Instrumentation & Measurement Magazine 24-6 - 6

How Can Technology Play a Central Role
in this Context?
It must be highlighted that the monitoring of falls and ADLs,
through the use of non-invasive and easy-to-use sensing systems,
could be strategic to achieve a real time awareness of user
status, in order to promptly perform actions to effectively support
end-users' needs.
Fall detectors can be classified into: wearables devices,
that exploit (inertial) sensors both in the form of customized
systems [8], [9] and smartphone-based platforms [10];
non-wearable systems (ambient sensors, vision sensors, and
radio-frequency sensors) [11]; and hybrid systems. An extensive
review of fall detection systems, including comparisons
among different approaches, is available in [12].
The state-of-the-art regarding solutions for postural analysis
propose different approaches both in terms of hardware
and signal processing. Some examples that exploit force platforms
for postural analysis are available in [13]. Also, visual
systems are nowadays considered a gold standard in the area
of motion analysis, with specific regards to gait analysis [14].
Despite performances of above-mentioned strategies, the
use of wearable devices to monitor postural instability has
been long recognized as a valuable and advantageous alternative
compared with other solutions [15]. The main motivation
supporting this statement relies in the main characteristic of
wearable devices, since they do not require any structured environment,
thus enabling their use at home.
Also, the research group working at the SensorLab of the
University of Catania, Italy, has developed many solutions in
the framework of AT, including Fall Detectors [16], [17], multisensor
systems for Postural Sway detection [18], platforms to
support weak people in indoor environments [19], [20] and autonomous
architectures that exploit advanced architecture for
energy harvesting [21].
A Smart Home solution is represented by outcomes generated
by the NATIFLife project [22], funded by the INTERREG
V-A Italy Malta Cooperation Programme, which aimed to realize
an advanced platform of AT, fostering the independence
and autonomy of elderly and people with impairments living
at home. The platform includes several solutions, from
the traditional monitoring of the environment and user biophysic
parameters to novel solutions for the monitoring of
users' habits, their activity rates, as well as the User-Environment
Interaction.
Reliability is one of the critical requirements of AT. Crucial
aspects related to the reliability of AT, including questions related
to the user trust in this kind of support, are addressed
in [7]. As always, a crystal-clear example is the one related to
fall detectors: " a robust fall detector is one that classifies falls
as falls and non-falls as non-falls, even under real-life conditions. "
One major challenge with this kind of monitoring
system is the generation of False Negatives and False Positives,
which may be very dangerous or unpleasant, and they
will be later addressed through this work. This is particularly
true in the case of solutions aimed to classify an unknown
pattern as belonging or not to a specific class of events (falls,
6
anomalous postural sway, ADLs). In this specific case, the reliability
is strictly related to the use of a robust classification
approach, which includes appropriate pre-post signal processing,
the generation of meaningful features and scores, as well
as an effective classification paradigm. These aspects are addressed
in the following sections.
Classification Strategies
A general representation of a classification strategy is presented
in Fig. 1. Many solutions use data provided by MEMS
(e.g., tri-axial accelerometers, inclinometers, or gyroscopes).
The pre-processing step elaborates raw measurements for the
next paradigms. A carefully designed pre-processing is mandatory
for the optimal extraction of useful information from
raw data. Results of pre-processing are represented by " features, "
e.g., each filtered component of inertial quantities or
their module. Features are then used for generating " scores. "
The latter are the real quantities exploited by the classification
algorithm.
Scores Generation: From Features Generation to
Similarity Measurement
A first approach to generate scores consists in combining information
provided by features. Typical examples can be found in
postural sway monitoring tasks, which are based on the analysis
of variation of the body's Center Of Pressure (COP). The
COP is built, considering the relative variation of the MedioLateral
(ML) and Antero-Posterior (AP) movements of the user
body that are derived by the acceleration components. Different
scores have been suggested by the scientific community
to extrapolate useful information on the user postural status,
such as: Mean Velocity, Rectangular area, Root Mean Square
(RMS) displacement, Displacement Range of the projection of
the center of mass, Total Displacement, and Elliptical area. A
description of such quantities can be found in [23].
Although (Non-)linear time-dependent scores have proven
to be valuable in assessing " body posture, " these may be not
suitable to provide information regarding specific body reactions
involved in postural control. An alternative solution to
generate scores can be represented by the use of Wavelet Decomposition,
which provides different timescales at different
time instants, conveying deep information on the user posture.
As an example, starting from inertial features (filtered components
or module) the following quantities, estimated for each
of the considered levels of transformation, have been successfully
adopted for the sake of postural sway monitoring: Mean
Value (MV), Standard Deviation (STD), and Energy (E) content
[24].
Another approach for scores estimation is based on EventDriven
strategies [16]. The idea behind this approach is that,
for each class of events, the time evolution of considered features
(e.g., acceleration components) are characterized by
well-defined shapes known as " signatures " [16]. The estimation
of a similarity index between features of the unknown
pattern and the whole set of signatures will generate the set
of scores.
IEEE Instrumentation & Measurement Magazine
September 2021
Instrumentation & Measurement Magazine 24-6

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-6
