BioDeep: A Deep Learning System for IMU-based Human Biometrics

Recognition

Abeer Mostafa

1 a

, Samir A. Elsagheer

1,3 b

and Walid Gomaa

1,2 c

Cyber-Physical Systems Lab, Egypt Japan University of Science and Technology, Alexandria, Egypt

Faculty of Engineering, Alexandria University, Alexandria, Egypt

Faculty of Engineering, Aswan University, Egypt

Keywords:

IMU, Transfer Learning, Convolutional Neural Networks, Age and Gender Recognition, Deep Learning.

Abstract:

Human biometrics recognition has been of wide interest recently due to its beneﬁts in various applications

such as health care and recommender systems. The rise of deep learning development, together with the mas-

sive data acquisition systems, made it feasible to reuse models trained on one task for solving another similar

task. In this work, we present a novel approach for age and gender recognition based on gait data acquired

from Inertial Measurement Unit (IMU). BioDeep design is composed of two phases, ﬁrst of which is applying

a statistical method for feature modelling, the autocorrelation function, then building a Convolutional Neural

Network (CNN) for age regression and gender classiﬁcation. We also use random forest as a baseline model to

compare the results achieved by both methods. We validate our models using four publicly available datasets.

The second phase is doing transfer learning over these diverse datasets. We train a CNN on one dataset and

reuse its feature maps over the other datasets for solving both age and gender recognition problems. Our

experimental evaluation over the four datasets separately shows very promising results. Furthermore, trans-

fer learning achieved 20 − 30x speedup in the training time in addition to keeping the acceptable prediction

accuracy.

1 INTRODUCTION

Recognition of human biometrics such as age and

gender has been widely been studied in the recent

decades due to its important use in many applica-

tions such as speech analysis (Markitantov, 2020),

(Albuquerque et al., 2021), recommendation sys-

tems (Sun et al., 2017), and of course health care

applications (Rosli et al., 2017). Researchers use

various types of data for performing age and gen-

der recognition starting from images, voice signals

to inertial measurement unit (IMU) signals. In this

paper, we build our age and gender recognition sys-

tem based on IMU data. IMUs have two essential

sensors: accelerometer and gyroscope and sometimes

more additional sensors are included. The accelerom-

eter sensor produces a tri-axical signal correspond-

ing to the proper acceleration of the moving body

accelerometer-X, accelerometer-Y, accelerometer-Z,

https://orcid.org/0000-0002-8971-4311

https://orcid.org/0000-0003-4388-1998

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

whereas the gyroscope sensor is used to determine the

angular velocity of the moving body, and also pro-

duces a tri-axical signal gyroscope-X, gyroscope-Y,

and gyroscope-Z.

The motivation for this work is that there is a big

gap in research regarding the analysis of human bio-

metrics based on IMU data since most of the literature

uses IMU data for human activity recognition only. In

addition, nowadays, IMUs are embedded in all of the

wearable devices we use in our everyday life so it will

be suitable for our analysis to work on IMU data.

Our contributions are illustrated as follows. We

propose a novel methodology for the analysis of hu-

man biometrics based on IMU data and applying it

speciﬁcally on age regression and gender classiﬁca-

tion using various datasets. In addition, our work is

the ﬁrst of its kind to perform cross-testing (transfer

learning) for age regression and gender classiﬁcation

using IMU data. Furthermore, we measure the tim-

ings for our experiments and report the computational

speedup gained in our approach.

We begin ﬁrstly by applying autocorrelation func-

tion on the accelerometer and gyroscope signals for

620

Mostafa, A., Elsagheer, S. and Gomaa, W.

BioDeep: A Deep Learning System for IMU-based Human Biometrics Recognition.

DOI: 10.5220/0010578806200629

In Proceedings of the 18th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2021), pages 620-629

ISBN: 978-989-758-522-7

feature modeling. The reason for choosing this statis-

tical feature is that it is simple, efﬁcient, and helps

in reducing the input data size which is better in

terms of processing time and memory. Afterwards,

we apply modern machine learning techniques for

age regression and gender classiﬁcation. We begin

by using random forest as it is considered a power-

ful machine learning model. Additionally, we investi-

gate the deep learning approach and compare both re-

sults. In the deep learning approach, we apply a Con-

volutional Neural Network (CNN) on each dataset.

Consequently, We apply transfer learning from one

dataset to another for the sake of minimizing the num-

ber of learning parameters and, hence, reducing train-

ing time dramatically.

We apply our proposed methodology on four pub-

licly available datasets: EJUST-GINR-1 (Mostafa

et al., 2020), OU-ISIR (Ngo et al., 2014), GEDS (Mi-

raldo et al., 2020), and HuGaDB (Chereshnev and

Kert

esz-Farkas, 2017). We begin by applying both

models (random forest and CNN) on each dataset sep-

arately for gender classiﬁcation and age regression,

then we perform cross testing by training the CNN on

one dataset then ﬁne-tune and test on the others.

The rest of the paper is organised as follows. In

section 2, we review the state of the art research work

that have been published in the area of age and gender

estimation and also in the area of transfer learning. In

section 3, we provide a brief description for all of the

four considered datasets. In section 4, we illustrate

our proposed methodology in detail. In section 5, we

explain the setup for our experiments. In section 6, we

show and discuss the results achieved by our method-

ology. Finally, we summarize our work in section 7

and conclude the paper.

2 RELATED WORK

In this section, we brieﬂy provide a literature review

on the latest work that have been published related to

two research areas. Firstly, research that addresses

solving the problem of age and gender estimation.

Secondly, research in the area of deep learning ap-

plied to solve classiﬁcation or regression problems us-

ing IMU signals with focus on doing transfer learning

among various datasets.

2.1 Age and Gender Recognition

Age and gender recognition has been a common re-

search area for many years. Almost all types of

data have been used in research to identify these

characteristics for various applications. The authors

in (Mostafa et al., 2020), proposed a robust method

for gender recognition based on IMU data acquired

from 8 sensors placed in 8 different positions on

the human body during gait activity. The proposed

method included using wavelet transform as a feature

extractor along with various classiﬁers. They eval-

uated their approach on EJUST-GINR-1 dataset and

had successfully reached accuracy up to 96.74% us-

ing the sensor placed on the left cube.

With reference to (Ngo et al., 2019), a competi-

tion on gender and age recognition was conducted us-

ing OU-ISIR Gait dataset (Ngo et al., 2014), which

includes IMU signals extracted during gait activity.

According to the competition results, most of the par-

ticipated teams got relatively low accuracy in gender

classiﬁcation. However, the results of age regression

were obviously better. The best results were presented

in (Garofalo et al., 2019) which showed accuracy of

gender classiﬁcation reaching 75.77% and age regres-

sion with mean absolute error equals to 5.3879 by

using orientation independent AE-GDI representation

along with a CNN.

In the work proposed by (Riaz et al., 2015), a sys-

tem for gender, age, and height estimation was pre-

sented based on IMU gait signals of four sensors lo-

cated on the moving body. The authors used ran-

dom forest classiﬁer along with two validation meth-

ods: 10-fold cross-validation and subject-wise cross-

validation. They categorized the age to three groups

to perform age classiﬁcation: less than 40 years, be-

tween 40 and 50, and older than 50 years. The high-

est accuracy for gender classiﬁcation was achieved by

the chest sensor and equals to 92.57%. Regarding

age classiﬁcation, the accuracy reached 88.82% using

chest and lower back sensors.

The authors in (Jain and Kanhangad, 2016) inves-

tigated gender recognition using accelerometer and

gyroscope data from the built-in smartphone sensors.

The authors used multi-level local pattern (MLP) and

local binary pattern (LBP) as feature extractors. To

classify the extracted features, the authors used sup-

port vector machine (SVM) and aggregate bootstrap-

ping (bagging). To evaluate these models, 252 gait

signals collected from 42 subjects were used. The ﬁ-

nal result for gender classiﬁcation reached 77.45% by

applying MLP along with bagging.

2.2 Deep Learning and Transfer

Learning on IMU Data

Research in deep learning has gained a huge popular-

ity in the recent few years and proved to be a powerful

methodology for solving machine learning problems.

Some researchers applied deep learning and transfer

BioDeep: A Deep Learning System for IMU-based Human Biometrics Recognition

621

learning on IMU signals to perform classiﬁcation or

regression tasks which resulted in good performance.

The authors in (Abdu-Aguye and Gomaa, 2019)

proposed a robust method for doing transfer learn-

ing over IMU data in the domain of human activity

recognition. Their approach was to train a CNN on

one dataset and use the convolutional ﬁlters as fea-

ture extractor, then train a feed forward neural net-

work as a classiﬁer on the extracted features for other

datasets. This method proved to be very promising

as the results were within 5% compared to the CNNs

which were trained from scratch, end-to-end, with

time speedup of 24 − 52x.

In (Fu et al., 2021), the authors designed a com-

pact wireless wearable sensing node that combines

an air pressure sensor and IMU to be used for hu-

man activity recognition. Their method is to ap-

ply a transfer learning algorithm that consists of a

joint probability domain adaptive method with im-

proved pseudo-labels (IPL-JPDA). The authors used

this method to recognise 7 daily human activities.

The average recognition accuracy of different subjects

reached 93.2%.

With reference to (Du et al., 2019), the authors

used cascade learning and compared it to end-to-end

learning in many transfer learning experiments ap-

plied to human activity recognition for IMU data.

The method proved to be robust and reliable. They

achieved 15% improvement of F1 score compared to

other methods.

The authors in (Ashry et al., 2020) Proposed a

deep learning method called CHARM-Deep to per-

form ofﬂine/online continuous human activity recog-

nition based on IMU data streams collected from

smartwatches. They built a cascaded bi-directional

long short term memory (Bi-LSTM) to classify the

feature vector extracted from IMU data. Statisti-

cal features such as autocorrelation, entropy, median

were extracted from the signals, then fed to the Bi-

LSTM. The proposed method achieved high accuracy

of 94.2% to 97.2% in classiﬁcation of all activities.

In addition, it reduced the processing time by 86%

compared to training Bi-LSTM on raw data with also

space reduction by 97.7%.

From the previous literature, it is apparent that

deep learning methods including transfer learning can

demonstrate very good performance on accelerometer

and gyroscope signals. However, these methods were

only used in human activity recognition. In this work,

we investigate the use of deep learning and, in particu-

lar, transfer learning on accelerometer and gyroscope

data to be applied for age and gender estimation and

seeing if deep learning can learn the features that dis-

tinguish the gender of a person and the features that

accurately estimate their age during the particular ac-

tivity of “walking”. In addition, we investigate doing

transfer learning (cross-testing) among four datasets

to see if the learned feature space from one dataset

can effectively be used on another dataset and gain

time speedup compared to training the network on

each dataset from scratch.

3 DATASETS

Before we illustrate our proposed methodology, we

provide a brief description of the datasets we used to

run our experiments to its test the validity and effec-

tiveness. We used four publicly available datasets.

Each dataset was collected in a different place with

different environmental setup. Furthermore, the

datasets were collected using different devices, so

they produce other modalities beside accelerometer

and gyroscope, however, we consider only accelerom-

eter and gyroscope signals in our analysis. These two

sensors are available in all of the considered datasets,

so it will make possible to do transfer learning. An-

other variation among the datasets is sensor place-

ment, meaning that the IMU sensors were placed dif-

ferently in each dataset, however, we consider the

placements they have in common. For example, if

a dataset has all sensors placed on legs only and the

other dataset has sensors placed on arms and legs, we

consider the data of leg sensors only to do transfer

learning. In the following four subsections, we brieﬂy

explain each dataset.

3.1 EJUST-GINR-1 Dataset

EJUST-GINR-1 dataset (Mostafa et al., 2020), (Adel

et al., 2020) was collected using six IMU sensors

called MetaMotionR in addition to two Apple smart

watches series-1 at each hand. The sensor placement,

as shown in Figure 1, are: waist, back, left upper arm,

right upper arm, left cube, and right cube. The data

acquisition system was able to synchronize the eight

sensing elements together when a subject was walk-

ing and capture the accelerometer and gyroscope sig-

nals with 50Hz frequency.

Twenty subjects participated in collecting this

dataset, 10 males and 10 females, with age ranging

from 19 years to 33 years. The procedure of collect-

ing the data was that each person walks naturally on a

straight ground for 20 minutes decomposed into small

sessions. The total number of samples in the dataset

is 5, 292 ﬁxed-length samples.

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

622

Figure 1: EJUST-GINR-1 dataset sensor placement.

3.2 OU-ISIR Gait Dataset

The OU-ISIR gait dataset (Ngo et al., 2014) is the

largest IMU gait dataset in the world. The dataset was

collected using three IMU sensors named as IMUZ

and a Motorola smartphone placed on a belt around

the waist of the person. The dataset includes a tri-

axial accelerometer and a triaxial gyroscope gait sig-

nals of 744 subjects (389 males and 355 females) with

age ranging from 2 years to 78 years. The data was

recorded at frequency of 100Hz. The dataset was col-

lected in an exhibition and formulated for many dif-

ferent research purposes. Each person provided two

signals of level walk gait. In addition, another proto-

col of the dataset was released that has up-slope and

down-slope data for 495 subjects. However, in this

paper, we consider only level walk data in order to

be used in transfer learning (cross-testing) with other

datasets. The total number of samples for level walk

only is 1, 488 samples. We obtained the permission to

use this dataset in our research by a signed agreement

from EJUST university to Osaka university.

3.3 Gait Events DataSet (GEDS)

The Gait Events DataSet (GEDS) (Miraldo et al.,

2020) is a publicly open dataset which was collected

using six wireless sensors. Four of them produce ac-

celerometer and gyroscope signals with other modal-

ities, and two of them are force-sensitive resistor

(FSR) sensors. In this paper, we work with the data

captured from the sensors placed at the tibialis ante-

rior muscles in the right and left legs which are named

(TaR) and (TaL), and the sensors placed over the tibia

bones at the right and the left legs which are named

(TbR) and (TbL). The number of subjects who partic-

ipated in collecting the dataset is 22 (10 males and

12 females) including one female with a foot drop

gait abnormality. The ages of the subjects range from

18 years to 50 years. The dataset contains a total of

9, 661 gait strides. It also contains gait data corre-

sponding to 3 different speeds: slow walk, fast walk,

and comfortable walk. In this work we consider the

comfortable gait style.

3.4 Human Gait Database (HuGaDB)

The Human Gait Database (HuGaDB) (Chereshnev

and Kert

esz-Farkas, 2017) consists of recordings for

12 human activities such as walking, running, sitting,

standing and so on. The dataset was collected us-

ing six IMU sensors that produce accelerometer and

gyroscope signals in addition to two electromyogra-

phy (EMG) sensors. The IMU sensors were located

at the right and left thighs, shins, and feet. Eighteen

healthy subjects participated in collecting the dataset

(14 males and 4 females), their ages range from 18 to

35 years. The sampling rate of the dataset is 56.35Hz.

The total records of the data equaled 10 hours, then

the dataset was segmented and annotated. In this

work, we consider only the walking activity data and

the IMU sensors.

4 PROPOSED SCHEME

We now describe our proposed methodology to build

a reliable age and gender recognition system. Figure 2

illustrates the stages of our proposed methodology.

We begin by loading the IMU data consisting of triax-

ial accelerometer (accelerometer-X, accelerometer-Y,

accelerometer-Z) and triaxial gyroscope (gyroscope-

X, gyroscope-Y, and gyroscope-Z). The data pre-

processing stage and feature extraction methods are

the same in both gender classiﬁcation and age regres-

sion. After the feature extraction phase, we feed the

feature vector to either a classiﬁer for learning to clas-

sify gender, or to a regressor to learn to estimate age.

4.1 Feature Modeling

To be able to extract useful information out of our raw

data, we propose to ﬁrst apply a statistical method in

the data preprocessing stage. One of the most use-

ful statistical properties in the analysis of timeseries

is the autocorrelation function. The autocorrelation

function is a measure of similarity among the signal

BioDeep: A Deep Learning System for IMU-based Human Biometrics Recognition

623

Figure 2: Proposed BioDeep System Architecture.

and time-shifted versions of itself. It is particularly

useful in the analysis of signals to ﬁnd the repeated

patterns such as gait data in our case as the motion is

repeated periodically. We calculate the autocorrela-

tion function acf for each sensory signal up to a cer-

tain lag which is speciﬁed by experiments. The sam-

ple autocorrelation function is calculated as indicated

in Equations (1) and (2).

ac f (h) =

γ(h)

γ(0)

(1)

γ(h) =

n−h

∑

t=1

t+h

− ¯x)(x

− ¯x) (2)

Where ¯x corresponds to the sample mean, n represents

the signal length, and h is the lag (Gomaa et al., 2017).

The output of the autocorrelation function is a vector

of length (6 x number of lags) as we have 6-axes sig-

nals. This vector is then considered the feature vector

to be fed to a classiﬁer or a regressor. Another beneﬁt

of using the autocorrelation function is dimensional-

ity reduction. At the beginning, our data consists of

6-dimensional timeseries. If we take EJUST-GINR-1

dataset as an example, each sample has 250 sequential

data points which result in dimensions of (6x250) for

each gait sample. However, after applying the auto-

correlation function, it results in a dramatic reduction

in the dimension. From the previous analysis, acf is

best suitable for modeling the features of our data.

4.2 Age Estimation

In this work, we consider age estimation as a regres-

sion problem to approximately estimate the exact age

of a person. A supervised machine learning approach

is employed as all of the considered datasets provide

age information attached to the gait signals. We ﬁx

the feature vector as the output of the autocorrelation

function; after that, we feed this feature vector to a

regressor. We ﬁrst try simple machine learning tech-

nique such as a random forest regressor, then we go

for the deep learning approach in which we use Con-

volutional Neural Network (CNN).

Random Forest is a well-known robust method

in machine learning. As shown in (Mehrang et al.,

2018), (Feng et al., 2015), (Mostafa et al., 2020),

and (Casale et al., 2011), random forest has proven

to be very promising applied to the analysis of IMU

signals. Random forest has such power because it re-

lies on ensemble learning, which means that not only

one big model is created to analyze the features of

the given dataset, but many different smaller models

are created using decision trees. Furthermore, each

model of those not only uses a subset of the data but

also performs random feature selection so as to reduce

the variance of prediction and hence increase accu-

racy (Mostafa et al., 2020). Here, we apply a random

forest technique for doing regression on the age with

model speciﬁcations illustrated in section 5.

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

624

Another method we use here for regression is

CNN. The reason for that is because we want to make

use of the power of deep learning for solving this su-

pervised learning problem and compare its results to

the results of a traditional method such as random

forest. CNNs are known to be good at solving ma-

chine learning problems based on multi-dimensional

data. In our work, we construct the network shown in

Figure 2. The input to the ﬁrst convolutional layer

is the feature matrix resulted from the autocorrela-

tion function which is considered six channels cor-

responding to the six timeseries autocorrelation func-

tions. The network is composed of 2 convolutional

layers in order to create the feature maps for age and

each one is followed by a max pooling layer to cre-

ate a summarized version of those feature maps. Fol-

lowing that, we ﬂatten the output of the second max

pooling layer then feed it to a fully-connected layer.

Consequently, we add a dropout layer for regulariza-

tion to prevent overﬁtting and to boost the computa-

tional performance by minimizing the large number

of learning parameters caused by the fully-connected

layer. Finally, the output layer consists of one neu-

ron that will be produce the prediction for the age.

More speciﬁcation about the model hyperparameters

are provided in section 5.

4.3 Gender Recognition

Gender recognition is considered a binary classiﬁca-

tion problem. We use the same scheme proposed for

age regression as illustrated in Figure 2 except for the

activation function in the output layer of the CNN and

the loss function for estimating the error.

4.4 Transfer Learning

Transfer learning is a machine learning paradigm

which has proved to be very effective recently. It

means that the model which had previously learned

on a speciﬁc task can be reused as a starting point to

learn another semi-similar task which means that the

learning time for the latter task will be reduced sig-

niﬁcantly. In this context, we apply transfer learn-

ing among the four previously mentioned datasets.

First, we train the CNN architecture shown in Fig-

ure 2. The top layers which are highlighted in green,

that include the convolutional and max pooling lay-

ers, learn the features of the input by creating the cor-

responding feature maps. Consequently, the bottom

layers which are highlighted in blue in Figure 2 are

then used to classify the gender or estimate the age ac-

cording to the application at that experiment (keeping

in mind that the activation function at the last layer

and the loss function will be different in regression

as illustrated in the model speciﬁcation in section 5).

The same pretrained top layers are taken after train-

ing on one dataset with the same parameters, then

we attach to them new bottom layers and do ﬁne-

tuning with this new model on another dataset. We

call this “cross-testing” The ﬁne-tuning process is not

like training the network from scratch. In ﬁne-tuning,

we train the model for 2 or 3 epochs only to capture

the features of the new dataset. Afterwards, we eval-

uate the model on the test samples from the second

dataset which the model haven’t seen before.

5 EXPERIMENTAL SETUP

To evaluate the effectiveness of our proposed method-

ology, we apply our model over the four previously

mentioned datasets. Due to the variation among the

datasets, we establish a systematic protocol of our ex-

periments. In this section, we illustrate the speciﬁ-

cations of those experiments. First, we calculate the

acf for all sensor data of each dataset. We take differ-

ent lag values up to 10. For each sample, we have 6-

dimensional signal: accelerometer-X, accelerometer-

Y, accelerometer-Z, gyroscope-X, gyroscope-Y, and

gyroscope-Z. So the resulted feature space is in R

We tried many lag values then chose the best value

that resulted in the highest accuracy which was 10.

Secondly, we train and test random forest and

CNN classiﬁcation models on each dataset separately

for gender classiﬁcation. Consequently, we train and

test random forest and CNN classiﬁcation models on

each dataset separately for age regression. We run

these experiments on each sensor data to compare the

performance and to recognize which body parts best

provide age and gender information. After that, we

perform cross-testing, that is, we train the CNN clas-

siﬁcation model on one dataset then ﬁne tune and test

on another dataset. We take the data from the same

sensor location in both datasets. We begin our training

on EJUST-GINR-1 dataset which includes the largest

number of sensor locations. We use the learned fea-

tures from the shin sensor in EJUST-GINR-1 (RC)

and transfer them to the data of taR sensor in GEDS

dataset. Similarly, we do the same on shin sensor in

HuGaDB which is (RS). The OU-ISIR dataset doesn’t

contain shin sensors, so we train on the waist sensor

data in EJUST-GINR-1 dataset and use the waist sen-

sor in OU-ISIR dataset. After that, we repeat these ex-

periments but with choosing another dataset for train-

ing so that we make a sort of cross testing among

the four datasets. Eventually, we compare the results

to those of random forest. Additionally, we perform

BioDeep: A Deep Learning System for IMU-based Human Biometrics Recognition

625

some timing evaluation experiments in order to ex-

plore the computational speed we gained by applying

transfer learning compared to training from scratch.

The experimental model speciﬁcations for the ran-

dom forest classiﬁer are as follows: we set the number

of decision trees in the model to be 100 trees because

it is a suitable choice to give a fair result as random

forest takes the majority vote. We also use Gini index

as a measure of the split quality. We apply bootstrap

aggregation to select random subsets of the data and

random subsets of the feature set thus prevent over-

ﬁtting. We run each experiment 10 times and write

down the average accuracy.

The CNN model speciﬁcations, shown in Fig-

ure 2, are as follows: The ﬁrst convolutional layer has

16 ﬁlters with stride 1, the second convolutional layer

has 32 ﬁlters with stride 2 and the fully-connected

layer has 64 neurons. These parameters were selected

experimentally by applying a speciﬁc setup and eval-

uating the model performance then choosing the best

evaluated setup. All of the three layers use Rectiﬁed

Linear Unit (ReLU) as activation function whereas

the ﬁnal single-neuron output layer uses sigmoid ac-

tivation in the case of gender classiﬁcation and linear

activation in the case of age regression. We use Adam

optimizer and set the batch size to 10 samples. We

apply binary cross-entropy as a loss function in the

case of gender classiﬁcation and mean absolute error

in the case of age regression. In addition, we scale

the values of the age to be from 0 to 1 in order to

have a better training process and faster convergence.

The code for all of these experiments was uploaded

on GitHub and available upon request.

6 RESULTS AND DISCUSSION

In this section, we include all the results achieved

by applying our methodology to the four mentioned

datasets. We ﬁrstly show the results of each sensor

in each dataset separately for age regression and gen-

der classiﬁcation. Consequently, we show the results

of cross testing i.e., training on a sensor data of one

dataset then testing on the other datasets.

6.1 Results of Age Estimation

The results of applying our approach for age regres-

sion are shown in Figure 3. We refer to EJUST-

GINR-1 dataset sensors right cube as RC, left cube

as LC, left upper arm as LUA, right upper arm as

RUA, left hand as LH, and right hand as RH. For

HuGaDB dataset, there was some corrupted data, so

we consider the right-side sensors only and refer to

right foot as RF, right shin as RS, and right thigh

as RT. For the GEDS dataset, we evaluate the sen-

sors placed at the tibialis anterior muscles in the right

and left legs and refer to them as TaR and TaL, and

the sensors placed over the tibia bones at the right

and the left legs referred to as TbR and TbL. For the

OU-ISIR database, the data was extracted automati-

cally for the subjects using the center IMUZ placed on

waist. Age regression evaluation was done by mea-

suring the mean absolute error. For consistency, the

results shown were achieved by taking the age range

from range 15 to 35 years as this is the most com-

mon range of subjects’ ages in all the datasets and any

other data outside this age range is considered out-

liers. This means we consider in our evaluation all of

20 subjects of EJUST-GINR-1 dataset, all of 18 sub-

jects of HuGaDB, 19 subjects of GEDS, and 385 sub-

jects of OU-ISIR database. Figure 3 shows the results

of using autocorrelation function followed by random

forest regressor in blue and the results of CNN in or-

ange using the same autocorrelation function features

for each sensor in each dataset. This allows us to com-

pare between using a traditional machine learning ap-

proach (random forest) and deep learning. It also al-

lows us to determine which sensor location(s) can be

used best to estimate age.

The overall performance indicates that the mean

absolute error in age estimation for EJUST-GINR-1

dataset lies between 1.9 for the LH sensor and 2.37

for the LUA sensor in the case of using random for-

est, and lies between 0.88 for the RC sensor and 1.8

for the RH sensor in the case of using CNN. For

HuGaDB, the mean absolute error lies between 1.38

for the RF sensor and 1.57 for the RT sensor in the

case of using random forest, and lies between 0.835

for the RS sensor and 0.93 for the RT sensor in the

case of using CNN. For GEDS, the mean absolute er-

ror lies between 2.2 for the TbR sensor and 2.29 for

the TaL sensor in the case of using random forest, and

lies between 1.3 for the TbR sensor and 2.37 for the

TaL sensor in the case of using CNN. For OU-ISIR

database, the result achieved using random forest was

3.99 and 4.28 using CNN.

From these results, we can observe that CNN per-

formed better than random forest in most of the cases

maybe this is due to the fact that CNN convolutional

ﬁlters can capture more complex features such that in

the case of mapping the gait signal to the correspond-

ing subject’s age. The results also show that lower

sensor locations on the body result in better age esti-

mation. That’s reasonable as most of gait patterns can

be featured from leg sensors.

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

626

Figure 3: Age regression results for each dataset evaluated in mean absolute error.

Figure 4: Gender classiﬁcation results for each dataset evaluated in percentage accuracy.

Table 1: Results of cross-testing in age regression evaluated in mean absolute error.

Training Dataset Testing Dataset

EJUST-GINR-1(RC) HuGaDB(RS) GEDS(TaR) OU-ISIR(Waist)

EJUST-GINR-1(RC) 0.883 2.0 2.5982 NA

HuGaDB(RS) 2.5685 0.835 3.6507 NA

GEDS(TaR) 2.232 2.19 1.963 NA

EJUST-GINR-1(Waist) NA NA NA 5.1209

Table 2: Results of cross-testing in gender classiﬁcation evaluated in percentage accuracy.

Training Dataset Testing Dataset

EJUST-GINR-1(RC) HuGaDB(RS) GEDS(TaR) OU-ISIR(Waist)

EJUST-GINR-1(RC) 97.05% 98.13% 96.56% NA

HuGaDB(RS) 95.58% 99.37% 96.2% NA

GEDS(TaR) 95.9% 98.03% 97.47% NA

EJUST-GINR-1(Waist) NA NA NA 67.93%

6.2 Results of Gender Classiﬁcation

The results of applying our approach in gender clas-

siﬁcation are shown in Figure 4. We use the same ab-

breviations mentioned in the previous subsection. Our

metric for evaluation here is the classiﬁcation percent-

age accuracy.

In Figure 4, we can see that the overall per-

formance for EJUST-GINR-1 dataset lies between

90.48% for the RH sensor and 96.44% for the waist

sensor in the case of using random forest, and lies be-

tween 92.08% for the LH sensor and 97.05% for the

RC sensor in the case of using CNN. For HuGaDB,

the classiﬁcation accuracy lies between 99.1% for the

RF and RT sensors and 99.37% for the RS sensor

in the case of using random forest, and lies between

99.37% for the RS and RT sensors and 99.69% for the

RF sensor in the case of using CNN. For GEDS, the

classiﬁcation accuracy lies between 95.44% for the

TbL sensor and 98.37% for the TbR sensor in the case

of using random forest, and lies between 94.94% for

the TbL and TaL sensors and 98.73% for the TaR sen-

sor in the case of using CNN. For OU-ISIR database,

the result achieved using random forest was 69.15%

and 70.43% using CNN. from these results, we can

observe that CNN had a slight improvement in per-

formance compared to random forest.

BioDeep: A Deep Learning System for IMU-based Human Biometrics Recognition

627

6.3 Cross-testing

The results of applying transfer learning, for cross-

testing, in age regression are shown in Table 1. The

diagonal elements represent training and testing over

the same dataset, so their results are considered the

baseline for our experimental comparisons. The off-

diagonal elements represent the results of transfer

learning across different datasets. For example, the

ﬁrst row shows the results of training on EJUST-

GINR-1 RC sensor and testing on other datasets with

the same sensor location. In the last row, we used the

waist sensor in EJUST-GINR-1 datast to train, then

test on OU-ISIR dataset. The overall average loss in

performance in the case of transfer learning is 1.2 in

mean absolute error compared to training and test-

ing on the same dataset, however, when measuring

the training time for each, transfer learning provided

20 − 30x speedup in the training time compared to

training from scratch. The training was performed

on the Nvidia GeForce GTX 1650 GPU with 4GB

of memory. The same conventions are applied in Ta-

ble 2 to show the results of applying transfer learning

in gender classiﬁcation evaluated in percentage accu-

racy. The overall average accuracy loss is 1.4% in

the case of transfer learning compared to the case of

training and testing on the same dataset, however, we

also achieved 20 − 30x speedup in the training time

for gender classiﬁcation.

It can be observed that EJUST-GINR-1 dataset

consistently has the highest transfer learning accuracy

in gender classiﬁcation and age regression. This can

be due to the fact that it consists of long sequences of

gait signals and a balanced age and gender distribu-

tion over the participated subjects causing the convo-

lutional ﬁlters of the CNN to efﬁciently model the fea-

tures. Additionally, it can be observed that although

OU-ISIR dataset has the largest number of subjects,

the results achieved by using it for testing have the

largest error in age regression and poorest accuracy

in gender classiﬁcation. Our reasoning for this is that

the dataset contains two gait sequences for each sub-

ject and each sequence consists of only a few seconds

which may not be enough for capturing the gender

and age pattern features for each subject.

7 CONCLUSION

In this work, we proposed a novel scheme for age and

gender recognition using IMU gait signals. Our de-

sign begins by applying the autocorrelation function

on the triaxical accelerometer and triaxical gyroscope

timeseries for feature modeling. Consequently, we

investigated, using two machine learning techniques

random forest and CNN for age regression and gen-

der classiﬁcation. Furthermore, we applied transfer

learning among datasets to reduce the training time

and validate our model generalizability. We train the

CNN on one dataset then ﬁne-tune and test on other

datasets. We used four publicly available datasets:

EJUST-GINR-1, OU-ISIR, GEDS, and HuGaDB.

The results obtained from our experimental eval-

uation indicate that our proposed methodology yields

a good performance in both age regression and gen-

der classiﬁcation. In addition, the transfer learning

experiments yielded outstanding results compared to

the baseline CNN-based models trained from scratch.

The cross-testing average loss in performance was

1.2 in mean absolute error compared to networks that

were trained from scratch in the case of age regression

and 1.4% loss in classiﬁcation accuracy in the case of

gender classiﬁcation. The speedup gained by trans-

fer learning reached 20 − 30x in the training time. We

believe these results should open the way in using pre-

trained models for age and gender recognition using

IMU timeseries the same way pretrained models are

used nowadays in computer vision.

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

vestigate other various human activities such as sit-

ting, climbing stairs, running, etc. We may also anal-

yse electroencephalogram (EEG) signals to do age

and gender recognition based on brain signals.

ACKNOWLEDGMENTS

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).

REFERENCES

Abdu-Aguye, M. G. and Gomaa, W. (2019). Versatl: Ver-

satile transfer learning for imu-based activity recogni-

tion using convolutional neural networks. In ICINCO

(1), pages 507–516.

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

Gait-based person identiﬁcation using multiple iner-

tial sensors.

Albuquerque, L., Oliveira, C., Teixeira, A., and Figueiredo,

D. (2021). Eppur si muove: Formant dynamics is

relevant for the study of speech aging effects. 14th

BIOSIGNALS, Online.

Ashry, S., Ogawa, T., and Gomaa, W. (2020). Charm-deep:

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

628

Continuous human activity recognition model based

on deep neural network using imu sensors of smart-

watch. IEEE Sensors Journal, 20(15):8757–8770.

Casale, P., Pujol, O., and Radeva, P. (2011). Human activity

recognition from accelerometer data using a wearable

device. In Iberian Conference on Pattern Recognition

and Image Analysis, pages 289–296. Springer.

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.

Du, X., Farrahi, K., and Niranjan, M. (2019). Trans-

fer learning across human activities using a cascade

neural network architecture. In Proceedings of the

23rd international symposium on wearable comput-

ers, pages 35–44.

Feng, Z., Mo, L., and Li, M. (2015). A random forest-based

ensemble method for activity recognition. In 2015

37th Annual International Conference of the IEEE En-

gineering in Medicine and Biology Society (EMBC),

pages 5074–5077. IEEE.

Fu, Z., He, X., Wang, E., Huo, J., Huang, J., and Wu,

D. (2021). Personalized human activity recogni-

tion based on integrated wearable sensor and transfer

learning. Sensors, 21(3):885.

Garofalo, G., Argones R

ua, E., Preuveneers, D., Joosen, W.,

et al. (2019). A systematic comparison of age and gen-

der prediction on imu sensor-based gait traces. Sen-

sors, 19(13):2945.

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.

Jain, A. and Kanhangad, V. (2016). Investigating gen-

der recognition in smartphones using accelerometer

and gyroscope sensor readings. In 2016 International

Conference on Computational Techniques in Infor-

mation and Communication Technologies (ICCTICT),

pages 597–602.

Markitantov, M. (2020). Transfer learning in speaker’s age

and gender recognition. In Karpov, A. and Potapova,

R., editors, Speech and Computer, pages 326–335,

Cham. Springer International Publishing.

Mehrang, S., Pietil

a, J., and Korhonen, I. (2018). An activ-

ity recognition framework deploying the random for-

est classiﬁer and a single optical heart rate monitor-

ing and triaxial accelerometer wrist-band. Sensors,

18(2):613.

Miraldo, D., Watanabe, R., and Duarte, M. (2020). An open

dataset of inertial, magnetic, foot-ground contact, and

electromyographic signals from wearable sensors dur-

ing walking.

Mostafa, A., Barghash, T. O., Assaf, A. A.-S., and Go-

maa, W. (2020). Multi-sensor gait analysis for gender

recognition.

Ngo, T. T., Ahad, M. A. R., Antar, A. D., Ahmed, M., Mura-

matsu, D., Makihara, Y., Yagi, Y., Inoue, S., Hossain,

T., and Hattori, Y. (2019). Ou-isir wearable sensor-

based gait challenge: Age and gender. In Proceedings

of the 12th IAPR International Conference on Biomet-

rics, ICB.

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.

Riaz, Q., V

ogele, A., Kr

uger, B., and Weber, A. (2015). One

small step for a man: Estimation of gender, age and

height from recordings of one step by a single inertial

sensor. Sensors, 15(12):31999–32019.

Rosli, N. A. I. M., Rahman, M. A. A., Balakrishnan, M.,

Komeda, T., Mazlan, S. A., and Zamzuri, H. (2017).

Improved gender recognition during stepping activity

for rehab application using the combinatorial fusion

approach of emg and hrv. Applied Sciences, 7(4):348.

Sun, M., Li, C., and Zha, H. (2017). Inferring private de-

mographics of new users in recommender systems. In

Proceedings of the 20th ACM International Confer-

ence on Modelling, Analysis and Simulation of Wire-

less and Mobile Systems, pages 237–244.

BioDeep: A Deep Learning System for IMU-based Human Biometrics Recognition

629