

entiate between places that are very proximal. As we

see in the confusion matrix Figure 3c, we can observe

that RT and LT have some wrong cross predictions

by about 4.997% for the RT and 4.759% for the LT.

Also, for the Right Shin (RS) and Left Shin (LS) they

cross-predict each other by 9.198%, for the RS and

7.414% for the LS. For the Left Foot (LF) and Right

Foot (RF) there is nearly negligible conﬂict between

them.

5.1.4 MMUSID

The MMUSID dataset is totally different and include

locations for the sensors the other three datasets did

not cover, also it more mimics sensor placement in

real-life(Permatasari et al., 2020) and as we see in

Figure 2d. The sensors are carried mobile phones, in

the left and right pockets (LP and RP), in the Back-

Pack (BP), being hold in the two hands (LH and RH),

and the last one has been in a hand carry bag (HCB).

Those 6 locations are common places to put your

phone in. Therefore, from the confusion matrix in

Figure 3d we can clearly see that because of the dis-

tance found between locations of the sensors there is

nearly no conﬂict found. However, we can see that

the HCB has a little conﬂict with the RP by 1.027%.

Also, the RH has little cross conﬂict with the LH by

1.002%.

5.2 Discussion

From the above results and observations, we can

conclude that in the three datasets EJUST-GINR-1,

Location Determination of On-body Inertial Sensors

697

(a) EJUST-GINR-1 (b)RealDisp

Figure 3: Confusion Matrices.

RealDisp and HuGaDB, there is high degree of con-

fusion between the right and left body parts i.e. LUA

and RUA, LLA and RLA , RS and LS, etc. Similarly,

there is a high degree of confusion between the sen-

sors on the same limb, for instance, the sensors on the

RUA and RLA and LUA and LLA.

Hence, it is apparent one can use the readings from

a sensor mounted on one side of the human body to in-

fer the readings from the opposite side. Also, it shows

that the difference between readings from proximate

body parts are subtle and should not make an issue for

a model trained on data from one of them. However,

the data readings, in the MMUSID, from the right and

left feet are unique, and separate models are required

to for each side.

On the other hand, in the MMUSID dataset, we

found that there is a false prediction in the place of

the HCB sensor with the RP sensor, which indicates

that nearly all the subjects carry their bags in the right

hand which makes the data biased towards the right

handed people only and this may give false predic-

tions in the experiments done on this dataset.

ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics

698

Table 2: Classiﬁcation report for Datasets.

Datasets Units Precision Recall F1-Score Accuracy

EJUST-GINER-1

LUA 0.90508 0.91734 0.91117

92.86%

Back 0.96232 0.97432 0.96828

RUA 0.88284 0.88958 0.88620

LC 0.99713 0.97407 0.98547

Waist 0.93180 0.93369 0.93275

RC 0.99869 0.98700 0.99281

LH 0.86200 0.87969 0.87075

RH 0.87805 0.86015 0.86901

RealDisp

RLA 0.86617 0.86994 0.86805

84%

RUA 0.78791 0.83141 0.80907

Back 0.88443 0.88506 0.88475

LUA 0.80887 0.82048 0.81464

LLA 0.88918 0.87701 0.88305

RC 0.82957 0.82937 0.82947

RT 0.84306 0.81188 0.82718

LT 0.82370 0.81449 0.81907

LC 0.84310 0.83337 0.83821

HuGaDB

RF 0.99526 0.98984 0.99304

93.98%

RS 0.88903 0.91019 0.89949

RT 0.93372 0.92804 0.93087

LF 0.98803 0.98938 0.98871

LS 0.90639 0.89236 0.89932

LT 0.92778 0.93029 0.92903

MMUSID

LP 0.99018 0.97882 0.98447

98.49%

RP 0.98244 0.97882 0.98447

HCB 0.98768 0.99175 0.98971

BP 0.98341 0.99213 0.98775

LH 0.98458 0.98139 0.98298

RH 0.98228 0.98837 0.98532

6 CONCLUSIONS AND FUTURE

WORK

In this work, we investigated the problem of detecting

the location of an Inertial Measurement Unit (IMU)

on the human body. The detection of the location

of these devices accurately can have a huge impact

on the accuracy of human activity recognition (HAR)

and person identiﬁcation. Therefore, we train a Ran-

dom Forest Classiﬁer on four different publicly avail-

able datasets separately and report the accuracy, preci-

sion, recall and F1-score on each one of them. More-

over, we show the confusion between the classiﬁ-

cation of each location and get some important in-

sights. Our model achieved 98.82% accuracy on the

MMUSID dataset which mimic real-life sensor place-

ments, 93.68% accuracy on HuGaDB which incor-

porates 13 activities, 92.85% accuracy on EJUST-

GINR-1 dataset and 83.86% accuracy on ReadlDisp

dataset which incorporate 33 activities. It is important

to notice that we train our model on the raw inertial

data without doing any feature engineering.

In the future, we plan to extend this work to in-

clude sensor location determination from many other

actions. From these we can see how different ac-

tions are dependent on the motion of different body

parts, which can be used, for example, in treatment

assessment for people with motion disabilities. An-

other line of research would be to map the signals

streamed from a given sensor mounted on a speciﬁc

place on the body to signals as if they were streamed

from other body parts. This can be considered as a

data augmentation that would extend the size of the

dataset without actually recollecting data using phys-

ical sensors mounted on the target body parts.

ACKNOWLEDGEMENT

This work is funded by the Information Technol-

ogy Industry Development Agency (ITIDA), Infor-

mation Technology Academia Collaboration (ITAC)

Program, Egypt – Grant Number (ARP2020.R29.2

- VCOACH: Virtual Coaching for Indoors and Out-

doors Sporting).

Location Determination of On-body Inertial Sensors

699

REFERENCES

Adel, O., Nafea, Y., Hesham, A., and Gomaa, W. (2020).

Gait-based person identiﬁcation using multiple iner-

tial sensors.

Banos, O. and T

oth, M. A. (2014). Realistic sensor dis-

placement benchmark dataset. Dataset manual, pages

1–4.

Breiman, L. (2001). Random forests machine learning, vol.

45.

Chereshnev, R. and Kert

esz-Farkas, A. (2017). Hugadb:

Human gait database for activity recognition from

wearable inertial sensor networks. In International

Conference on Analysis of Images, Social Networks

and Texts, pages 131–141. Springer.

Kunze, K., Lukowicz, P., Junker, H., and Tr

oster, G.

(2005). Where am i: Recognizing on-body posi-

tions of wearable sensors. In International Symposium

on Location-and Context-Awareness, pages 264–275.

Springer.

Permatasari, J., Connie, T., and Song, O. T. (2020). The

mmuisd gait database and performance evaluation

compared to public inertial sensor gait databases. In

Computational Science and Technology, pages 189–

198. Springer.

Primo, A., Phoha, V. V., Kumar, R., and Serwadda, A.

(2014). Context-aware active authentication using

smartphone accelerometer measurements. In Pro-

ceedings of the IEEE conference on computer vision

and pattern recognition workshops, pages 98–105.

Sztyler, T. and Stuckenschmidt, H. (2016). On-body lo-

calization of wearable devices: An investigation of

position-aware activity recognition. In 2016 IEEE In-

ternational Conference on Pervasive Computing and

Communications (PerCom), pages 1–9. IEEE.

Vahdatpour, A., Amini, N., and Sarrafzadeh, M. (2011).

On-body device localization for health and medical

monitoring applications. In 2011 IEEE International

Conference on Pervasive Computing and Communi-

cations (PerCom), pages 37–44. IEEE.

Weenk, D., Van Beijnum, B.-J. F., Baten, C. T., Hermens,

H. J., and Veltink, P. H. (2013). Automatic identiﬁca-

tion of inertial sensor placement on human body seg-

ments during walking. Journal of neuroengineering

and rehabilitation, 10(1):1–9.

ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics

700