Feature-based Analysis of Gait Signals for Biometric Recognition

Automatic Extraction and Selection of Features from Accelerometer Signals

Maria De Marsico, Eduard Gabriel Fartade and Alessio Mecca

Department of Computer Science, Sapienza University of Rome, Via Salaria 113, Rome, Italy

Keywords:

Biometric Authentication, Gait Recognition, Automatic Feature Extraction.

Abstract:

Gait recognition has been traditionally tackled by computer vision techniques. As a matter of fact, this is a still

very active research ﬁeld. More recently, the spreading use of smart mobile devices with embedded sensors

has also spurred the interest of the research community for alternative methods based on the gait dynamics

captured by those sensors. In particular, signals from the accelerometer seem to be the most suited for rec-

ognizing the identity of the subject carrying the mobile device. Different approaches have been investigated

to achieve a sufﬁcient recognition ability. This paper proposes an automatic extraction of the most relevant

features computed from the three raw accelerometer signals (one for each axis). It also presents the results

of comparing this approach with a plain Dynamic Time Warping (DTW) matching. The latter is computa-

tionally more demanding, and this is to take into account when considering the resources of a mobile device.

Moreover, though being a kind of basic approach, it is still used in literature due to the possibility to easily

implement it even directly on mobile platforms, which are the new frontier of biometric recognition.

1 INTRODUCTION

Biometric traits are traditionally classiﬁed into phys-

ical and behavioral, and further denoted as strong

and soft, mainly according to uniqueness and perma-

nence. Strong traits mainly belong to the physical

group, and the soft ones in this group generally iden-

tify classes of users instead of individuals (e.g., hair

color, height, face shape, etc.). Almost all behavioral

traits rather belong to the soft category, due to lack

of sufﬁcient permanence over the long period. This

is because they can be affected by mood and speed of

action execution, and in some cases by temporary im-

pairment of the body part involved. Even if soft traits

are not accurate and permanent as the strong ones,

their analysis can be fused to enforce recognition ac-

curacy. Behavioral traits have also the advantage to

be more difﬁcult to forge and replicate.

Gait recognition can be considered to belong to

the behavioural biometrics, even if, especially in re-

lation with computer vision-based approaches, it also

presents some physical/visual characteristics. While

this kind of approaches has been the ﬁrst one adopted

for recognition, the problem has recently gained new

interest thanks to the new researches based on new

wearable sensors or wider availability of existing

ones. This paper focuses on Wearable Sensors-based

techniques, exploiting sensors built in modern smart-

phones. In particular, the techniques proposed in this

paper carry out recognition by the signals captured by

the embedded 3-axes accelerometers.

Like other traits, gait recognition suffers from

both intra-personal variations, either intrinsic or ex-

trinsic to the walking person, and inter-personal sim-

ilarities, possibly causing a subject to be confused

with another. Variations of the walking pattern from

the same individual mainly depend on speed, ground

slope, kind of worn shoes (e.g., heels for women

shoes, or heavy boats), and also on some temporary

(if not permanent) illness, such as contusions or other

problems affecting legs, articulations or feet. In addi-

tion to those factors, image-based techniques applied

to video sequences, can be further affected by vary-

ing illumination, occlusion, pose, perspective with re-

spect to the camera, and large clothes. Finally, a com-

mon problem, though producing different effects ac-

cording to the adopted sensors, is the presence of car-

ried objects, that modify the silhouette, generally ex-

ploited by computer vision approaches, as well as the

walking dynamics, especially if heavy. A further dis-

advantage of computer vision-based approaches is the

impossibility to carry out veriﬁcation of the walker di-

rectly on a personal device, since videos are necessar-

ily captured by an external device.

630

Marsico, M., Fartade, E. and Mecca, A.

Feature-based Analysis of Gait Signals for Biometric Recognition - Automatic Extraction and Selection of Features from Accelerometer Signals.

DOI: 10.5220/0006719106300637

In Proceedings of the 7th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2018), pages 630-637

ISBN: 978-989-758-276-9

Notwithstanding the above limitations, it is quite

difﬁcult to copy or forge the gait pattern produced

by someone else. In addition, while gait recognition

can operate at a distance (even 10 meters or more)

with computer vision applications, there is not even

this limitation with wearable sensors, since they are

directly attached to the user body. Moreover, while

video capture can mingle silhouettes of different sub-

jects in the same frames, depending on the point of

view and relative positions of these subjects, signals

captured by wearable sensors are by deﬁnition iso-

lated and independent from each other. Last but not

least, in both cases the user is not asked to do any

speciﬁc action but walk, and gait recognition can be

effectively combined with other ”strong” biometrics

to both enforce recognition accuracy and as an anti-

spooﬁng support.

The aim of this paper is to report the results of

investigations regarding the possibility to reduce the

computational time for matching with respect to pure

Dynamic Time Warping (DTW) that is still quite used

in literature. In particular, a feature extraction, anal-

ysis and selection procedure is devised in order to

achieve a faster yet still accurate recognition. The

added value of this approach is the possibility to pro-

cess the gait signal directly on a low-medium level

smartphone without need of sending data to an ex-

ternal server. Of course, this holds for the veriﬁ-

cation of the owner identity (1:1 identity matching),

which is implicitly assumed to be the person carry-

ing the device. As for identiﬁcation (1:N identity

matching), privacy as well as security considerations

would not allow anyway maintaining templates from

other users (gallery) in one’s own device and process-

ing them locally. Experimental setup exploits one of

the largest available wearable sensor-based datasets,

in order to provide a common benchmark for com-

parison with other works. Unfortunately, at present

this is not always possible due to in-house and often

private datasets used in several papers.

The paper continues as follows. Section 2 pro-

vides a summary of present research lines in this ﬁeld.

Section 3 describes the paper proposal, entailing the

process of signal capture followed by feature extrac-

tion, analysis and selection. Section 4 presents the

results from experiments carried out, and some com-

parison with raw signal matching. Finally, Section 5

draws some conclusions and sketches future work.

2 RELATED WORK

It is worth anticipating that most works in litera-

ture, addressing wearable sensor-based gait recogni-

tion, locate the acquisition device in different body

locations, and use different (often in-house) datasets

with a different number of subjects, so that a fair

comparison is not always feasible. Moreover, most

works only consider veriﬁcation modality (1:1 iden-

tity matching with either implicit or explicit identity

claim), a few ones also consider closed set identiﬁ-

cation modality (1:N identity matching with no iden-

tity claim but the assumption that all probes belong to

enrolled users), and almost none considers open set

identiﬁcation modality (1:N identity matching with

no identity claim and reject option). This is related

to the main intended use of this kind of biometrics,

devised to authenticate the owner of a mobile device

against a stored template. However, it is worth con-

sidering that the possibility to automatically trigger

signal capture by Bluetooth devices, and to transmit

the captured signal to a remote server, allows hypoth-

esizing a wider use of this biometrics for access con-

trol to restricted areas by user identiﬁcation.

As for wearable sensor-based recognition, it is

possible to identify two main categories of ap-

proaches.

The attempts in the ﬁrst group generally try to

match the template/templates of the users by sig-

nal matching techniques, such as Euclidean Distance

(ED), Dynamic Time Warping (DTW), or Histogram

Similarity (HS). These kinds of techniques, if applied

to match the entire signals, generally provide quite

good results. However, they drastically lose accu-

racy when the templates signiﬁcantly differ in terms

of number of samples. For this reason, some works in

this group use the adopted matching strategies after

dividing the signals into steps/cycles. This generally

entails a segmentation procedure that tries to automat-

ically identify the start and stop of a step/cycle in the

walk signal, where step is intended as a single foot

step, while cycle is intended as a pair of them (left-

right or right-left). In these cases, it is worth noticing

that the quality of the results of this operation have

an high impact on the recognition accuracy, and de-

veloping a good segmentation algorithm is a crucial

point.

In (Gafurov et al., 2010), the authors show the im-

pact of different kinds of shoes in the recognition,

using a in-house dataset of 30 persons. The pro-

posed method uses a ﬁxed threshold to identify cycles

and then normalizes their sample length. Recogni-

tion entails measuring ED between all pairs of cycles,

composed of one cycle extracted from a gallery sam-

ple and one cycle from the probe to match; the best

achieved distance has to meet the ﬁxed threshold for

the acceptance. Results are in terms of Equal Error

Rate (EER), the value where False Acceptance Rate

Feature-based Analysis of Gait Signals for Biometric Recognition - Automatic Extraction and Selection of Features from Accelerometer

Signals

631

is equal to False Rejection Rate, therefore the lower,

the better. Values range from 1.6% with the lighter

shoes to 5% with the heavier ones.

The work presented in (Gafurov et al., 2006) pro-

poses two different approaches for gait recognition,

tested on a in-house dataset of 21 users with only one

single walk each, further divided into two parts. This

causes a lack of most intraclass variation factors, and

therefore the dataset hardly represents a realistic sce-

nario. In real life settings, even the position of the

sensors cannot be completely controlled and is not

identical, and this fact in itself can produce variations

in the captured signal. For both proposed methods,

the 3-dimensional raw signals from the accelerometer

are combined into a single 1-dimensional vector using

the following ad hoc formula: v

= arcsin(

√

)

where i represents the index of the sample within the

signal. The ﬁrst matching strategy exploits HS. To

this aim, the obtained values are stored in a histogram

representing the derived biometric template. Match-

ing between the obtained histograms achieves a 5% of

EER. The second attempt uses the 1D vector for cycle

comparison and achieves a 9% of EER.

In (De Marsico and Mecca, 2016), the authors

present a novel step segmentation procedure and show

the performance of ﬁve different algorithms based on

DTW, one dealing with the entire signal and the oth-

ers using different strategies to match the detected

steps. In this case no ﬁxed threshold is used for seg-

mentation. The algorithm exploiting the entire signal

achieves 92.8% of Recognition Rate (RR) in closed

set identiﬁcation modality and 9.26% of EER in ver-

iﬁcation modality; the algorithms exploiting the step

segmentation procedure achieve up to 82.7% of RR

and up to 10.3% of EER. It is to underline that results

in this latter work are obtained over a very large pub-

lic dataset including 175 subjects with 12 walks each

(Zhang et al., 2015).

The approaches in the second group, instead, try

to exploit machine learning algorithms/classiﬁers in

order to get the correct match between templates.

These proposals generally work only in veriﬁcation

modality, training a classiﬁer for each subject. A com-

mon pre-processing step is to fragment the signal in

chunks with a ﬁxed length (in terms of either time or

number of samples, with or without overlap) in order

to extract more data for the training phase. Of course,

these approaches do not require any step/cycle seg-

mentation procedure. After the training of the clas-

siﬁer, that can occur on a more powerful device, the

trained system can be executed directly on a smart-

phone due to the low computational cost of the sin-

gle recognition operation. In fact, several works in

this category are generally designed to be executed di-

rectly on the mobile device in order to unlock it only

for its owner, as an alternative to pins or passwords.

Two solutions in this category are presented in

(Nickel et al., 2011b) and in (Nickel et al., 2011a).

Both works exploit the same dataset, collected by a

Google G1 phone. Such dataset contains walk signals

from 48 subjects, 4 walks each. In the ﬁrst work, sig-

nals are re-sampled to 200Hz. Afterwards, they are

divided into fragments of 3 seconds with no overlap.

Fragments are then grouped into two sets, one for the

training and the other one for testing. The recognition

is carried out by the HVITE tool and each subject is

used one time as genuine and forty-seven times as im-

postor. This strategy reports an EER of about 10%.

In the second work, the walking signals are inter-

polated to 100Hz and then are fragmented into chunks

of 7 seconds with an overlap of 50%. Each fragment

is used as feature vector, adding some extra statistical

parameters, and the Mel and the Bark frequency cep-

stral coefﬁcients. Training and recognition are car-

ried out exploiting the SVM classiﬁer. This approach

achieves a 5.9% of False Match Rate (FMR) with

6.3% of False Non Match Rate (FNMR).

In (Nickel et al., 2012), the authors try to exploit

the k-NN algorithm for recognition. Walking data are

collected during two sessions. Each of the 36 users

is asked to walk 12 times at normal pace, 16 times

at fast pace and again other 12 times at normal pace

on a ﬂat hallway. Each such group of walk signals is

captured by a single recording operation. Single sig-

nals are divided using an automatic procedure accord-

ing to some stop periods decided in advance, and the

result is eventually manually corrected. All signals

are then interpolated at 127Hz (this value is empiri-

cally chosen). After this preprocessing, the interpo-

lated signals are fragmented. This work exploits three

different fragment sizes, namely 3s, 5s, and 7.5s. In

all three cases, the fragments have an overlap of 50%.

The feature vectors are then created using some sta-

tistical parameters and by Mel and Berk coefﬁcients.

The recognition exploits the implementation of k-NN

algorithm included in the WEKA library. The work

reports a FMR of about 4% with a very high FNMR

of about 22-23%, resulting in a Half-Total Error Rate

(HTER - the average of the two) of about 13%. In or-

der to improve performances, the authors try a voting

approach using different fragments of the same sub-

ject; this signiﬁcantly reduces the FNMR while in-

creasing minimally the FMR, so achieving an HTER

of about 8.5%.

The proposal in (Zhang et al., 2015) presents a

large dataset of 175 subjects with 12 walks per per-

son (the same used as benchmark by our proposal)

and tests the extraction and the use of signature points

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

632

(SP). SPs are described as the ﬁducial positions within

gait signals that should be both stable for the same

person and distinctive for different persons. The au-

thors propose to sparsely represent the SPs and then

to create clusters, labelling them to make up a dictio-

nary in a linear combination, in order to have a subject

for each cluster. Recognition is described as a con-

ditional probability problem solved by a sparse-code

classiﬁer. The result reaches an up to 95.8% of RR in

closed set modality, and up to 2.2% of EER in veriﬁ-

cation modality. In this case, it is worth considering

that the authors use 5 accelerometers of the same kind

positioned in different body locations.

3 AUTOMATIC FEATURE

EXTRACTION AND TEST

SCENARIOS

3.1 Extracted Features

Notwithstanding the variety of solutions proposed in

literature, DTW still plays a relevant role for wearable

sensor-based gait recognition. Problems to address

are related to the possible different length of walk sig-

nals, and to noisy acquisition. Moreover, computa-

tional complexity is not negligible when considering

mobile processing. This work investigates the possi-

ble application of Machine Learning procedures in or-

der to extract aggregate features from the signals, and

to select the most relevant ones. The aim is twofold:

from one side, to discard less robust or less informa-

tive features, i.e., those more subject to distortions, or

that present quite ﬂat values across signals; from the

other side, achieving the goal of a lighter though ac-

curate recognition procedure would be better suited to

mobile settings.

In order to evaluate the possible inﬂuence of spe-

ciﬁc feature selection choices, 4 different scenarios

have been conﬁgured, with different characteristics

regarding both the use of training information and the

way to exploit such information.

In all the test scenarios considered for the exper-

iments presented below, the same Python libraries

have been used for feature extraction and analysis.

Tsfresh

library is used to automatically extract a

large number of features from temporal series. It

is usually exploited together with Pandas

for data

analysis, and with Scikit-learn

library for Machine

https://tsfresh.readthedocs.io/en/latest/index.html

http://pandas.pydata.org/index.html

http://scikit-learn.org/stable/

Learning. The extracted features can be later ex-

ploited to create regression or classiﬁcation models,

and to cluster or match time series.

Tsfresh includes library functions to extract a

huge number of features (222) from a time series. Of

course not all of them were taken into account for our

experiments. Some examples follow, but it is not pos-

sible to provide the complete list of them.

• abs energy: returns the absolute energy of the

time series: E =

∑

i=1

with n=number of

points in the time series

• absolute sum of change: returns the sum of ab-

solute values of subsequent variations in the se-

ries: E =

∑

i=1

i+1

−x

• approximate entropy: returns the approximate

entropy of the signal

• ar coefﬁcient: returns the coefﬁcient of the Auto

Regressive (AR) process for a given conﬁguration

passed as parameter

• augmented dickey fuller: returns the result of

Dickey-Fuller test

• autocorrelation: returns autocorrelation given a

certain lag

• count above mean: returns the number of values

in the time series higher than its mean

• count below mean: returns the number of values

in the time series lower than its mean

• cwt coefﬁcients: computes the wavelet transform

using this formula

√

3aπ

(1 −

)exp(−

))

• fft coefﬁcient: computes the Fourier coefﬁcients

applying Fourier Tranform

• mean: returns the mean of the signal

• mean abs change: returns the mean of absolute

values of consecutive changes in the time series

∑

n−1

i=1

i+1

−x

• standard deviation: returns standard deviation

• variance: returns the variance

• median: returns the median value

• skewness: returns the skewness (computed with

the Fisher-Person standardized coefﬁcient)

• kurtosis: returns the kurtosis (computed with the

Fisher-Person standardized coefﬁcient)

It is worth noticing that features are separately

extracted from the signals produced by the three ac-

celerometer axes, and the difference between test sce-

narios also regards the way to take their possible cor-

relation into account.

Feature-based Analysis of Gait Signals for Biometric Recognition - Automatic Extraction and Selection of Features from Accelerometer

Signals

633

3.2 Test Scenarios and Feature Selection

This work analyzes 4 different test scenarios. Test

Scenario 1 (T1) does not use any training phase, while

the others do. Considering the different domains and

scale values of extracted features, a standardization

procedure is exploited to build homogeneous vec-

tors, using the well-known Gaussian normalization

formula. For each feature, the average µ and the stan-

dard deviation σ are computed over gallery templates

and then, for each value x, the resulting standardized

value z is obtained by the formula:

z =

(x −µ)

(1)

The µ and σ values are then stored, in order to normal-

ize the further incoming probes used for testing with

the same gallery. The galleries that are used in turn

for the experiments each contain a number of tem-

plates (more than 450) that allows considering these

parameters stable enough to avoid recomputing them

for each probe. All test scenarios entail recognition in

multiple instance veriﬁcation modality: each subject

has more than one template, all of them are matched

against the incoming probe, and the best match among

the gallery and the probe is returned as veriﬁcation

result. A probe set vs. gallery set distance matrix is

produced in order to evaluate the performances. For

each scenario, distances are computed between pairs

of vectors built according to the scenario setting. Ex-

periments are carried out using both Manhattan and

Euclidean Distance as alternative metrics.

3.2.1 Test Scenario 1 (T1)

In T1, all feature extracted by Tsfresh tool are ex-

ploited. For each axes, all 222 feature are taken into

account, for a total of 666 features in the template

vector. This scenario allowed to have a baseline per-

formance. In order to get a fair comparison with the

other scenarios, the templates from samples in the

training set are not used during testing.

3.2.2 Test Scenario 2 (T2)

The strategy adopted in T2 aims at selecting and keep-

ing only the most relevant features. For each axis,

only the features that have a probability of at least

80% of changing across vectors are taken into ac-

count. In other words, only features presenting the

highest variance are maintained. This analysis is car-

ried out by Scikit-learn library. The next selection

step entails a further pruning, that discards features

that do not present this property for all axes, i.e., those

that are informative enough but only for a subset of

axes. This provides a total of 55 feature per axis,

summing up to 165 features. This feature selection

is carried out in the training phase, so that in testing

only the identiﬁed features are taken into account for

both gallery and probe sets.

3.2.3 Test Scenario 3 (T3)

In T3 the same ﬁrst step of variance-based pruning is

performed as in T2. As a second step of feature se-

lection, the complement of the features identiﬁed in

T2 is maintained. After discarding features that show

a too high homogeneity of values across the training

set, only the features that are relevant for a strict sub-

set of axes (1 or 2) are maintained. In this case, this

sums up to 24 features, 9 from the x axis, 10 from the

y axis, and 5 from z axis. Even in this case, selection

is carried out during training, and the features identi-

ﬁed are then extracted from gallery and probe in the

testing step.

3.2.4 Test Scenario 4 (T4)

T4 uses a totally different approach for feature selec-

tion. In this case the choice of features to be kept

is based on the Principal Feature Analysis (PFA)(Lu

et al., 2007). It uses the same principles of the well-

known Principal Component Analysis (PCA), also

exploited, e.g., in face recognition for feature space

reduction. The same PCA criteria are applied to se-

lect a subset of dimension q of the most representa-

tive features from the complete original set. During

training, the best results are obtained with q=60 and

q=62.

Figure 1: Body locations available from ZJU-gaitacc

dataset. The red circle (pelvis zone), is the one exploited

to get the experimental results.

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

634

Table 1: Results in term of EER for the 4 test scenarios.

TEST

EUCLIDEAN

DISTANCE

MANHATTAN

DISTANCE

T1 24.6% 22.4%

T2 22.5% 20.2%

T3 31.5% 30.6%

T4 19.6% 18.7%

4 EXPERIMENTAL RESULTS

4.1 Dataset

The dataset exploited for the experiments of this work

is the one proposed in (Zhang et al., 2015), named

ZJU-gaitacc. It collects walk signals from 175 sub-

jects, out of which only the 153 selected for this work

have 12 walks each, equally divided into two ses-

sions, with data captured from 5 different body loca-

tions: upper arm, pelvis, wrist, thigh, ankle. In order

to maintain consistency with real life use and possi-

ble positioning of smartphones, only the subset cap-

tured from pelvis has been used. Furthermore, this

accelerometer positioning provides the best results. It

is worth noticing that the proposal of this paper ad-

dresses the use of smartphones for gait authentication,

and that embedded accelerometers presently achieve

a higher signal resolution. However, this dataset is

the best one viable for a fair comparison of meth-

ods. It presents the best characteristics in terms of

both number and length of samples, and also provides

data from two different sessions, so allowing to take

into account time-related variability too.

The dataset has been divided into training and test-

ing sets. The training set contains the ﬁrst 3 walks

of each session, while the testing set contains the re-

maining ones. As for probe and gallery partition of

the testing set, in order to get more results, the walks

from the ﬁrst and from the second session has been

used in turn as either probe or gallery. The obtained

results have been averaged to produce the ﬁnal perfor-

mance measures. Figure 1 shows the body locations

available from the dataset, and the red circle indicates

the one used in this work.

4.2 Results and Discussion

The system has been tested in multiple instance ver-

iﬁcation modality (the best matching between the in-

coming probe and the gallery of the claimed iden-

tity is returned). Performances are reported in terms

of Equal Error Rate (EER). T1 achieves EER=24.6%

with Euclidean Distance (ED) and a slightly bet-

ter EER= 22.6% with Manhattan Distance (MD). T2

shows an improvement, achieving EER=22.5% with

ED and EER=20.2% with MD. This seems to demon-

strate that selecting the features that provide the high-

est information for all axes improves recognition per-

formance. On the contrary, T3 achieves worse per-

formance than T1, namely EER=31.5% with ED and

EER=30.6% with MD. This is probably due to the

too low number of features (24) and possibly to the

uneven distribution across axes. T4 achieves the

best results, represented by EER=19.6% with ED and

EER=18.7% with MD. Overall, the best results are

always obtained by MD, independently from the test

scenario. The performance of PFA demonstrates that

reduction techniques exploiting data correlation are

effective with this kind of temporal series, obtaining

an improvement of about 7% over T2. This is not dra-

matically signiﬁcant, but encourages continuing in-

vestigating along this line. Table 1 summarizes the

obtained results. Figure 2 plots the EER trend of test

scenarios. Figure 3 shows FAR and FRR curves for

all scenarios with both exploited metrics.

It is interesting to make a comparison with re-

sults reported in (De Marsico and Mecca, 2016), ob-

tained on the same dataset using pure DTW for the

same multiple instance veriﬁcation modality. On

one hand, the algorithm matching the whole signal

(EER=9.2%), as well as the best of those exploiting

step segmentation (EER=10.25%), got better results

than the approaches presented here. These two al-

gorithms are the slowest and most computational de-

manding in (De Marsico and Mecca, 2016). More-

over, the ﬁrst one requires signals to be not dramati-

cally different in terms of length. On the other hand,

the other three proposals in (De Marsico and Mecca,

2016) based on step segmentation compute distance

with a matching strategy comparable in terms of com-

putational costs and speed to those proposed here, and

allow releasing the constraint of a similar number of

steps. However, they show lower performances (EER

of 0.328%, 0.4104%, and 0.3625% respectively) than

the approaches presented here. Approaches based on

feature extraction work on a kind of aggregated in-

formation that does not depend on the signal length,

given that it is long enough. The above comparison

seems to suggest a possible compromise between dif-

ferent application constraints, that deserves more in-

vestigation. As a further comparison, the results in

(Zhang et al., 2015) are reported. We considered only

the single accelerometer scenario (i.e., entailing the

same setting of our experiments). That work achieves

an EER that ranges from 8.6% to 13% (depending on

the chosen body location), that is generally better than

Feature-based Analysis of Gait Signals for Biometric Recognition - Automatic Extraction and Selection of Features from Accelerometer

Signals

635

Figure 2: EER trend of the two metrics exploited.

our approach, but it is worth noticing that their proce-

dure requires about 0.3 second to run on a powerful

pc, while our approach is devised to work on smart-

phones that have lower computational power. It is

further worth underlining that making more compar-

isons with other state of the art methods is not pos-

sible at present, due to the different (and generally

much smaller) datasets used. The larger dataset intro-

duced in (Ngo et al., 2014) contains very short signals

that have been manually segmented from longer ones,

acquired in a single session and with a single walk.

Therefore we preferred to use data presenting more

challenging variations.

5 CONCLUSIONS

The main aim of this work has been to try a feature

based approach for gait recognition based on wear-

able sensors. The achieved results fall in between

those provided on a similar dataset using pure DTW.

It is worth noticing that, with respect to better ones

in literature, they are produced by approaches espe-

cially devised to run on personal mobile devices. This

entails aiming at lowest computational costs and ex-

ecution times. In any case, performances are quite

encouraging, and call for more investigation. As a fu-

ture attempt, we plan to apply PFA to an already par-

tially reduced feature set, possibly applying the best

strategy presented here based on features which are

equally informative for signals from all accelerome-

ter axes. In addition, more and possibly wider feature

sets can be investigated.

REFERENCES

De Marsico, M. and Mecca, A. (2016). Biometric walk

recognizer - gait recognition by a single smartphone

accelerometer. Multimedia Tools and Applications.

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

gait authentication using accelerometer sensor. Jour-

nal of computers, 1(7):51–59.

Gafurov, D., Snekkenes, E., and Bours, P. (2010). Improved

gait recognition performance using cycle matching. In

Advanced Information Networking and Applications

Workshops (WAINA), 2010 IEEE 24th International

Conference on, pages 836–841. IEEE.

Lu, Y., Cohen, I., Zhou, X. S., and Tian, Q. (2007). Fea-

ture selection using principal feature analysis. In Pro-

ceedings of the 15th ACM international conference on

Multimedia, pages 301–304. ACM.

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.

Nickel, C., Brandt, H., and Busch, C. (2011a). Classiﬁca-

tion of acceleration data for biometric gait recognition

on mobile devices. BIOSIG, 11:57–66.

Nickel, C., Busch, C., Rangarajan, S., and M

obius,

M. (2011b). Using hidden markov models for

accelerometer-based biometric gait recognition. In

Signal Processing and its Applications (CSPA), 2011

IEEE 7th International Colloquium on, pages 58–63.

IEEE.

Nickel, C., Wirtl, T., and Busch, C. (2012). Authenti-

cation of smartphone users based on the way they

walk using k-nn algorithm. In Intelligent Information

Hiding and Multimedia Signal Processing (IIH-MSP),

2012 Eighth International Conference on, pages 16–

20. IEEE.

Zhang, Y., Pan, G., Jia, K., Lu, M., Wang, Y., and Wu,

Z. (2015). Accelerometer-based gait recognition by

sparse representation of signature points with clusters.

IEEE transactions on cybernetics, 45(9):1864–1875.

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

636

Figure 3: FAR and FRR for all test scenarios using both Euclidean Distance and Manhattan Distance.

Feature-based Analysis of Gait Signals for Biometric Recognition - Automatic Extraction and Selection of Features from Accelerometer

Signals

637