This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2016.2582723, IEEE Internet of
Things Journal

Gender Classification of Walkers Via Underfloor

Accelerometer Measurements
Dustin Bales, Pablo Tarazaga, Mary Kasarda, Dhruv Batra, A.G. Woolard, J.D. Poston, V.V.N.S
Malladi

ABSTRACT – The ability to classify the gender of occupants I. INTRODUCTION

in a building has far-reaching applications including security
and retail sales. The authors demonstrate the success of
machine learning techniques for gender classification. High-
sensitivity accelerometers mounted non-invasively beneath an
actual building floor provide the input for these machine
learning methods. While other approaches using gait
measurements, such as vision systems and wearable sensors,
provide the potential for gender classification, they each face
limitations. These limitations include an invasion of privacy,
occupant compliance, required line of sight, and/or high sensor
density. Underfloor mounted accelerometers overcome these
limitations. The authors utilize the highly-instrumented
Goodwin Hall smart building on the Virginia Tech campus to
measure vibrations of the walking surface caused by walkers.
In this study, the gait of fifteen individual walkers was recorded
as they, alone, walked down the instrumented hallway.
Fourteen accelerometers, mounted underneath the walking
surface, recorded walking trials with the placement of the
sensors unknown to the walker. This work studies Bagged
Decision Trees, Boosted Decision Trees, Support Vector
Machines (SVMs), and Neural Networks as the machine
learning techniques for their ability to classify gender. A ten-
fold-cross-validation method is used to comment on the validity
of the algorithm’s ability to generalize to new walkers. This
work demonstrates that a gender classification accuracy of 88%
is achievable using the underfloor vibration data from the
Virginia Tech Goodwin Hall by using Decision Tree
approaches.
Index Terms – Application Platform, Gait Measurement,
Goodwin Hall, Smart Buildings, Vibration
_____________________________________________
The authors wish to acknowledge the support as well as the collaborative
efforts provided by our sponsors, VTI Instruments, PCB Piezotronics, Inc.;
Dytran Instruments, Inc.; and Oregano Systems. The authors are particularly
appreciative of the support provided by the College of Engineering at Virginia ability of an individual to mask their identity, privacy concerns,
Tech through Dean Richard Benson and Associate Dean Ed Nelson as well as
VT Capital Project Manager, Todd Shelton. The work was conducted through
the support of the Virginia Tech Smart Infrastructure Laboratory and its
nd
members. Submitted for review on the 2 of February 2016.
Dustin Bales is in the Mechanical Engineering Department at the Virginia
Polytechnic Institute and State University, Blacksburg, VA 24061 USA (email:
dbbales@vt.edu). All authors are affiliated with Virginia Tech.
Pablo Tarazaga, Mary Kasarda, A.G. Woolard, J.D. Poston, V.V.N.S
Malladi are also in the Mechanical Engineering Department (respective emails:
patarazag@vt.edu, maryk@vt.edu, awool012@vt.edu, poston@vt.edu,
sriram@vt.edu.)
Dhruv Batra is in the Electrical and Computer Engineering Department
(email: dbatra@vt.edu).
Copyright (c) 2012 IEEE. Personal use of this material is permitted. However, permission to use this material for
any other purposes must be obtained from the IEEE by sending a request to pubs-permissions@ieee.org.
Building occupant classification from non-invasive gait
measurement has the potential to be of prime importance in
numerous applications such as improving sales in a retail space,
advancing security and threat detection, and improving
emergency response. There are various methods of gait
measurement (e.g. computer vision, wearable accelerometry,
etc.), and many have been used to classify a variety of
characteristics of building occupants [1-8]. The novelty of this
work includes the use of non-invasive underfloor mounted
accelerometers as a method of measuring gait and the
application of machine learning techniques to classify the
gender of each walker.
Gait, the manner in which humans walk, consists of the
interaction between hundreds of muscles and joints in the body
[9]. The patterns associated with this movement are similar in
humans, but not identical [9]. These differences in gait lead to
the assumption that gender can be inferred from its analysis.
Nixon and Carter [10] support this idea by suggesting that
gender classification is achievable by observation with cameras
in conjunction with computer vision techniques due to
differences in gender behavior such as larger hip swing for
women and shoulder swing for men.
One approach to organizing existing gait measurement
methods in the literature is to broadly label an approach as
either “visual based” or “non-visual based.”
Visual based methods involve some form of computer vision
to measure the movement of a walker. These measurements are
broken down further into active and passive methods. The
active method, known as Structure from Motion, tracks specific
points on the body to characterize its motion [11]. Passive
methods are not interested in the underlying physical structure
of a body’s motion and treat each frame of the video as a unique
pose made by a walker [12]. Visual based methods have
limitations associated with video recording and computer vision
processing (e.g., line of sight requirement, large data sets, the
etc.).
Non-visual methods achieve gait measurement through a
variety of techniques including step loggers, accelerometer
instrumented vests, mobile force plates, instrumented shoes,
walker mounted accelerometers, sound measurement, and
sensitive floors [5, 8, 13-21]. These technologies have various
limitations that restrict wide adoption. For example, they may
require the active participation of a walker (i.e. wearing a
sensor), may have significant potential for the violation of
privacy (i.e. recording of conversations), and/or pre-installation
of the system is very expensive and/or difficult. Another

2327-4662 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2016.2582723, IEEE Internet of
Things Journal

IoT – 0902 - 2016
subcategory under non-visual methods is ground measurement.
Pan et al. [1, 22] and Sabatier and Ekimov [23, 24] measure
vertical velocity (using geophones) or the vertical acceleration
(using accelerometers) of the ground caused by footsteps. These
works mount sensors on the walking surface where they can
become a trip hazard, and each could be destroyed or evaded
because of his or her obvious location. The underfloor mounted
accelerometers in this work have locations that are unknown to
building occupants. Retrofitting underfloor accelerometers on
new or existing buildings can be relatively easy if floors are
accessible from below.
Much of the data produced in these previous works have been
used for the purpose of classification of different individual
characteristics from walker identification to diagnosing
2
neurodegenerative diseases [4, 7]. The algorithms used ranged
from simple thresholding of a single term to complex non-linear
learning strategies. Yun [8] contains a list of gait classification
examples. Additionally, Table 1 contains an extended list of
gait measurement classification works not included in Yun [8].
The most similar effort in gender classification to our work is
by DeLoney [18] and Li et al. [19]. Both use acoustic
measurements of walkers with gender classification accuracies
that were either not reported (NR) [18] and 75% [19],
respectively. Our work uses machine learning techniques to
interpret the vibrational data measured by the novel underfloor
mounted accelerometers in an operational building (Goodwin
Hall) to classify the gender of a walker.

Table 1. Gait classification overview of previous studies. The reference, the gait measurement method, what characteristics were classified,
accuracy, algorithm, and notes.

Accuracy
Gait Measurement Classification Machine Learning
Source Achieved Special Notes
Method Goal Algorithm
(%)
Walking Surface Footsteps/
[1] 99.55 / 85 Decision Trees N/A
Mounted Geophones Occupancy
Individual
[3] Video Surveillance 93 k-nearest neighbors Leave-one-out-cross-validation
Identification
Human
[4] Video Surveillance 95 k-nearest neighbors Generalized Symmetry Operator
Recognition
Individual
[2] Pressure Sensors 80 Decision Trees N/A
Identification
Walking
97.72 /
Walker Mounted upstairs/ Discrete Dyadic Wavelet
[5] 93.18 / Decision Trees
Accelerometers downstairs/ Transformation
93.93
level ground
Individual 90.40 - Self-organizing-
[6] Video Surveillance N/A
Identification 99.07 maps (SOM)
Various
Database/ Video 79.04 - Support Vector
[7] Neurodegenera- 10-fold cross-validation
Surveillance 93.96 Machines
tive Diseases
Gender/
Microphone Support Vector
[18] Individual NR / 60 Modulation Features
Recording Machines
Identification
Microphone Discriminant
[19] Gender 75 Principal Component Analysis
Recording Analysis

II. EXPERIMENTAL SETUP
The 5-story, 155,000 sq. foot Goodwin Hall on the Virginia
Tech campus has been instrumented with 212 accelerometers,
with more in various stages of installation and commissioning.
All accelerometers are mounted to structural members of the
building via steel mounting beams as shown in Fig. 1d. The
instrumentation of this operational building creates the ideal
“real-world” test bed for walker classification. These
measurements will encounter complications possibly not
present in a simplified, controlled laboratory system. The varied
nature of the building (i.e. classrooms, laboratories,
auditoriums, and staff offices) allows for the developed
technologies to translate directly into many other types of
buildings. Various initial works describe the studies completed
in the building [25-30]. For this study, data was collected in a
single hallway with each participant walking alone. Fifteen
individuals (eight males and seven females) walked in this
study. Walkers were asked to state whether they considered
their shoes hard or soft soled. Three females and two males
were asked to walk in both hard and soft soled shoes while the
rest walked in either hard or soft soled shoes. This strategy was
used to reduce bias in walker characteristics while accounting
for some differences in footwear. Although in previous studies,
footwear has been shown to have a small effect on gait
measurements [19, 31]. In summary, thirteen soft sole and
seven hard sole walkers were tested. Each of the 20 walkers
completed six trials, totaling 120 walking trials. Participants

2327-4662 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2016.2582723, IEEE Internet of
Things Journal

IoT – 0902 - 2016
walked in both directions of the hallway: Eastward (East to
West) and Westward (West to East) for three trials each.
Fig. 1 shows the fourth-floor hallway where the study was
3
multiple features. The dimensionality of the raw data for a
single walking trial is ℝ𝑛×𝑚 where 𝑛 is nominally the number
of samples required to complete a walking trial and 𝑚 is the

Fig. 1. The hallway used for testing: (a) shows a floor plan of the 4th floor depicting the location of the sensors mounted
under the floor, (b) the blue dashed area shows the section used in this study, and (c) show the view from the east side
of the building facing westward. Building north is labeled. (d) a tri-axis mount with three accelerometers. Only a subset
of the sensors in (b) were used. Only vertical accelerometers were used in this study.
completed. Only fourteen of the 212 PCB393B04
accelerometers with nominal 1000 mV/g sensitivity, sampling 0.3
at 51,200 Hz were used. The relatively high sampling rate
0.25
allows for high fidelity time signal resolution while filtering can 0.2
Acc. (g)

reduce resonant effects of each sensor in post-processing. 0.2

0
The time taken to complete a trial was dependent on the 0.15
walking speed of the individual as researchers gave no
Acceleration (g)

0.1 -0.2
instructions about appropriate walking speed. These times 16.25 16.3 16.35 16.4
Time (s)
ranged from approximately 20 seconds to 35 seconds for 0.05

individuals to traverse the hallway. The recorded data was 0

processed by removing the DC offset of each accelerometer, -0.05
giving a zero initial value for each signal, and applying a low
-0.1
pass filter. A Butterworth filter of order three with a 500 Hz
cutoff frequency was used to capture all the expected frequency -0.15
content of the normal force for measurement in the vertical -0.2
0 2 4 6 8 10 12 14 16 18 20
direction [17]. Time (s)
Special care was taken to ensure the only walking recorded
Fig. 2. A typical walking signal from one sensor. This
during the experiment was that of the study participants. Future
particular signal is the first trial of Female A wearing hard soled
work will study methods that enable multiple walkers walking
shoes. The peaks of the signal represent footsteps. The inlaid
along the hallway.
plot shows a portion of the same signal.
III. FEATURE EXPLORATION number of accelerometers used in the study (𝑚 = 14). A typical
A feature set is the extracted data used as inputs to the walking signal is shown in Fig. 2 for reference.
machine learning algorithms. One of the aims of this work is to Due to the large dimensionality, the raw data is not an ideal
understand what makes a useful feature type. There is little a set of features for a machine learning algorithm. Because large
priori knowledge about which walking features will yield the features lead to high computation time and low accuracies in
highest gender classification rate necessitating definition of classification due to overfitting [32]. To reduce the

2327-4662 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2016.2582723, IEEE Internet of
Things Journal

IoT – 0902 - 2016
dimensionality of the data multiple feature sets were developed:
Sensor Averaged Step (SAS), Multiple Sensor Step (MSS), and
Single Sensor Step (SSS). Each of the three feature types was
constructed in the time domain and then transformed to the
frequency domain using a fast Fourier transform (FFT). A
description of the construction of each feature set follows, and
Fig. 3. Representation of initial data in a single
walking trial. Each column represents a single sensor
and a spike in the signal represents a footstep.
Table 2 shows a pictorial summary for each feature type. Table
2 shows the construction of these various feature based on a
schematic walking signal shown in Fig. 3.
The first feature set examined, the SAS feature set, is
constructed by selecting the 𝑛 most prevalent steps (based on
amplitude) recorded by a single sensor during the entirety of a
single walking trial. These 𝑥 steps were maximum value
aligned and averaged. The most appropriate number of steps to
average is unknown a priori, and therefore, 𝑥 = 10, 5, and 3 are
chosen to construct three SAS feature sets (denoted SAS10,
SAS5, and SAS3). The SAS feature sets were then transformed
to the frequency domain, using an FFT, to study domain type
influence on feature representation.
4
The second feature set examined, MSS, is constructed by
taking the most prevalent step (based on amplitude) recorded
by accelerometer 4 and then recording the same footstep in time
measured by accelerometers 2 and 6. These three steps are
concatenated so that the signal contains the three steps in the
specified order (2, 4, 6 for Westward trials and 6, 4, 2 for
Eastward trials). Fig. 4 details a representative layout of sensors
2, 4, and 6. The MSS feature set was constructed after
recognizing that the averaging of SAS may adversely affect the
machine learning capabilities. The averaging caused the loss of
high-frequency content and spatial information.
The third feature set examined, SSS, is constructed by taking
the most prevalent single step recorded by sensor 1, see Fig. 4.
This feature set is used as a control to determine the ability of a
single accelerometer and single step to classify gender. A single
footstep in all time feature sets was defined as the 61.0 ms
before a footstep peak and 152.6 ms after the peak, as shown in
the inlaid figure of Fig. 4. The development of the features was
completed with the MATLAB ‘peakfinder’ command to
identify footsteps to then peak align and average.
In summary, there are ten sets of data. The SAS10, SAS5,
SAS3, MSS, and SSS in the time domain (denoted time), and
the frequency domain (denoted frequency).
1
6 4 2
Fig. 4. Illustration of sensor arrangement. The spatial relation
between sensor 1, 2, 4, and 6 in the testing hallway. These
sensors are used in the construction of the MSS and SSS data
types.

Table 2. Illustration of the feature type construction. The steps required to develop the three feature types (SAS, MSS, and
SSS) in the time domain are illustrated schematically for ease of understanding.

Feature
Step 1 Step 1 Representation Step 2 Step 2 Representation
Type

Averagi- Sensor 1 Sensor 2 Sensor 3 Sensor 14

Sensor 1 Sensor 2 Sensor 3 Sensor 14
ng of 𝑛
Concatenati-
steps of
SAS on
each
sensor

Dimension: ℝ7001×14 Dimension: ℝ98014×1

Sensor 2 Sensor 4 Sensor 6
Most
prominen Concatenati- Sensor 2 Sensor 4 Sensor 6
-t step on * Based
MSS recorded on walking
by direction
Sensor 4

Dimension: ℝ7001×3 Dimension: ℝ21003×1

2327-4662 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2016.2582723, IEEE Internet of
Things Journal
IoT – 0902 - 2016
Single
most
prominen
-t step
SSS recorded
from
only
Sensor 1
Dimension: ℝ7001×1
IV. MACHINE LEARNING TECHNIQUES
This work uses supervised machine learning to classify
gender. The machine learning algorithms used in this
work require three steps: training, validation, and testing.
The training step develops a generalized predictive
model, the validation step sets or updates ‘hyper-
parameters’ or model design choices, and the test step
determines the ability of the algorithm to classify new
walker measurements. The features used in this work are
the ten previously discussed feature sets and the gender
labels for each walker. The machine learning algorithms
develop models that make predictions on new. The error
(i.e. the percentage of gender misclassifications) of each
step of machine learning is tracked, and only test errors
are reported throughout this work.
To reduce the effect of feature initialization (i.e. which
trials of the walking features are sorted into each of the
three steps above) this work uses a ten-fold-cross-
validation method. This method randomly partitions
feature sets into ten, nearly equal parts, known as folds.
A single fold is held out, and the remaining nine folds
are used in the training and validation steps. The
developed model is used to classify the held out tenth
fold, and the test error is recorded. Then, a different fold
is held out, with this process repeating until all folds have
been held out once. The averaged test error represents
the error that can be expected on a never before seen
walker [33].
An aim of this work is to understand the effect of type
of machine learning algorithm on the final gender
classification abilities. The machine learning techniques
used in this work are Bagged Decision Trees, Boosted
Decision Trees, Support Vector Machines (SVMs), and
Neural Networks (NNs). Each algorithm type is briefly
discussed.
A) Decision Trees
Decision trees involve the splitting of feature sets at a
node until there is a reasonably high probability of a
specific classification. A node represents the splitting
point for a single feature, and a branch represents the
flow of features to further nodes. The training and
validation steps of these algorithms determine which
features are the most important to split on and the
appropriate depth of a tree (i.e. how many nodes are
necessary for an acceptable level of classification). A
2327-4662 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
5
Sensor 1
new walker’s features would be evaluated at each node
until classification occurs at a leaf to test the
classification ability of the tree. Further reading on
decision trees is found in Quinlan [34].
Choosing the most appropriate features to split on is
difficult a priori and leads to the desire to combine many
decision trees through ensemble learning methods:
bagged decision trees and boosted decision trees.
Bagged and boosted machine learning algorithms are
examples of ensemble methods that combine multiple
learning algorithms to solve complex problems. These
two examples manipulate the feature used in training to
generate multiple models [35]. For background before
further discussion, bias is an error due to assumptions in
the model and variance is an error due to small changes
in the training dataset.
Bagging machine learning algorithms aim to reduce
the variance associated with high variance, low bias
algorithms [35, 36]. The process of bagging takes the
original training dataset and constructs multiple datasets
by sampling instances (with replacement) from the
original dataset. It can be shown that this process results
in approximately 63% of unique samples in each newly
created “bagged” dataset [36]. A decision tree is
developed using the labels and features of this bagged
dataset. This process repeats for a specific number of
bags, and the test errors of these models are averaged for
a representative test error for the entire system.
Boosted machine learning algorithms try to reduce the
bias associated with high bias, low variance systems
[37]. In the process of boosted learning, every simple
model (i.e. a decision tree) is weighted based on its
ability to correctly classify. The final model is a
weighted combination of the results of the simple
models. This work uses the AdaBoost method [35].
B) Neural Networks
Neural Networks (NNs) are machine learning
algorithms that, at the highest level, map features to
hidden neurons and then to an output label through a
series of non-linear functions (i.e. Sigmoid, etc.) [38].
Each feature contributes a weight to a neuron in the
hidden layer. This work uses a linear mapping of features
to neurons in a hidden layer. A single hidden layer was
used in this work as more layers can add unnecessary
computational complexity [39]. All of the weights at a
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2016.2582723, IEEE Internet of
Things Journal
IoT – 0902 - 2016
single neuron then pass through a non-linear function,
and the outputs are weighted and combined into the
output layer. A final, chosen non-linear function
generates a classification label for gender in this work.
C) Support Vector Machines
These algorithms ‘learn’ a hyperplane between the
classes (i.e. female and male), and in particular, they
learn the hyperplane that maximizes the margin between
the two classes. Often associated with SVMs is the use
of what is called the kernel trick. The kernel trick allows
for the replacement of inner products (i.e. measures of
similarity) with kernel functions [40]. Kernels map the
feature sets into a higher dimensional space, via a chosen
non-linear mapping. The non-linear mapping (i.e.
polynomial) is chosen a priori to develop an optimal
hyperplane that separates the classes [41, 42]. Based on
their wide use and high rates of success, Radial Basis
Function (RBF) and polynomials of order 1-25 SVMs
are used in this work to understand how the type of SVM
affects classification rate.
The machine learning algorithms were developed
using MATLAB commands: ‘fitensemble’, ‘patternnet’
and ‘fitcsvm’ respectively. The ten-fold-cross-validation
was completed either internally to the machine learning
command or through the ‘KFold’ command. Each of the
four learning algorithms were combined with the ten
feature types to examine the effectiveness of each
combination to successfully classify gender.
V. RESULTS
The ten previously discussed feature types (SAS10,
SAS5, SAS3, MSS, SSS, in both time and frequency)
were all used as input with each of the four machine
learning algorithms. This combination allowed for the
answering of 1) which feature type is most informative,
2) which machine learning algorithm is most effective,
and 3) whether the time or frequency domain houses the
most important classification information. The overall
range accuracies for all feature set types and machine
learning algorithm combinations was 95 – 42%.
Fig. 5 shows the classification error results for all
combinations of feature types and machine learning
algorithms in the study. There are two broad columns
representing the time and frequency domains and the
columns further subdivided show feature types. The
dependent axis is the average error for each algorithm
feature type combination.
To better determine the significance of the results
presented in Fig. 5, a statistical test is applied.
Specifically, a one-way analysis of variance, ANOVA,
and a Tukey test with a confidence of 5%. The Tukey
test examines all possible pairs of means (in this work
average classification error) to find pairs that are
statistically different from each other to evaluate
performance.
2327-4662 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
6
A) Comparison of Feature Type Selection
The statistical testing reveals that there is
indistinguishability between the SAS5 and SAS10
feature types. While the mean error of the SAS5 type in
the time and frequency domain is the lowest reported at
22.9% and 33.8%, respectively. All feature type errors
were developed while considering all four machine
learning algorithm types. This method is to ensure
robustness regardless of machine learning type.
B) Comparison of Machine Learning Types
We find that Bagged and Boosted Decision Trees as
well as SVMs are all effective machine learning
algorithms for gender classification in our dataset.
Neural Networks perform statistically worse than the
other algorithms. When examining the mean error of the
machine learning algorithms, Bagged Decision Trees
produced a classification error rate of 21.5% and 27.8%
across all time and frequency domain feature types,
respectively. Considering the statistical analysis across
the averaging across feature types is to build robust
machine learning algorithms independent of which
feature type.
C) Comparison of Time Frequency Domain
As shown in Fig. 5, the time domain data input gives a
lower misclassification rate than the frequency domain.
The overall accuracy of the time domain is 70% while
the frequency domain is 64%. This result may be partly
due to the relative richness of input data in the time
domain versus the frequency domain. Generally,
dynamicists “simplify” complex time series data by
transforming into the frequency domain to understand
underlying phenomena. However, a larger number of
peaks in the time domain are represented only by a
dominating peak in the frequency domain. This result
means that the time domain has a larger amount of
potentially interesting (non-zero) features than the
frequency domain. This hypothesis is supported by
Ekimov and Sabatier [17], showing that the energy in the
frequency domain of gait is bounded by 500 Hz for the
normal force interaction between the foot and the
ground. The frequency domain in this study contains
zero amplitude higher frequency terms. These zero
values may only add noise to the machine learning
algorithm. The sensitivity of specific machine learning
algorithms to the features used may also explain some of
the differences between the two domains.
D) Machine Learning – Feature Combination
The single best performing feature type and machine
learning combination was the boosted decision trees
using SAS5 time as features, with an accuracy of 88.3%
in the classification of the gender of an individual
walker. Similar results may be possible using either the
bagged decision tree or SVM as the learning algorithm
and either SAS5 or SAS10 features, as these produce
statistically indeterminate results.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2016.2582723, IEEE Internet of
Things Journal

7
IoT – 0902 - 2016
the large number of features and the desire to keep the
developed neural network single layered has led to the
relatively poor results. A more complex NN or
successfully selecting lower dimensionality features is
predicted to improve algorithm performance [32]. This
is beyond the scope of the current work.
The optimum polynomial degree for the SVMs (24)
and the optimum number of neurons in the neural
networks (2) were determined by iterative search. There
is relatively low change in accuracy of classification as
the degree of polynomial increases, as shown in Fig. 6
suggesting robustness to overfitting. The number of
neurons has little effect on the accuracy of classification
leaving little confidence in robust classifying with the
NN single-layer scheme chosen for the study.
Parameter Study

0.5

0.45

0.4
SVM
Percent Error 0.35 Neural Networks

0.3

0.25

0.2

0.15

0.1
5 10 15 20 25
SVM Polynomial Order / NN Number of Neurons
Fig. 5. Average error result of all feature types and
machine learning algorithms. Fig. 6. SVM and NN parameter study. The independent
axis is either the order of the SVM polynomial or the
E) Notable Observations
number of neurons in the neural network each plotted
Additional observations beyond the primary test goals
against the percent error for gender classification.
were made regarding the success of the “control” SSS
feature type, poor NN performance, and parameter
VI. CONCLUSIONS
studies of SVMs and NNs.
Using data from underfloor mounted accelerometers as
The investigation of the control feature set, SSS, reveals
input feature sets for machine learning techniques shows
that with using data from only a single footstep and
promising results for classifying the gender of an
accelerometer shows that gender classification is
individual walker. The use of vibration data from this
possible with little gait information. The boosted
novel mounting of accelerometers is out of sight,
decision trees produced the highest accuracy rate of 80%
allowing for an additional level of both privacy and
accuracy with this SSS feature set. This result implies
tamper resistance not achievable with visual systems or
that a high density of gender classification information
microphones used in current gait measurement
lies within a single footstep. If higher classification rates
techniques. The primary results of this work are that a
are desired, then multiple footsteps should be recorded
Bagged or Boosted Decision Tree and Support Vector
as shown by the better performance of the SAS. The
Machines produce statistically equivalent gender
addition of 4 more steps (SAS5) only increases accuracy
classification results and seem robust to the type of
by 8%, although there is four times as much data, and
feature set used as an input. All of the feature types
therefore, an increase of computation time and storage.
constructed in the time domain produce better gender
The low accuracies produced by NNs are not
classification results. The time domain provides more
surprising due to the number of weights that must be
useful features for successful gender classification than
learned for this type of algorithm. This particular form of
the frequency domain. The single best classifier created
the algorithm must learn 𝑛𝑑 + 𝑛 number of weights,
was a Bagged Decision Tree using the SAS5 time feature
where 𝑛 is the number of neurons and 𝑑 is the
set (this feature set is created by peak aligning the five
dimensionality of the feature set. It is hypothesized that
most prevalent steps recorded by each sensor and

2327-4662 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2016.2582723, IEEE Internet of
Things Journal
IoT – 0902 - 2016
averaging) with an accuracy of 88% in gender
classification.
Ongoing studies have shown that robust features are
capable of accommodating multiple floor types (e.g.,
concrete and hardwood). Future work will address the
robustness of the proposed feature types for
classification on different floor types.
REFERENCES
[1] S. Pan, A. Bonde, J. Jing, L. Zhang, P. Zhang, and H. Y.
Noh, "BOES: Building Occupancy Estimation System
using sparse ambient vibration monitoring," 2014, p. NA.
[2] L. Middleton, A. A. Buss, A. Bazin, and M. S. Nixon, "A
floor sensor system for gait recognition," in Fourth IEEE
Workshop on Automatic Identification Advanced
Technologies, 2005, pp. 171-176.
[3] C. BenAbdelkader, R. Cutler, H. Nanda, and L. Davis,
"EigenGait: Motion-Based Recognition of People Using
Image Self-Similarity," in Audio- and Video-Based
Biometric Person Authentication, J. Bigun and F. Smeraldi,
Eds., ed: Springer Berlin Heidelberg, 2001, pp. 284-294.
[4] J. B. Hayfron-Acquah, M. S. Nixon, and J. N. Carter,
"Recognising human and animal movement by symmetry,"
in International Conference on Image Processing.
Proceedings, 2001, pp. 290-293.
[5] M. N. Nyan, F. E. H. Tay, K. H. W. Seah, and Y. Y. Sitoh,
"Classification of gait patterns in the time–frequency
domain," Journal of Biomechanics, vol. 39, pp. 2647-2656,
2006.
[6] Y.-C. Lin, B.-S. Yang, Y.-T. Lin, and Y.-T. Yang, "Human
Recognition Based on Kinematics and Kinetics of Gait,"
Journal of Medical and Biological Engineering, vol. 31, pp.
225-263, 2010.
[7] M. Yang, H. Zheng, H. Wang, and S. McClean, "Feature
selection and construction for the discrimination of
neurodegenerative diseases based on gait analysis," in 3rd
International Conference on Pervasive Computing
Technologies for Healthcare, 2009. PervasiveHealth 2009,
2009, pp. 1-7.
[8] J. Yun, "User Identification Using Gait Patterns on
UbiFloorII," Sensors (Basel, Switzerland), vol. 11, pp.
2611-2639, 2011.
[9] C. BenAbdelkader, R. Cutler, and L. Davis, "Stride and
cadence as a biometric in automatic person identification
and verification," in Fifth IEEE International Conference
on Automatic Face and Gesture Recognition. Proceedings,
2002, pp. 372-377.
[10] M. S. Nixon and J. N. Carter, "Automatic Recognition by
Gait," Proceedings of the IEEE, vol. 94, pp. 2013-2024,
2006.
[11] D. D. Hoffman and B. E. Flinchbaugh, "The interpretation
of biological motion," Biological Cybernetics, vol. 42, pp.
195-204, 1982.
[12] C. Cedras and M. Shah, "A survey of motion analysis from
moving light displays," in, 1994 IEEE Computer Society
Conference on Computer Vision and Pattern Recognition,
1994. Proceedings CVPR '94, 1994, pp. 214-221.
[13] T. Liu, Y. Inoue, K. Shibata, and K. Shiojima, "A Mobile
Force Plate and Three-Dimensional Motion Analysis
System for Three-Dimensional Gait Assessment," IEEE
Sensors Journal, vol. 12, pp. 1461-1467, 2012.
[14] M. Sousa, A. Techmer, A. Steinhage, C. Lauterbach, and P.
Lukowicz, "Human Tracking and Identification using a
Sensitive Floor and Wearable Accelerometers," IEEE
International Conference on Pervasive Computing and
Communications (PerCom), pp. 166-171, 2013.
[15] J. Sinclair, S. J. Hobbs, L. Protheroe, C. J. Edmundson, and
A. Greenhalgh, "Determination of gait events using an
externally mounted shank accelerometer," Journal of
Applied Biomechanics, vol. 29, pp. 118-122, Feb 2013
2013.
2327-4662 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
8
[16] A. Godfrey, R. Conway, D. Meagher, and G. ÓLaighin,
"Direct measurement of human movement by
accelerometry," Medical Engineering & Physics, vol. 30,
pp. 1364-1386, 2008.
[17] A. Ekimov and J. M. Sabatier, "Vibration and sound
signatures of human footsteps in buildings," The Journal of
the Acoustical Society of America, vol. 120, pp. 762-768,
2006.
[18] C. DeLoney, "Person Identification and Gender
Recognition from Footstep Sound using Modulation
Analysis," p. NA, 2008.
[19] X. Li, R. J. Logan, and R. E. Pastore, "Perception of
acoustic source characteristics: Walking sounds," The
Journal of the Acoustical Society of America, vol. 90, pp.
3036-3049, 1991.
[20] G. Qian, J. Zhang, and A. Kidané, "People Identification
Using Floor Pressure Sensing and Analysis," IEEE Sensors
Journal, vol. 10, pp. 1447-1460, 2010.
[21] M. Kohle and D. Merkl, "Identification of gait patterns with
self-organizing maps based on ground reaction force.,"
1996, pp. 24-26.
[22] S. Pan, N. Wang, Y. Qian, I. Velibeyoglu, H. Y. Noh, and
P. Zhang, "Indoor Person Identification Through Footstep
Induced Structural Vibration," presented at the Proceedings
of the 16th International Workshop on Mobile Computing
Systems and Applications, New York, NY, USA, 2015.
[23] J. M. Sabatier and A. E. Ekimov, "A Review of Human
Signatures in Urban Environments Using Seismic and
Acoustic Methods," in 2008 IEEE Conference on
Technologies for Homeland Security, 2008, pp. 215-220.
[24] J. M. Sabatier and A. E. Ekimov, "Range limitation for
seismic footstep detection," 2008, pp. 69630V-69630V-12.
[25] J. M. Hamilton, B. S. Joyce, M. E. Kasarda, and P. A.
Tarazaga, "Characterization of Human Motion Through
Floor Vibration," in IMAC XXXII, Orlando, FL, February
3-6, 2014.
[26] B. S. Joyce and P. A. Tarazaga, "Estimating Noise Spectra
for Data from an Instrumented Building," in IMAC XXXIV
A Conference and Exposition on Structural Dynamics,
Orlando, Florida, January 25-28, 2016.
[27] M. E. Kasarda, P. A. Tarazaga, M. Embree, S. Gugercin, A.
Woolard, B. S. Joyce, et al., "Detection and Identification
of Firearms Upon Discharge Using Floor-Based
Accelerometers," in IMAC XXXIV A Conference and
Exposition on Structural Dynamics, Orlando, Florida,
January 25-28, 2016.
[28] J. D. Poston, R. M. Buehrer, V. S. Malladi, A. G. Woolard,
and P. A. Tarazaga, "Vibration Sensing in Smart
Buildings," in International Conference on Indoor
Positioning and Indoor Navigation (IPIN), Banff, Alberta,
Canada, 13-16 October, 2015.
[29] J. D. Poston, J. Schloemann, R. M. Buehrer, V. S. Malladi,
V. S. Woolard, and P. A. Tarazaga, "Towards Indoor
Localization of Pedestrians via Smart Building Vibration
Sensing," in International Conference on Localization and
GNSS, June 22-24, 2015.
[30] J. Schloemann, V. V. N. S. Malladi, A. G. Woolard, J. M.
Hamilton, R. M. Buehrer, and P. A. Tarazaga, "Vibration
Event Localization in an Instrumented Building," in
International Modal Analysis Conference (IMAC) XXXIII,
Orlando, Florida, February 2-5, 2015.
[31] R. J. Orr and G. D. Abowd, "The Smart Floor: A
Mechanism for Natural User Identification and Tracking,"
2000, pp. 275–276.
[32] I. Czarnowski and P. Jȩdrzejowicz, "Data Reduction
Algorithm for Machine Learning and Data Mining," in New
Frontiers in Applied Artificial Intelligence, N. T. Nguyen,
L. Borzemski, A. Grzech, and M. Ali, Eds., ed: Springer
Berlin Heidelberg, 2008, pp. 276-285.
[33] B. Parmanto, P. W. Munro, and H. R. Doyle, "Reducing
Variance of Committee Prediction with Resampling
Techniques," Connection Science, vol. 8, pp. 405-426,
1996.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2016.2582723, IEEE Internet of
Things Journal

9
IoT – 0902 - 2016
[34] J. R. Quinlan, "Induction of Decision Trees," Machine [39] G. Cybenko, "Approximation by superpositions of a
Learning, vol. 1, pp. 81-106, 1986. sigmoidal function," Mathematics of Control, Signals and
[35] T. G. Dietterich, "Ensemble Methods in Machine Systems, vol. 2, pp. 303-314, 1989.
Learning," in Multiple Classifier Systems, ed: Springer [40] K. P. Murphy, Machine Learning: A Probabilistic
Berlin Heidelberg, 2000, pp. 1-15. Perspective: MIT Press, 2012.
[36] L. Breiman, "Bagging Predictors," Machine Learning, vol. [41] B. E. Boser, I. M. Guyon, and V. N. Vapnik, "A Training
24, pp. 123-140, 1996. Algorithm for Optimal Margin Classifiers," 1992, pp. 144–
[37] R. E. Schapire, "The strength of weak learnability," 152.
Machine Learning, vol. 5, pp. 197-227, 1990. [42] V. N. Vapnik, The Nature of Statistical Learning Theory.
[38] S. Haykin, A comprehensive foundation: Neural Networks, New York, NY: Springer New York, 2000.
2004.
Prof. Batra is a recipient of Carnegie Mellon Dean's
Biographies Fellowship (2007), two Google Faculty Research
Dustin Bales received his B.S. from the University of Awards (2013, 2015), Virginia Tech Teacher of the
Wyoming in 2014, and his M.S. from Virginia Tech in Week (2013), Army Research Office (ARO) Young
2016, both in engineering. Investigator Program (YIP) award (2014), the National
Science Foundation (NSF) CAREER award (2014), and
His research includes gait analysis and characteristic Virginia Tech CoE Outstanding New Assistant Professor
classification through advance computational methods. award (2015). Research from his lab has been featured
He currently serves as Chief Business Officer for GAiTE in Bloomberg Business, The Boston Globe, MIT
LLC, a company progressing the Internet of Things. Technology Review, Newsweek, WVTF Radio IQ, and
a number of popular press magazines and newspapers.
Pablo A. Tarazaga received his B.S. degree from the
University of Puerto Rico at Mayaguez in 2002, and his A.G. Woolard is a PhD candidate in
M.S. and Ph.D. degrees from Virginia Tech in 2004 and Mechanical Engineering at Virginia Tech. His research
2009, respectively all in Mechanical Engineering. area is in wave propagations in elastic solids for the
purpose of tracking people in instrumented buildings
He is currently an Assistant Professor in the Department through footstep induced floor vibrations.
of Mechanical Engineering at Virginia Tech where he
directs the Virginia Tech Smart Infrastructure J. D. Poston is currently a Ph.D. student with research
Laboratory and the Vibrations, Adaptive Structures and interests in novel indoor localization techniques and their
Testing Laboratory. He is the receipt of the 2015 Air privacy implications. Also, he is interested in the
Force Office of Scientific Research Young Investigator application of machine learning and signal processing
Award. techniques to cyber-physical systems.

Mary Kasarda received her B.S., M.S., and PhD from Previously, he worked in industry in the fields of
the University of Virginia in 1984, 1988, and 1997, telecommunications and defense.
respectively. She worked as a mechanical engineer at He received a B.S. degree from Virginia Tech and M.S.
Ingersoll-Rand, Du Pont, and Rotor Bearing Dynamics, degree from George Washington University.
between her academic degrees. She is currently an He is a Senior member of the IEEE.
Associate Professor in the Department of Mechanical
Engineering at Virginia Tech. Her research interests V. V. N. S. Malladi received the B.Tech degree in
include Smart Building Technologies, Machinery Health Mining Machinery engineering from Indian Institute of
Monitoring, and Engineering Education. She is an NSF Technology (or ISM) – Dhanbad, India, in 2011, and
CAREER Award winner and a Fellow of the American M.S. and PhD degrees in Mechanical Engineering from
Society of Mechanical Engineers (ASME). Virginia Polytechnic Institute and State University,
Blacksburg in 2013 and 2016 respectively. He is
Dhruv Batra is an Assistant Professor at the Bradley currently a postdoctoral researcher working on modal
Department of Electrical and Computer Engineering at testing, waves, and gait analysis.
Virginia Tech, where he leads the VT Machine Learning
& Perception group. His research interests lie at the
intersection of machine learning, computer vision, and
AI, with a focus on developing intelligent systems that
are able to concisely summarize their beliefs about the
world, integrate information and beliefs across different
sub-components or `modules' of AI (vision, language,
reasoning) to extract a holistic view of the world, and
explain why they believe what they believe.

2327-4662 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.