Athlete Identiﬁcation using Acceleration and Electrocardiographic

Measurements Recorded with a Wireless Body Sensor

Peter Christ

, Felix Werner

, Ulrich R

uckert

and J

org Mielebacher

Cognitronics and Sensor Systems Group, CITEC, Bielefeld University, Bielefeld, Germany

Mielebacher Informatiksysteme, Fahrenheitstr. 1, Bremen, Germany

Keywords:

Human Identiﬁcation, Accelerometer, Electrocardiograph (ECG), Wireless Body Sensor (WBS), Pattern

Recognition.

Abstract:

In this paper we propose a biometric method for identifying humans during walking and jogging. We use

acceleration and electrocardiographic measurements recorded with a wireless body sensor attached to a chest

strap. Our method does not require a particular acquisition setup. Information on the gait style and on the

physiology is combined to identify a human despite severe motion related artefacts in the electrocardiograph

and variations in the gait patterns. We propose to identify humans using features extracted in time and fre-

quency domain and a standard classiﬁer. With the collected data of 22 subjects on a treadmill at velocities

from 3 to 9 km/h we obtained an accuracy of 98.1 %. The sensitivity of the identiﬁcation ranged between 94.6

to 99.5 % for the different subjects and the speciﬁcity was higher than 99.7 %.

1 INTRODUCTION

The identiﬁcation of humans is important for various

applications such as surveillance systems, authoriza-

tion checks at doors or electronic devices (e.g. com-

puter, smartphone). A variety of biometric character-

istics have been investigated such as information from

ﬁngerprint, iris and retina, human face, voice, gait or

electrocardiograph.

Previous work has shown that discerning, repro-

ducible information on the human is found in the

ECG waveform, especially around the QRS com-

plex (Conover, 2002; Chan et al., 2008). Moreover,

biomechanical differences between the gait style of

humans have been investigated and used for identiﬁ-

cation within video and acceleration sensor based ap-

plications (Gafurov et al., 2006; Nixon et al., 2006).

We propose a biometric measure combining both

sources of information: characteristics in the electro-

cardiograph (ECG) waveform and the gait style. Un-

like other applications, our approach focuses on the

identiﬁcation of athletes during physical exercise us-

ing a compact wireless body sensor (WBS) which is

worn around the chest (see Figure 1). The WBS is

typically used to measure the heart-rate and the body

accelerations of athletes. Our identiﬁcation method

additionally utilizes the sensor measurements to iden-

tify the athlete, enabling an automatic annotation of

sensor data with the subject’s identity. Our goal is

to overcome the drawbacks of a manual annotation

of measurements for applications in sports medicine

and athlete training research. Furthermore, recogniz-

ing the subject allows to automatically load personal

settings on the WBS or the sport equipment for a cus-

tomized training. Our identiﬁcation method is in par-

ticular interesting for a WBS which is used with sev-

eral athletes of a mid-sized group.

Our identiﬁcation method uses features in time

and frequency domain to extract characteristics on the

subject which are used as input to a classiﬁer for iden-

tiﬁcation. By combining information from gait and

ECG we can successfully identify subjects despite of

artefacts in the ECG caused by a slipping of the ECG

electrodes and severe variations in the gait patterns

between walking and jogging.

Previous work in this ﬁeld focused on the identiﬁ-

cation of humans from either gait or ECG waveform

characteristics. Mainly ECGs were used which were

recorded at rest or with a clinical acquisition setup.

The gait based identiﬁcation was carried out for walk-

ing velocities.

Rong et. al (2007) proposed a method which

uses measurements recorded during walking with an

accelerometer located at the subject’s waist. The

method utilises a segmentation into gait cycles to ex-

tract gait patterns. Dynamic time warping is applied

Christ P., Werner F., Rückert U. and Mielebacher J..

Athlete Identiﬁcation using Acceleration and Electrocardiographic Measurements Recorded with a Wireless Body Sensor.

DOI: 10.5220/0004190300110019

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2013), pages 11-19

ISBN: 978-989-8565-36-5

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Our self-made wireless body sensor (WBS) and

its integration into a chest strap. The WBS can measure a

person’s electrocardiograph (ECG) and accelerations of the

body along three orthogonal axes.

to compensate natural changes in walking speed. The

actual gait segment is then compared with a reference

pattern of the subject and a 1-nearest neighbour clas-

siﬁer is used to recognize the subject. Ailisto et. al

(2005) evaluated an accelerometer based identiﬁca-

tion based on similarities between gait segments to

protect portable devices. M

antyj

arvi et. al (2005)

evaluated a gait based identiﬁcation for different

walking velocities using correlation coefﬁcients de-

rived from a template comparison, frequency coefﬁ-

cients and a histogram based comparison. Gafurov et.

al (2006) proposed two methods based on histogram

similarity and gait cycle length to distinguish acceler-

ation measurements recorded at the lower leg.

Several methods have been proposed to identify

a human based on ECG measurements. Biel et.

al (2001) used data from a standard 12-lead ECG

recorded during rest to identify subjects using mul-

tivariate analysis. Furthermore, the study showed that

identiﬁcation is possible with even one-lead ECGs.

Shen et. al (2002) also utilises data from one-lead

ECGs to distinguish subjects using a template match-

ing and a decision-based neural network. Chan et.

al (2008) identiﬁes subjects based on ECGs recorded

within a non-clinical acquisition setup where the sub-

jects were holding two electrodes on the pads of their

thumbs. For classiﬁcation, three qualitative measures

were used: percent residual difference, correlation

coefﬁcient, and a novel distance measure based on

wavelet transform.

This paper is organized as follows: Section 2 de-

scribes the identiﬁcation of a subject based on accel-

eration and ECG measurements. Information on pre-

processing, feature extraction and used classiﬁers is

given. Section 3 explains the conducted experiment

for data collection. Section 4 presents the experimen-

tal results of our identiﬁcation method. The results are

summarised and discussed in section 5 and a prospect

on our future work is given.

Figure 2: Vertical acceleration data of a subject walking and

jogging at velocities from 3 to 9 km/h. Each stride is rep-

resented by two consecutive peaks which correspond to the

heel strike (square) and the toe strike (triangle). These peaks

are marked for 9 km/h (red) and for 3 km/h (black). Veloc-

ity can be increased with either longer strides (increase in

signal amplitude) or a higher step frequency.

Figure 3: Alignment of 100 consecutive strides of four

subjects jogging at 9 km/h. The vertical acceleration sig-

nals were automatically segmented into strides and cross-

correlation was used to align the strides. The peaks related

to the heel strikes (square) and toe strikes (triangle) signiﬁ-

cantly differ in shape between the subjects.

2 IDENTIFICATION OF A

SUBJECT

This section describes the identiﬁcation of a subject

based on gait style and ECG waveform characteris-

tics. We describe the preprocessing of the signals, the

feature extraction and the classiﬁers used for identiﬁ-

cation.

2.1 Gait Analysis for Identiﬁcation

Previous work has shown that gait differs between hu-

mans and that the gait style is fairly stable for a sub-

ject (Bianchi et al., 1998; Nixon et al., 2006). Bianchi

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

Figure 4: Comparison of heartbeat segments of six subjects

(different colours). The DC-offset was removed and the

heartbeat segments were aligned using cross-correlation.

We use inter-subject variations in the ECG waveform to

identify subjects.

et. al (1998) stated that the variability across humans

depends on different kinematic strategies rather than

on biomechanical characteristics. Their study showed

that subjects are different in the ability of minimising

energy oscillations of their body segments for trans-

ferring mechanical energy.

In order to measure these inter-subject differences,

severe intra-subject variations in the gait patterns be-

tween walking and jogging have to be taken into ac-

count. The intra-subject variations are a result of

an adaptation of the gait to achieve different veloci-

ties. The velocity of a person is described by stride

length and stride frequency. According to Weyand

et. al (2000), longer strides are achieved by apply-

ing greater support forces to the ground which sig-

niﬁcantly increases the amplitude of the vertical ac-

celeration signal, whereas the step frequency changes

frequency components of the signal.

Samples of vertical acceleration data of one sub-

ject walking and jogging at different velocities on a

treadmill are shown in Figure 2. Strides are pre-

sented by two consecutive peaks corresponding to the

heel and toe strikes. Signiﬁcant changes in amplitude

and an almost doubling of the step frequency can be

observed between walking at 3 km/h and jogging at

9 km/h.

Despite of this intra-subject variability in the gait

patterns, we observed inter-subject variations in ac-

celeration signals recorded during walking and jog-

ging (Christ et al., 2011). In particular, heel and toe

strikes differ in the vertical acceleration signal’s shape

between subjects (see Figure 3). The peak accelera-

tion of the heel strikes varies between the four sub-

jects about 2 m/s

2.2 ECG Analysis for Identiﬁcation

Inter-subject variability is also found in the ECG’s

waveform. The variations depend on position, size

and anatomy of the heart, age, sex, relative body

weight, chest conﬁguration and various other fac-

tors (Green et al., 1985; Simon and Eswaran, 1997).

Figure 4 shows sample heartbeat segments from six

subjects recorded with our WBS. The ECG reﬂects

the electrical activity of the heart and consists of the

P wave followed by the QRS complex and the T

wave (Conover, 2002, chap. 2). Discerning informa-

tion on the subjects is found in the QRS complex, the

P and the T wave.

Chan et. al (2008) observed a high degree of re-

producibility of information extracted from the QRS

complex of a person through several sessions of

recording. Furthermore, a higher identiﬁcation accu-

racy was determined for the P wave than the T wave.

During physical exercise these characteristics can

be superposed by motion related artefacts. These arte-

facts are caused by a slipping of the ECG electrodes

and variations in the contact resistance during body

movements (Christ et al., 2010). Figure 7 shows dis-

turbances in the ECGs of two subjects recorded dur-

ing jogging on a treadmill.

2.3 Preprocessing of Acceleration and

ECG Signals

ECGs recorded with our WBSs showed hardware-

related differences in the DC-offset making an ECG

associable to a WBS. Furthermore, using textile ECG

electrodes, the skin contact resistance decreases over

time because of an increased transpiration which re-

sults in changes in the DC-offset. In order to avoid

classiﬁcation errors, we removed the DC-offset using

a 4th-order high-pass butterworth ﬁlter with a cutoff

frequency of f

= 0.67 Hz. Additionally, we applied a

low-pass ﬁlter with a cutoff frequency of f

= 40Hz

to remove noise in the ECG signal.

With a decrease in skin contact resistance after

a few minutes of exercise, we observed an increase

in the ECG signal’s amplitude which improved the

signal-to-noise ratio. We normalised the signal’s am-

plitude to assure that ECG segments are comparable.

The results of the ECG preprocessing are shown in

Figures 5 a and b.

For the frequency analysis of the acceleration

measurements, we approximated the dynamic accel-

erations by applying a 4th-order butterworth high-

pass ﬁlter with a cutoff frequency of f

= 0.1 Hz

to the magnitude of the acceleration vector a =

); a

denotes anteroposterior accelera-

AthleteIdentificationusingAccelerationandElectrocardiographicMeasurementsRecordedwithaWirelessBodySensor

(a) ADC output and preprocessed ECG of subject 1.

(b) ADC output and preprocessed ECG of subject 2.

the high-pass ﬁlter.

Figure 5: The 12 bit analog-to-digital converter (ADC) out-

put and the preprocessed ECG in comparison. An offset of

300 between the ADC output of the two different subjects

was removed by the preprocessing. In Figure 5 c the offset

due to the static acceleration of gravity and a sensor-related

zero-g-level offset are reduced after preprocessing.

tions, a

mediolateral accelerations and a

vertical

(up-down) accelerations. The high-pass ﬁlter reduced

the impact of the static acceleration due to gravity

and a sensor-related offset (zero-g level offset). The

results of this preprocessing step are shown in Fig-

ure 5 c.

2.4 Feature Extraction for Identiﬁcation

In order to access characteristics of a subject in the ac-

celeration and ECG measurements, we extracted fea-

tures in the time and the frequency domain.

The features were calculated within a sliding win-

dow with no overlap and length N. Each window

at time t consists of N measurements x(t : t +N −

1) = x(t),x(t+1), ...,x(t+N−1). We empirically de-

termined an appropriate window length of two sec-

onds (N = 300).

Time Domain Features. In the time-domain we

calculated the variance, amplitude, mean and root

mean square (RMS) along the three orthogonal axes

and a

of the windowed acceleration signals.

Figure 6: Visualisation of time domain features ex-

tracted from mediolateral accelerations a

of ten subjects

at 9 km/h. Clusters are observable for the different subjects.

In our feature selection we obtained a good identiﬁcation

performance based on the mean, the variance, the amplitude

and the root-mean-square (RMS) features (see Table 3).

Figure 7: Alignment of 20 heartbeat segments of four sub-

jects recorded during jogging on a treadmill. A correct

placement of the chest strap is important for an identiﬁca-

tion based on a similarity measure between heartbeat seg-

ments. Motion related artefacts and poor skin contact can

disturb the ECG-signal (see subjects 5 and 13).

The variance, mean and amplitude of a

are visu-

alised in Figure 6. Discriminative clusters can be ob-

served for the different subjects. From the ECG sig-

nal we calculated a feature measuring the closeness of

an unknown heartbeat segment to ﬁve reference pat-

terns stored for each subject. This step requires a seg-

mentation of the ECG signal into heartbeats. We used

a QRS detection based on the algorithm of (Afonso

et al., 1999) in its implementation of Schloegl in the

BioSig toolbox (Vidaurre et al., 2011). The ﬁve refer-

ence heartbeat segments were chosen randomly from

the ECG data of each subject. However, we assured

that only heartbeat segments without severe distur-

bances were chosen. For identiﬁcation, an unknown

segment x was aligned to each reference segment y

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

Figure 8: The FFT amplitude spectra of the ECG sig-

nals of three subjects during walking (3 km/h) and jog-

ging (9 km/h). The amplitude spectra show differences be-

tween the subjects but also vary with the velocity.

using cross-correlation:

(m) =

N−m−1

∑

j=0

y( j+m)x( j) (1)

where N is the length of a segment and m the offset

with m = 0,1,...,2N−1. We calculated the Pearson’s

correlation coefﬁcient as a measure of similarity be-

tween the two segments. The Pearson’s correlation

coefﬁcient is deﬁned as the covariance (cov) of the

two segments divided by the product of their standard

deviation σ:

r(x, y) =

cov(x,y)

(2)

Figure 7 shows the alignment of 20 heartbeat seg-

ments of four subjects. The QRS-detection and

the alignment are sensitive to motion-related arte-

facts (see subjects 5 and 13).

For heartbeat segments without major distur-

bances the alignment centred the segments around the

QRS complex. The discerning information in this re-

gion of the ECG is fairly stable in relation to morphol-

ogy changes in the ECG waveform during effort.

Frequency Domain Features. In the frequency do-

main we use the discrete Fourier transform (DFT) to

extract frequency components of each window. The

DFT is deﬁned as:

X(k) =

t+K−1

∑

j=t

x( j)e

−i2πk

, k =0,...,K−1 (3)

where K is the number of outputs X(k). We used a

512-point fast Fourier transform (FFT) algorithm to

compute the DFT efﬁciently for our windows of the

length N = 300. Therefore, each window x(t :t +N−

1) was padded with trailing zeros to the length of K =

512. Before calculating the FFT, a Hamming window

function was applied to each window to reduce the

spectral leakage.

Figure 8 shows the FFT amplitude spectra of

ECGs of three subjects recorded during walk-

ing (3 km/h) and jogging (9 km/h). Despite of velocity

related variations in the amplitude spectra, differences

can be observed between the three subjects.

We calculated additional frequency domain fea-

tures from the amplitude spectrum (FFT features): the

variance, the mean, the Fourier coefﬁcient with the

highest amplitude and the Shannon entropy SE:

SE = −

K−1

∑

k=0

X(k)

log

(

X(k)

) (4)

where X(k) is the output of the DFT of length K.

2.5 Methods for Classiﬁcation

We used a standard classiﬁer to identify the subject

based on the extracted features. The identiﬁcation

performance was determined by evaluating three dif-

ferent classiﬁers: artiﬁcial neural network (ANN),

support vector machine (SVM), and random for-

est (RF).

Artiﬁcial Neural Network (ANN). We used a

feed-forward ANN with 25 neurons with tangent sig-

moid activation functions in one hidden layer to asso-

ciate the extracted features with the subjects’ identi-

ties. The ANN was trained using back-propagation

which is a supervised learning method (Han and

Kamber, 2006). During training the prediction of

the network is compared to the known target value

(subject’s identity) and the weights are modiﬁed to

minimize the mean square error. These errors propa-

gate backwards from the output layer to the hidden

layer (Han and Kamber, 2006). The network was

trained using the scaled conjugate gradient algorithm

described in (Møller, 1993). The weights and bias

values of the neurons were updated using a gradient

descent with momentum.

Support Vector Machine (SVM). We used a ν-

SVM (Sch

olkopf et al., 2000) with a sigmoid kernel in

its implementation in the LIBSVM

(Chang and Lin,

2011). SVMs are fundamentally a two-class classi-

ﬁer. Various methods have been proposed how to use

SVMs for multi-class problems (Bishop, 2006, chap.

LIBSVM: library for support vector machines.

AthleteIdentificationusingAccelerationandElectrocardiographicMeasurementsRecordedwithaWirelessBodySensor

Table 1: Characteristics of the 22 subjects (15 men, 7

women) who participated in the data collection.

Characteristic Mean ± SD Range

Age (yr) 26.6 ± 4.0 18-33

Height (cm) 179.8 ± 9.6 160-198

Weight (kg) 76.7 ± 11.1 58-108

7). We used a one-against-one method which con-

structs n(n − 1)/2 classiﬁers where n is the number

of classes to distinguish. Each classiﬁer is trained on

tuples from two classes. A voting strategy is then ap-

plied to determine the winning class (Hsu and Lin,

2002).

Random Forest (RF). A random forest is a clas-

siﬁer consisting of a combination of tree predictors.

The growth of each tree is governed by independently

and identically distributed random vectors (Breiman,

2001). Each tree votes for one class and the class

which occurs most frequently is the output of the clas-

siﬁer. RF classiﬁers are fast in the training phase and

the training time is linear to the number of trees used.

The testing of an unknown tuple is performed on each

tree independently and is therefore parallelisable. We

used a RF consisting of 100 trees, with each tree being

constructed of ten randomly chosen features.

3 SUBJECTS AND DATA

COLLECTION

Twenty-two healthy subjects (15 men, 7 women) vol-

unteered to participate in the study. The subjects

were informed verbally and in writing in advance

and signed an informed consent document. With a

questionnaire we additionally obtained information

on age, height and weight (see Table 1).

The data was collected using the treadmills in the

gymnasium of our university. Velocities between 3

to 9 km/h were chosen to cover slow, normal, and

fast walking as well as jogging. The treadmill was

set to no incline and the velocity was manually in-

creased by 2 km/h every two minutes. This procedure

was repeated twice for each subject in order to collect

enough data.

The accelerations of the upper body and the

ECG were recorded with a self-made WBS (see Fig-

ure 1). The WBS measures accelerations within a

range of ±6 m/s

along three orthogonally oriented

axes using a commercial off-the-shelf accelerometer

(ST LIS3LV02DL). The ECG is digitized using the

analog-to-digital converter of a TI MSP430 micro-

controller. Body accelerations and ECG were mea-

sured with a 150 Hz sampling rate and a 12 bit resolu-

tion (range 0 to 4095). The measurements were sent

wirelessly to a nearby receiver for recording.

The subjects were given an explanation as how to

place the chest strap with the WBS tightly around the

chest. However, we didn’t verify the correct place-

ment of the WBS to assure real world conditions. Fur-

thermore, no instructions were given on how to per-

form the exercise.

4 RESULTS

This section describes the evaluation of the athlete

identiﬁcation on data collected from 22 subjects dur-

ing walking and jogging on a treadmill.

Evaluation Methods. All features were calculated

on windows of acceleration and ECG measurements

of two seconds. No overlap of the windows was cho-

sen to ensure fully discriminative training and test-

ing data. We concatenated features of two consecu-

tive windows to have samples of four seconds of data

to identify the subject. Our dataset consists of 4048

samples (184 samples per subject). The identiﬁcation

performance was determined using a ten-fold cross-

validation. For evaluation, we used three statistical

measures: sensitivity, speciﬁcity and accuracy. In or-

der to calculate the statistics we obtained the number

of true positive samples TP

, true negative samples

, false positive samples FP

, and false negative

samples FN

from the classiﬁer’s output. For a class i

the sensitivity R

is deﬁned as:

+ FN

∗ 100 (5)

The sensitivity (also referred to as recall) measures

the percentage of correctly classiﬁed positive samples

in relation to all positive samples. For negative sam-

ples the speciﬁcity S

is deﬁned as:

+ FP

∗ 100 (6)

We calculated the overall sensitivity R and the overall

speciﬁcity S as a class-based weighted average. For

our multi-class problem we refer to the overall sensi-

tivity as the accuracy of the classiﬁer:

ACC = R =

∑

i=1

(7)

where n denotes the number of classes and p

the

probability of the occurrence of the class in the test

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

Table 2: Accuracy (ACC) and overall speciﬁcity (S) results

determined with different classiﬁers on a feature space com-

bining acceleration and ECG features (combination C8, see

Table 3). The highest performance was obtained with the

RF classiﬁer.

Classiﬁer ACC S

ANN 94.2 % 99.8 %

SVM 90.4 % 99.5 %

RF 98.1 % 99.9 %

Figure 9: Class-speciﬁc sensitivity (stars) and speciﬁcity

(squares) results of the identiﬁcation of the 22 subjects (RF

classiﬁer, feature combination C8). The sensitivity varied

between 94.6 to 99.5 %. The speciﬁcity was over 99.7 %

for all subjects.

data. In our test data the samples of the 22 classes are

equally distributed (p

= 1/22, ∀i). The overall speci-

ﬁcity S is calculated accordingly. The optimum of the

statistical measures is 100 %.

Results of the Athlete Identiﬁcation. We deter-

mined the identiﬁcation performance for three stan-

dard classiﬁers: ANN, SVM and RF. The classiﬁers

and their parametrization are described in section 2.5.

We achieved up to 98.1 % accuracy (see Table 2) with

the RF classiﬁer using a feature space combining ac-

celeration and ECG based features. The lowest accu-

racy of 90.4 % was obtained with the SVM. For all

three classiﬁers, we obtained an overall speciﬁcity S

of more than 99 %.

The class-speciﬁc sensitivity (see Equation 5) of

the identiﬁcation varied between 94.6 to 99.5 % for

the different subjects (RF classiﬁer, see Figure 9). We

observed only low deviations in the identiﬁcation’s

speciﬁcity between the 22 subjects. A class-speciﬁc

speciﬁcity (see Equation 6) of more than 99.7 % was

achieved for all subjects.

We performed a feature selection using the ANN

classiﬁer to determine the impact of the different fea-

tures and to identify combinations C with a high clas-

siﬁcation performance (see Table 3). We obtained a

similar identiﬁcation accuracy based on acceleration

(86.6 %, C6) and ECG (84.8 %, C4) measurements. In

combination, the accuracy improved to 94.2 % (C8).

The ECG contained more information on the sub-

ject in the frequency domain than the acceleration

Table 3: Identiﬁcation accuracy of different feature combi-

nations C. We obtained a similar accuracy with accelera-

tion and ECG based features (see C4, C6). Combining both

improved the accuracy (see C8). The feature selection was

performed using the ANN classiﬁer. * denotes the use of the

average over the ﬁve correlation coefﬁcients per subject.

C Acceleration feat. ECG feat. ACC

Time FFT FFT FFT FFT Corr.

dom. coef. feat. coef. feat. coef.

C1 - - - x x - 72.4 %

C2 - - - - - x 80.3 %

C3 x - - - - - 83.3 %

C4 - - - x x x 84.8 %

C5 x - - - x - 86.5 %

C6 x x x - - - 86.6 %

C7 x - - - - x 93.6 %

C8 x x x x x x* 94.2 %

measurements (12.3 % higher accuracy). Frequencies

of up to 10 Hz contained the most discriminant in-

formation of the acceleration measurements. A re-

duction of the frequency band from 40 to 10 Hz re-

duced the identiﬁcation accuracy by only 3.8 %. For

the ECG measurements, a reduction from 40 Hz to

15 Hz resulted in a 8.8 % lower accuracy. Overall, we

obtained an accuracy of 72.4 % (C1) for features ex-

tracted from the ECG in the frequency domain.

We found that correlation coefﬁcients describing

the similarity between heartbeat segments provide

useful insights to identify subjects (80.3 % accuracy,

C2). To reduce the dimensionality of the feature

space, we averaged the correlation coefﬁcients corre-

sponding to the ﬁve reference segments per subject.

This averaging resulted in a 7.7 % lower accuracy.

However, in combination with other features this dif-

ference was negligible (0.4 % for C8).

The time domain features calculated from the ac-

celeration signals showed a good accuracy (83.3 %,

C3). Additional information on the gait in the fre-

quency domain improved the identiﬁcation accuracy

to 86.6 % (C6).

By combining the time domain features of the ac-

celeration data with the correlation coefﬁcients de-

rived from the ECG, we achieved a high accuracy

of 93.6 % (C7), which is only 0.6 % less than using

the full feature set (C8).

For the time domain features extracted from the

acceleration signals, we analysed the impact of the

different acceleration axes on the subject’s identiﬁ-

cation accuracy. The highest accuracy was obtained

for the anteroposterior accelerations (a

). The medi-

olateral accelerations (a

) showed a 4.4 % and the

vertical accelerations (a

) a 16.2 % lower accuracy.

We additionally evaluated our approach using a

hold-out validation for which the data set was split

AthleteIdentificationusingAccelerationandElectrocardiographicMeasurementsRecordedwithaWirelessBodySensor

Table 4: Identiﬁcation accuracy (ACC) and equal error ratio

(ERR) of the RF classiﬁer obtained with a ten-fold cross-

validation and a hold-out validation (percentage split: 66 %

training, 34 % testing) on feature combination C8.

Validation ACC ERR

Ten-fold cross-validation 98.1 % 1.1 %

Hold-out validation 97.2 % 3.4 %

Table 5: Equal error ratio (ERR) of other gait based iden-

tiﬁcation methods. N denotes the number of subjects who

participated in the experiments.

Velocities N ERR

antyj

arvi et. al (2005) slow, normal and

fast walking

36 7 %

Ailisto et. al (2005) normal walking 36 6.4 %

Gafurov et. al (2006) normal walking 21 5 %

Rong et. al (2007) normal walking 21 5.6 %

in 66 % training data and 34 % testing data. A

hold-out validation avoids temporal proximity be-

tween training and testing data and allows therefore a

more accurate estimation of the generalization perfor-

mance. We noted only a slight decrease in accuracy

by 0.9 % for the RF classiﬁer (see Table 4).

To estimate the impact of the number of subjects

in the data set on the identiﬁcation performance, we

repeated our test with eleven out of the twenty-two

subjects. For a group of eleven subjects which were

randomly chosen, the overall accuracy could be im-

proved to 99.3 % (RF classiﬁer).

In order to compare our results with existing work,

we additionally calculated the equal error rate (ERR)

of the RF classiﬁer on feature combination C8. The

ERR is the rate at which both accept and reject errors

are equal. For our dataset containing ECG and gait

characteristics, we obtained an ERR of 1.1 % for the

ten-fold cross-validation and 3.4 % for the hold-out

validation. Compared to other approaches which are

based on only gait characteristics our achieved ERR

is lower (see Table 5). For a comparison of our ap-

proach with an ECG based identiﬁcation we have cho-

sen the method of Chan et. al (2008) because the re-

sults are also based on data from non-clinical ECGs.

With 98.1 % our accuracy is higher than Chan et. al

(2008) results (89 %). However, with an identiﬁcation

on ECG characteristics only, we obtained a lower ac-

curacy (84.8 %, C4). Overall, our high performance

is achieved by combining ECG and gait characteris-

tics. We believe that motion related artefacts in the

ECG, and a high variability in the gait patterns be-

tween changing from slow walking to jogging, reduce

the identiﬁcation performance when we use only one

source of information.

5 DISCUSSION AND

CONCLUSIONS

This paper is concerned with the identiﬁcation of hu-

mans during walking and jogging using a single wire-

less body sensor module attached to a chest strap. Our

approach focuses on recognising a human using a bio-

metric measure based on the characteristics in the gait

style and the ECG of the human and is hence indepen-

dent of the used hardware. Thus, our system over-

comes the drawbacks of an identiﬁcation based on

the WBS’s serial number or an radio-frequency based

identiﬁcation (RFID) which recognises the hardware

but not the subject itself.

We have collected data from 22 subjects on a

treadmill at velocities from 3 to 9 km/h using a WBS

attached to a chest strap. To assure real world condi-

tions, no advice was given on how to perform the ex-

ercise and the correct placement of the chest strap was

not veriﬁed. Despite severe variations in the gait pat-

terns and motion-related artefacts in the ECG, which

occur due to real world conditions and physical exer-

cise, our method achieves up to 98 % accuracy.

We obtained a good identiﬁcation accuracy for

time domain features extracted from the acceleration

signals. By using simple and low-dimensional fea-

tures on the acceleration signal our method can poten-

tially be implemented on computationally constrained

platforms, such as a microcontroller on a WBS.

Our identiﬁcation method can presumably not be

extended to an unlimited number of subjects. The in-

dividual characteristics in the subject’s ECG and gait

patterns are extremely difﬁcult to capture and may

change over time because of an adaptation to physi-

cal exercise. However, we believe our method is well

suited to provide an automatic annotation of sensor

measurements from several WBSs with the subject’s

identity for use in sports medicine and athletic train-

ing research. Moreover, our method helps to cus-

tomize a training session by loading personal settings

of the recognized athlete on the WBS or other sport

equipment.

Our future work includes the evaluation of the

identiﬁcation method within team sports. In partic-

ular, we want to recognize handball players in order

to support a real-time vision-based tracking of these

players.

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

ACKNOWLEDGEMENTS

This research was supported by the DFG CoE 277:

Cognitive Interaction Technology (CITEC)

REFERENCES

Afonso, V. X., Tompkins, W. J., Nguyen, T. Q., and Luo, S.

(1999). ECG beat detection using ﬁlter banks. Trans-

actions on Biomedical Engineering, 46(2):192–202.

Ailisto, H. J., Lindholm, M., Mantyjarvi, J., Vildjiounaite,

E., and Makela, S. M. (2005). Identifying people

from gait pattern with accelerometers. In Society

of Photo-Optical Instrumentation Engineers, volume

5779, pages 7–14.

Bianchi, L., Angelini, D., and Lacquaniti, F. (1998). In-

dividual characteristics of human walking mechan-

ics. Pﬂ

ugers Archiv European Journal of Physiology,

436:343–356.

Biel, L., Pettersson, O., Philipson, L., and Wide, P. (2001).

ECG analysis: a new approach in human identiﬁca-

tion. IEEE Transactions on Instrumentation and Mea-

surement, 50(3):808–812.

Bishop, C. M. (2006). Pattern Recognition and Ma-

chine Learning (Information Science and Statistics).

Springer-Verlag New York, Inc., Secaucus, NJ, USA.

Breiman, L. (2001). Random forests. Machine learning,

45(1):5–32.

Chan, A. D. C., Hamdy, M. M., Badre, A., and Badee, V.

(2008). Wavelet distance measure for person identiﬁ-

cation using electrocardiograms. IEEE Transactions

on Instrumentation and Measurement, 57(2):248–

253.

Chang, C.-C. and Lin, C.-J. (2011). LIBSVM: A library

for support vector machines. ACM Transactions on

Intelligent System Technology, 2:27:1–27:27.

Christ, P., Mielebacher, J., Haag, M., and R

uckert, U.

(2010). Detection of body movement and measure-

ment of physiological stress with a mobile chest mod-

ule in obesity prevention. In Proc. of the 10th Aus-

tralasian Conf. on Mathematics and Computers in

Sport, pages 67–74.

Christ, P., Werner, F., R

uckert, U., and Mielebacher, J.

(2011). An approach for determining linear veloc-

ities of athletes from acceleration measurements us-

ing a neural network. In Proc. of the 6th IASTED Int.

Conf. on Biomechanics, pages 105–112. ACTA Press.

Conover, M. B. (2002). Understanding electrocardiogra-

phy. Mosby.

Gafurov, D., Helkala, K., and Søndrol, T. (2006). Biometric

gait authentication using accelerometer sensor. Jour-

nal of Computers, 1(7):51–59.

Green, L. S., Lux, R. L., Haws, C. W., Williams, R. R.,

Hunt, S. C., and Burgess, M. J. (1985). Effects of

age, sex, and body habitus on QRS and ST-T po-

tential maps of 1100 normal subjects. Circulation,

71(2):244–253.

Han, J. and Kamber, M. (2006). Data mining: concepts

and techniques. The Morgan Kaufmann series in data

management systems. Elsevier.

Hsu, C. W. and Lin, C. J. (2002). A comparison of methods

for multiclass support vector machines. IEEE Trans-

actions on Neural Networks, 13(2):415–425.

antyj

arvi, J., Lindholm, M., Vildjiounaite, E., M

akel

a, S.-

M., and Ailisto, H. A. (2005). Identifying users of

portable devices from gait pattern with accelerome-

ters. In IEEE Int. Conf. on Acoustics, Speech, and

Signal Processing, volume 2, pages ii/973–ii/976.

Møller, M. F. (1993). A scaled conjugate gradient algo-

rithm for fast supervised learning. Neural Networks,

6(4):525–533.

Nixon, M. S., Tan, T., and Chellappa, R. (2006). Hu-

man identiﬁcation based on gait, volume 4. Springer-

Verlag New York, Inc.

Rong, L., Jianzhong, Z., Ming, L., and Xiangfeng, H.

(2007). A wearable acceleration sensor system for gait

recognition. In 2nd IEEE Conf. on Industrial Elec-

tronics and Applications, pages 2654–2659.

Sch

olkopf, B., Smola, A. J., Williamson, R. C., and Bartlett,

P. L. (2000). New support vector algorithms. Neural

computation, 12(5):1207–1245.

Shen, T. W., Tompkins, W. J., and Hu, Y. H. (2002). One-

lead ECG for identity veriﬁcation. In Proc. of the 2nd

IEEE Int. Joint Conf. on Engineering in Medicine and

Biology Society, volume 1, pages 62–63. IEEE.

Simon, B. P. and Eswaran, C. (1997). An ECG classiﬁer de-

signed using modiﬁed decision based neural networks.

Computers and Biomedical Research, 30(4):257–272.

Vidaurre, C., Sander, T. H., and Schl

ogl, A. (2011). BioSig:

The free and open source software library for biomed-

ical signal processing. Computational Intelligence

and Neuroscience, 2011:12.

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

S. (2000). Faster top running speeds are achieved with

greater ground forces not more rapid leg movements.

Journal of Applied Physiology, 89(5):1991–1999.

AthleteIdentificationusingAccelerationandElectrocardiographicMeasurementsRecordedwithaWirelessBodySensor