Gait-based Person Identiﬁcation using Multiple Inertial Sensors

Osama Adel

1 a

, Yousef Nafea

, Ahmed Hesham

and Walid Gomaa

1,2 b

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:

Gait, Person Identiﬁcation, Multi-sensory, Intertial Sensors, Human-centered Computing.

Abstract:

Inertial sensors such as accelerometers and gyroscopes have gained popularity in recent years for their use

in human activity recognition. However, little work has been done on using these sensors for gait-based

person identiﬁcation. Gait-based person identiﬁcation turns out to be important in applications such as where

different people share the same wearable device and it is desirable to identify who is using the device at a

given time while walking. In this research, we present the ﬁrst multi-sensory gait-based person identiﬁcation

dataset EJUST-GINR-1 and present our work on gait-based person identiﬁcation using multi-sensory data,

by mounting 8 wearable inertial sensory devices on different body locations and use this data to identify the

person using it. Two of these sensors are smart watches worn on both wrists. We explore the correlation

between each body location and the identiﬁcation accuracy, as well as exploring the effect of fusing pairs of

sensory units in different locations, on the ﬁnal classiﬁcation performance.

1 INTRODUCTION

Inertial sensors such as accelerometers and gyro-

scopes have gained a renowned reputation in re-

cent years for their use in human activity recognition

(HAR). These sensors can capture the motion dynam-

ics of different parts of the human body, and stream

this data to a processing unit for further analysis. By

analyzing the streamed data (either in real-time or of-

ﬂine), we can classify the motion done by a human as

one of different activities, which proved to be helpful

in a plethora of applications speciﬁcally, in the sports

and healthcare ﬁelds (Patel et al., 2012).

However, inertial sensors haven’t received enough

attention in Gait-based Person Identiﬁcation. Gait-

based Person Identiﬁcation using inertial sensors is

concerned with the problem of identifying a person

from a predeﬁned set of people using one or more in-

ertial measurement units mounted on the body while

walking. It’s different from person authentication or

veriﬁcation in the sense that the later addresses the

problem of identifying if a certain person is proper

one wearing the (embedded) inertial sensors or not.

Since gait patterns are unique for each individual,

given the inertial sensors readings of an individual’s

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

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

gait patterns, that individual should be correctly and

uniquely identiﬁed (Tao et al., 2018).

Machine learning algorithms (including deep

learning) have been widely used in human activity

recognition (Lara and Labrador, 2013) and should

show good results in person identiﬁcation as well.

Despite the wide use of deep learning techniques,

some traditional machine learning still achieves com-

parable performance in activity recognition espe-

cially, with much less computational demands, when

the amount of data available is small (Abdu-Aguye

and Gomaa, 2019).

Person identiﬁcation is of interest, for example,

when a wearable device is shared among a group of

people, and it is desirable to distinguish the person

who is wearing or using the device. It can be espe-

cially helpful in sports teams, where activity tracking

devices are shared among different members and an

automatic detection of the person wearing the device

at a given moment is desired to automatically stream

data to her proﬁle. Hence, it has become really im-

portant to identify the person who uses the device at

that moment.

In this work, we focus on person identiﬁcation us-

ing multiple inertial sensory devices mounted on dif-

ferent locations on human body. We try to answer two

questions:

Adel, O., Nafea, Y., Hesham, A. and Gomaa, W.

Gait-based Person Identiﬁcation using Multiple Inertial Sensors.

DOI: 10.5220/0009791506210628

In Proceedings of the 17th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2020), pages 621-628

ISBN: 978-989-758-442-8

621

1. Which body locations/placements for an inertial

sensory device, are the most indicative of a person

identity assuming the action of ‘walking’?

2. What is the impact of fusing pairs of those sensory

devices on the ﬁnal classiﬁcation performance?

To answer these two questions, we constructed a

new dataset, EJUST-GINR-1

, of 20 persons using

8 inertial wearable devices mounted on 8 ﬁxed loca-

tions on each person’s body. We used three differ-

ent machine learning models for person identiﬁcation:

Random Forest Classiﬁer, Support Vector Machine

(SVM), and Convolutional Neural Network (CNN).

Each was trained in two different ways: on 8 single

sensory device data in order to investigate the impact

of differing the location of the sensor on the classiﬁ-

cation performance, and on a fused pairs of devices,

to explore the effect of sensor fusion on person iden-

tiﬁcation.

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

tion 1 is an introduction. We present the related work

in Section 2. Section 3 is a detailed demonstration of

the data collection and cleaning process. In Section 4

we demonstrate our data pre-processing and feature

selection procedures, in Section 5 we present our ex-

periments including our machine learning models and

the metrics used for evaluating them. Then we present

our results in Section 6, and ﬁnally Section 7 is our

conclusion.

2 RELATED WORK

(Kwapisz et al., 2010) collected 3-axis accelermoeter

data from different Android smartphones inserted in

the pockets of 36 users while they perform different

activities including walking, then they divide the col-

lected data into 10-second examples, do feature ex-

traction using six basic statistical features on each

axis ending up with 43-D feature vector, and feed this

feature vector to two classiﬁers: a decision tree and a

neural network. Their work showed that person iden-

tiﬁcation was possible with accuracy up to 90% on

walking data using neural networks.

(Ngo et al., 2014) presented the largest inertial

sensors gait database for personal authentication. The

authors collected data from 744 subjects with ages

ranging from 2 to 78 years, an almost balanced male-

to-female ratio - using 4 sensors, 3 IMUZ sensors on

the left, right and center of the waist, in addition to a

smartphone on center of the waist. Two 6-D 1-min se-

quences were recorded for each subject during walk-

ing a designed path on his normal speed. The path

Dataset and code are available upon request.

has a ﬂat and a slope ground, and one of the two se-

quences was recorded on the slope ground. Then, for

authentication, a template (or gallery) was created for

the owner’s data and to decide if a new sequence of

data (a probe pattern) belonged to the same person, a

distance function was used to calculate the difference

between the gallery and the probe pattern and if the

difference was below a certain threshold, the user was

considered the legitimate one.

(Cola et al., 2016) demonstrated a light-weight

method for gait-based user authentication using data

collected from an accelermoeter wrist sensor. A pro-

ﬁle of a user is created on ﬁrst-time use, then anoma-

lies in acceleration data could detect a stranger us-

ing the device. Evaluated successfully with 15 vol-

unteers with Equal Error Rate (EER) of 2.9%, they

showed that person authentication using wrist-worn

device can achieve high accuracy.

(Mondol et al., 2017) used a wrist-worn ac-

celerometer and gyroscope sensors embedded in a

smartwatch while doing a gesture in the air to authen-

ticate the user of the smartwatch. The purpose of this

work was to provide an alternative method for authen-

ticating the true owner of a smartwatch or a smart-

phone using a unique gesture. Each participant in this

study performed a unique gesture in the air, such as

performing a signature, a number of times to create

a template for his unique gesture. Then, a new user

would try to replicate his signature, and the Dynamic

Time Warping (DTW) was used to measure the devi-

ation between the two data gestures. The signature is

accepted only if the deviation is below a predeﬁned

threshold.

As far as we know, this is the ﬁrst work done on

gait-based person identiﬁcation using multiple inertial

sensory devices mounted on different upper and lower

body parts, and the ﬁrst to study the impact of the lo-

cation and fusion of sensory devices at different loca-

tions on the ﬁnal identiﬁcation performance, as well

as the ﬁrst to provide a dataset collected for multi-

sensory gait-based person identiﬁcation with sufﬁ-

cient number of samples and subjects - publicly avail-

able for the research community.

3 DATASET CONSTRUCTION

We recorded walking activity data from inertial sen-

sory devices mounted on 8 locations (Figure 1) on 20

subjects, to use these data to identify the person wear-

ing the sensory devices.

All the participants in data collection of EJUST-

GINR-1 dataset are students (undergraduate and post-

graduate) at our university, with ages ranging from

ICINCO 2020 - 17th International Conference on Informatics in Control, Automation and Robotics

622

19 − 26, weights from 56 − 130 kg, heights from

146 − 187 cm, and a 1:1 male-to-female ratio. Table

1 shows an overview of the subjects in our dataset.

Table 1: An overview of the subjects in EJUST-GINR-1.

Features Min Max Mean Stdv

Age 19 26 21.15 2.30

Height (cm) 146 187 169.4 10.15

Weight (kg) 56 130 76.37 16.46

Lap time (min) 4 4.3 4.2 0.025

Figure 1: Locations of sensory devices on a subject’s body.

Figure 2: Data capture system. Two Apple watches on the

left, an iPhone 7 and 6 MMR devices in the middle. Three

velcro sleeves on the right.

We used two types of sensory devices in our data

capture system: Apple Watches Series 1 and Mbient-

Lab MetaMotionR (MMR) sensory devices. Two Ap-

ple watches were worn on the right and on the left

wrists (RW and LW) of the subjects, and 6 MMR

sensory devices were mounted on the following 6 lo-

cations on each subject’s body: the right upper arm

(RUA), the left upper arm (LUA), the back, the waist,

the left and right calfs (LC and RC). These locations

are illustrated in Figure 1. Additionally, we have used

an iPhone 7 in the recording activity, as all MMR de-

vices had to be connected to it during the recording

session until the data recorded is transferred, and be-

cause the data collected on the Apple watches had

also to be transferred to the iPhone after the end of

the data recording activity. Figure 2 shows the entire

data capture system described above.

Both Apple watches and MMR devices came

packed with a large number of embedded sensors

such as: 3-axis accelerometer, 3-axis gyroscope, 3-

axis magentometer, pressure sensor, humidity sensor

and others. However, we only considered the record-

ings of the 3-axis accelerometer and 3-axis gyroscope

from both the Apple watch and MMR devices at sam-

pling rate of 50 Hz.

For every recording session, each subject walked

a path (around our university’s HQ building) of about

330 meters long taking about 4 minutes on aver-

age. Each subject recorded 4 sessions with a total

of about 16 minutes of walking activity, and since

we recorded 4 laps for each of the 20 subjects, our

dataset is composed of about 5 hours of walking ac-

tivity data collected from 8 different sensory devices

synchronously.

3.1 Data Cleaning

Figure 3 shows the ﬁrst 30 seconds of the 3-axis ac-

celerometer’s signals recorded from the sensory de-

vice on the Left Upper Arm (LUA) of two randomly

selected subjects (number 13 and 14). From this ﬁg-

ure, it is obvious that there is a period (about 600 sam-

ples or 12 seconds) in the start of each session where

there is no activity done. This period of non-walking

activity is problematic since it will be fed to our mod-

els as walking activity. Certainly, this will confuse

our models and decrease their performance and con-

sequently, they must be removed. This was done by

manually inspecting the recording session of every

sensory device for every subject then trimming the

parts that showed no-activity or high levels of noise.

The ﬁnal structure of our dataset is as follows:

each sensory device has a separate folder containing

20 ﬁles, each representing the data collected from one

subject and each subject’s ﬁle contains all 4 recording

sessions appended to each other. We used the Numpy

framework to store each subject’s data in .npy format,

as Numpy is the most popular Python framework for

multi-dimensional array manipulation. Each .npy ﬁle

is a M × 6 matrix, where M represents the number of

entries/samples per subject while 6 corresponds to the

concatenation of 3-axis accelerometer and 3-axis gy-

roscope columns.

Gait-based Person Identiﬁcation using Multiple Inertial Sensors

623

Figure 3: The ﬁrst 30 seconds of two sequences of walking activities of two randomly selected subjects recorded from the

Left Upper Arm (LUA) device.

4 METHODOLOGY

4.1 Data Preprocessing

We devised a preprocessing stage composed of 4

steps: data loading and labeling, segmentation, shuf-

ﬂing and one-hot encoding. In the data loading and

labeling step, given a speciﬁc sensory device, we load

all the ‘.npy’ ﬁles for all subjects in the given device’s

folder, then we concatenate the corresponding labels

to them. Next is the segmentation step, where we di-

vide the loaded data into non-overlapping t-second

segments. Since we use 50 Hz as the sampling rate

across all devices and recording sessions, the total

number of segments M resulting from segmentation

of the data into t-second segments can be calculated

from the following formula:

M =

50 ∗t

(1)

where L is the length of samples recorded for all sub-

jects given a single sensory device. Based on the work

in (Murray et al., 1964) the average duration of one

gait cycle for men with ages from 20 to 65 is about

1.03 seconds. Since our subjects are both men and

women, and we use sensors on many different loca-

tions, we decided to set the duration of a single seg-

ment to be 5 seconds (t = 5) to make sure that each

segment contains enough information for sound clas-

siﬁcation. Therefore, given that each sensory device

recorded about 868511 samples (for all subjects) on

average, the total number of segments for each device

is M = 868511/(5 ∗ 50) ≈ 3474. After that, we shuf-

ﬂed the data segments randomly because they were

ordered ascendingly according to the labels (from 0

to 19) and randomization would help the models train

better without bias. Furthermore, some classiﬁers

such as Convolutional Neural Networks require the

labels to be in one-hot encoding format, hence, we

had to convert the labels from a M-D vector to a

M × 20 one-hot encoded matrix, where M is the num-

ber of segments and 20 is the number of classes.

The result of the preprocessing stage is a list X of

M segments (calculated from Equation 1), each seg-

ment is a 250 × 6 matrix, where 250 is the number of

samples per 5-second segment and 6 is the number of

features/columns/axes per segment - in addition to a

M ×20 one-hot encoded labels matrix.

4.2 Feature Extraction (FE)

Given that each segment in list X is 250 × 6 matrix,

the dimension of the input space to our machine learn-

ing models must be in R

1500

which is huge. Moreover,

each segment represents the raw data collected from

the inertial sensors which are not useful on them-

selves. Hence, an additional step to reduce the dimen-

sionality of the input space and to extract meaningful

features is required. In this work, we use the autocor-

relation statistical feature extractor based on the work

done in (Gomaa et al., 2017).

The Auto-Correlation Function (ACF) of a sam-

pled time-series signal x

calculates the Pearson cor-

relation between x

and a shifted version of itself x

t+h

by a lag h, assuming x

is weakly stationary. This

can give us information about the dependency of the

signal to its shifted version (Shumway and Stoffer, ),

which can be particularly useful in periodic activities

such as walking. Therefore, given different input lags,

the Auto-Correlation Function can be calculated for

each lag h using the following formulas:

ICINCO 2020 - 17th International Conference on Informatics in Control, Automation and Robotics

624

ac f

(h) =

(h)

(0)

(2)

(h) =

T −h

∑

t=1

t+h

− ¯x)(x

− ¯x) (3)

where ¯x is the sample mean of the signal x and T is

the length of the sampled signal x.

Assuming that our input data sample has K fea-

tures/columns/axes, we can calculate the ACF over

each column in the input segment over a set of dif-

ferent lags. The output will be a scalar for each lag

that summarizes the T entries (where T = 250) along

one column of the input segment. We calculate the

ACF for each one of the K columns for N lags (start-

ing from N = 0), which will result in a K(N + 1)-D

feature vector as follows:



ac f

(0)... ac f

(0) ac f

(1)... ac f

(N)



To calculate the ACF over all M segments in our

input list X for a set of N lags, which would result in

a M × K(N + 1) features matrix, we implemented the

following algorithm (Algorithm 1):

Algorithm 1: Feature Extraction Algorithm.

input : List X, N ∈ R

output:

X ∈ R

M×KN

Initialize

X with zeros, i ← 0;

foreach s ∈ X do

for h ← 0 to N do

for k ← 0 to K − 1 do

i,Kh+k

= ac f

(h) (From Eq. 2)

end

i ← i + 1;

end

5 EXPERIMENTS

We conducted two experiments: the ﬁrst aims at ex-

ploring the effect of a single location/placement of

an inertial sensory device on the ﬁnal person identi-

ﬁcation performance, whereas the goal of the second

was to investigate how fusing different pairs of iner-

tial sensory devices worn on 8 different locations on

the human body while walking (Figure 1) - affects the

person identiﬁcation performance.

In both experiments, we started by loading the

data (individual devices in the ﬁrst experiment, and

pairs of devices in the second), then we pre-processed

the loaded data as discussed in Section 4.1, did fea-

ture extraction as shown in Section 4.2 (choosing the

number of columns K to be 6 in the ﬁrst experiment,

and to 12 in the second), and the number of lags N

to be 10 based on the work done in (Gomaa et al.,

2017), split the features matrix and the correspond-

ing labels into training and test sets (80 − 20% train-

test split), trained 3 classiﬁcation models: a Random

Forest (criterion for a split is the gini-index, number

of grown trees = 100, minimum number of samples

to split a node = 2 and using bootstrapping), a Sup-

port Vector Machine (linear kernel, hinge as the loss

function, regularization parameter C = 1, maximum

number of iterations = 1000 and One-Vs-The-Rest as

the multi-class SVM implementation) and a Convo-

lutional Neural Network (as shown in Figure 5) - us-

ing the training set, evaluated the models performance

using two metrics: accuracy and f1-measure - on the

test set and eventually stored the evaluation metrics

results for each model for further analysis.

The up-mentioned pipeline (illustrated in Figure

4) was repeated for every sensory device (in the ﬁrst

experiment) and each 28 possible pairs of sensors (in

the second experiment) saving the results separately.

Moreover, to ensure that our results are consistent, we

ran the previous steps 15 times, each on a different

and random train-test split (again, 80 −20% train-test

split), then taking the average of the results over the

15 trials.

6 RESULTS AND DISCUSSION

In order to interpret the results obtained in both exper-

iments, a Fusion Matrix has been constructed, where

every entry f

i, j

represents the score achieved by fus-

ing device i with device j while the entries at the di-

agonal where i = j represent the scores of individual

devices i obtained from the ﬁrst experiment. The Fu-

sion Matrix is a symmetric matrix since the fusion of

device i with device j is the same in reverse. For each

classiﬁer of the 3 we used, two Fusion Matrices have

been constructed, one for the accuracy and another for

the F1-measure.

By examining the diagonal of the 6 Fusion Ma-

trices shown in Figure 6 and 7, it is clear that the

sensory devices mounted on either the left calf (LC)

or right calf (RC) achieved the highest accuracy and

F1-measure. Using CNN, the highest accuracy and

F1-measure were achieved on the LC with 98.6% and

0.99, whereas the accuracy and F1-measure achieved

using CNN on the RC were 97.3% and 0.97, respec-

tively. On the other hand, SVM achieved 96.8% accu-

racy on LC, 96.4% accuracy on RC, 0.97 F1-measure

on LC and 0.96 F1-measure on RC. Our Random For-

est model , despite being simple, achieved 95% ac-

Gait-based Person Identiﬁcation using Multiple Inertial Sensors

625

Figure 4: Stages of our experiments pipeline.

Figure 5: CNN architecture.

curacy on LC and 95.5% on RC, in addition to 0.95

F1-measure on both LC and RC. Likewise, the device

mounted on the waist was able to score 96.9% accu-

racy and 0.97 F1-measure using CNN. At about 94%

accuracy, SVM was the second best before Random

Forest, which achieved 92.7% accuracy on the same

device.

One the other hand, the sensory devices worn

on both the left wrist (LW) and right wrist (RW)

showed the worst results. Again, CNN performed bet-

ter than SVM and Random Forest, achieving 82.8%

accuracy for the LW and 82.3% for the RW. Simi-

larly, it achieved 0.83 and 0.82 F1-measures on the

LW and RW, respectively. The performance of the

other two models was close, with an edge to SVM at

about 77% and 76.4% accuracy for the LW and RW

whereas Random Forest achieved 75.6% and 74% ac-

curacy, respectively. The F1-measure of both models

achieved similar results.

Surprisingly, the identiﬁcation accuracy achieved

using the data collected from the sensory device on

the back was about 96.8% with CNN, 94% with SVM

and 93% with Random Forest - despite being the fur-

thest from the lower body sensory devices. The same

goes with F1-measure with 0.97, 0.94 and 0.93, re-

spectively. This is the third best accuracy achieved

over all the devices, which means that the gait pat-

tern was clearly captured by this device, and that this

location is the best in the upper body for person iden-

tiﬁcation.

The classiﬁcation accuracy of the fusion of two

sensory devices can be seen in the off-diagonal en-

tries in the 3 Fusion Matrices shown in Figure 6. As

we can see, in all cases the fusion of two sensory de-

vices results in an overall accuracy higher than that of

the single lower-accuracy device. For example, The

Random Forest Classiﬁer accuracy on data collected

from the device on the RW individually is about 74%,

however, when combined with any other device, the

accuracy drastically increases, up to 96% when com-

bined with either the RC or LC. The same is true with

SVM, as the accuracy jumps from 76% for RW up to

98% when combined with LC, which is substantially

higher than the average of their combined accuracy

(86.5%).

On the other side, this means that combining one

of the devices with very high individual accuracy with

another very low individual accuracy can result in a

degradation of the classiﬁcation power gained from

the high-accuracy device. Speciﬁcally, fusing one of

the devices mounted on the lower body with another

device worn on the upper body results, in most cases,

ICINCO 2020 - 17th International Conference on Informatics in Control, Automation and Robotics

626

Figure 6: Accuracy Fusion Matrix for Random Forest, SVM and CNN.

Figure 7: F1-Measure Fusion Matrix for Random Forest, SVM and CNN.

in a lower accuracy than that of the individual lower

body device. For example, the device worn on the LC

achieves 95%, 97% and 99% accuracy with Random

Forest, SVM and CNN, respectively. When combin-

ing this device with RUA or LUA devices, the ac-

curacy drops to 94% with Random Forest Classiﬁer,

96% with SVM and 95% with CNN.

Therefore, the best fusion of devices will be be-

tween two lower-body devices especially, LC and RC,

since they produce up to 97% accuracy even with sim-

ple classiﬁers such as Random Forest, and up to 98%

accuracy with SVM and CNN. However, fusing the

data from the device on the back with data from ei-

ther RC or LC also results in an accuracy up to 98%

with SVM. Besides, fusing either RC or LC devices

with RW or LW devices appears to achieve higher ac-

curacy than fusing RC or LC with RUA or LUA, up

to 98% (LC + RW under SVM).

From the previous results, it can be deduced that

the devices mounted on the lower part of the body

achieve better classiﬁcation performance in general

and are more indicative of the person wearing them

while walking - than the devices mounted on the up-

per body parts. This can be attributed to the fact

that people have more freedom to move their hands

and arms while walking in different ways than they

have with their legs and waist while walking, which

may induce noise in the inertial sensors readings from

these locations. Possibly, being able to extract dis-

criminative features from such noisy data might help

increase the accuracy of models trained on such data.

However, recording data from an inertial sensory de-

vice placed on the back (of the neck) achieves second-

best performance after sensory devices placed on the

lower body. One explanation might be that the gait

patterns were transmitted almost intact through the

backbone up to this location.

Generally, SVM achieves better results compared

to CNN in the fusion of pairs of devices. This

may have stemmed from the fact that we merge the

columns in the input data in order to merge two de-

vices. However, the architecture of the CNN was de-

signed to accommodate only 6 columns. This is a

result of using the Global Max Pooling layer which

takes the maximum value out of each feature map and

returns a vector with size equal to the number of fea-

ture maps. Since the architecture of the network is

the same, doing the same convolution operations on

a larger input feature space results in a larger feature

maps at the last layer, and taking the maximum di-

rectly may have resulted in a great loss of information

and as a result, loss in classiﬁcation power as well. On

the other hand, the SVM we used is with a linear ker-

nel thus, it is a linear classiﬁer, which beneﬁts from

Gait-based Person Identiﬁcation using Multiple Inertial Sensors

627

the increase in the number of features as it adds more

separability power to it. The F1-measure closely fol-

lows the accuracy in almost all cases. This means that

our models correctly classify each person correctly

without any bias towards a single class/individual.

7 CONCLUSIONS AND FUTURE

WORK

In this work, a novel multi-sensory dataset was pre-

sented with detailed demonstration of all the steps

taken for its construction. Additionally, three differ-

ent machine learning algorithms were used to eval-

uate the new dataset, and to explore the correlation

between the on-body location of sensory device -

with the model accuracy as well as, the effect of fus-

ing pairs of sensory data from different body loca-

tions on the resultant predictive performance. It turns

out that, the sensory devices mounted on the lower

body achieved better performance (up to 99% accu-

racy) than sensors worn on the wrist or on the upper

arm. However, the sensory device mounted on the

back achieved a comparable accuracy despite being

far away from the gait.

It was also shown that fusing sensory data from

two different locations help increasing the classiﬁca-

tion accuracy above the low-accuracy single location,

and increases the accuracy signiﬁcantly if fused with

a device mounted on one of the three locations in the

lower body (up to 98%).

Finally, the data collected in this work included

only walking activity. However, in the future, we

would like to explore more activities such as: brush-

ing teeth, climbing stairs, jogging, etc., and investi-

gate if they can constitute bio-metric features for dif-

ferent people. We would also like to study the impact

of sensory devices’ location on the identiﬁcation per-

formance for each activity.

ACKNOWLEDGEMENTS

This work is funded by the Information Technol-

ogy Industry Development Agency (ITIDA), Infor-

mation Technology Academia Collaboration (ITAC)

Program, Egypt – Grant Number (PRP2019.R26.1 - A

Robust Wearable Activity Recognition System based

on IMU Signals).

REFERENCES

Abdu-Aguye, M. G. and Gomaa, W. (2019). Competitive

Feature Extraction for Activity Recognition based on

Wavelet Transforms and Adaptive Pooling. In 2019

International Joint Conference on Neural Networks

(IJCNN), pages 1–8. ISSN: 2161-4393.

Cola, G., Avvenuti, M., Musso, F., and Vecchio, A. (2016).

Gait-based authentication using a wrist-worn device.

In Proceedings of the 13th International Conference

on Mobile and Ubiquitous Systems: Computing, Net-

working and Services, MOBIQUITOUS 2016, page

208–217, New York, NY, USA. Association for Com-

puting Machinery.

Gomaa, W., Elbasiony, R., and Ashry, S. (2017). Adl

classiﬁcation based on autocorrelation function of

inertial signals. In 2017 16th IEEE International

Conference on Machine Learning and Applications

(ICMLA), pages 833–837.

Kwapisz, J. R., Weiss, G. M., and Moore, S. A. (2010).

Cell phone-based biometric identiﬁcation. In 2010

Fourth IEEE International Conference on Biometrics:

Theory, Applications and Systems (BTAS), pages 1–7.

ISSN: null.

Lara, O. D. and Labrador, M. A. (2013). A Sur-

vey on Human Activity Recognition using Wearable

Sensors. IEEE Communications Surveys Tutorials,

15(3):1192–1209.

Mondol, M. A. S., Emi, I. A., Preum, S. M., and Stankovic,

J. A. (2017). Poster Abstract: User Authentication

Using Wrist Mounted Inertial Sensors. In 2017 16th

ACM/IEEE International Conference on Information

Processing in Sensor Networks (IPSN), pages 309–

310. ISSN: null.

Murray, M., Drought, A., and Kory, R. (1964). Walking

patterns of normal men. The Journal of bone and joint

surgery. American volume, 46:335—360.

Ngo, T. T., Makihara, Y., Nagahara, H., Mukaigawa, Y.,

and Yagi, Y. (2014). The largest inertial sensor-

based gait database and performance evaluation of

gait-based personal authentication. Pattern Recogni-

tion, 47(1):228–237.

Patel, S., Park, H., Bonato, P., Chan, L., and Rodgers, M.

(2012). A review of wearable sensors and systems

with application in rehabilitation. Journal of Neuro-

Engineering and Rehabilitation, 9(1):21.

Shumway, R. H. and Stoffer, D. S. Time series analysis and

its applications : with R examples. Springer, Cham.

Tao, S., Zhang, X., Cai, H., Lv, Z., Hu, C., and Xie, H.

(2018). Gait based biometric personal authentication

by using MEMS inertial sensors. Journal of Ambient

Intelligence and Humanized Computing, 9(5):1705–

1712.

ICINCO 2020 - 17th International Conference on Informatics in Control, Automation and Robotics

628