Aerospace and Electronic Systems Magazine May 2018 - 17

Di Domenico et al.
a moving person in the environment causes a higher spreading
of Doppler spectrum than that of a static person. Starting from
that result, which confirms the intuition behind our key idea, it
seems feasible to find some features that quantify the spreading of the Doppler spectrum and allow recognizing different
scenarios.

FEATURE EXTRACTION FROM THE DOPPLER SPECTRUM
The mean Doppler spectrum D(k) provides the mean contribution of all the subcarriers to the Doppler frequency k. The
Doppler spectrum is a measure of the weighted distribution of
the speed of the scatterers. Because most of the scatterers are
usually fixed, the mean Doppler spectrum D(k) is concentrated
on the 0-Hz Doppler frequency. The presence detection of a
person in a static or dynamic condition is achieved through a
classification task applied to a selected set of spectral features
of the mean Doppler spectrum. A nonexhaustive list of spectral
features that can be computed on the mean Doppler spectrum
is shown in Table 1. See [15] for a more detailed description of
the features.

FEATURE SELECTION AND CLASSIFICATION
In this work, we applied a naïve Bayes classifier to a set of features
computed on the mean Doppler spectrum. The data set containing
the overall set of features has been divided into a training set (50%)
and a test set (50%).
Starting from the list of features provided in Table 1, we
have computed the mean classification accuracy over the test
environments for each pair of features. Then, we have chosen
the pair of features that provides the best classification accuracy
over both environments. As a result, the selected pair of features are spectral spread and Doppler ratio of the mean Doppler
spectrum. The scatter plots for the selected features are shown
in Figures 4-6 for Setup 1, Setup 2, and Setup 3, respectively.

Table 1.

List of Features Extracted from the Mean Doppler
Spectrum
Feature

Definition

Root mean square
Doppler bandwidth

Is the square root of the
second moment of the
Doppler spectrum

Spectral spread

Is the second central
moment of the Doppler
spectrum with log
frequencies

Spectral roll-off

Is the 95th percentile of
the Doppler spectrum
distribution across
frequencies

Spectral slope

Is the rate of falloff of the
Doppler spectrum

Spectral flatness

Is a measure of the flatness
of the Doppler spectrum

Spectral skewness

Is a measure of the
asymmetry of the Doppler
spectrum about its mean

Spectral kurtosis

Is a measure of the
tailedness of the Doppler
spectrum

Doppler ratio

Is the ratio between the
amplitude of the third and
the first component the
Doppler spectrum
Using these features, we note from
the scatter plots that different classes
are grouped in various clusters in the
feature space, and this allows us to
achieve very good performance during the classification task, as will be
shown in the next section.

EXPERIMENTAL RESULTS

Figure 3.

Magnitude of average Doppler spectrum for empty room, presence (static), and presence (dynamic)
scenarios obtained in Setup 2.

MAY - JUNE 2018

IEEE A&E SYSTEMS MAGAZINE

As explained in the previous section,
experiments have been carried out
in different positions of a person, in
different static conditions (sitting or
standing firm), and in three different
setups. The classifier has been trained
separately for each position, using
50% of each data set. We have consid17

Table of Contents for the Digital Edition of Aerospace and Electronic Systems Magazine May 2018