Stress and anxiety are part of the human mental process which is often unavoidably yield by circumstances and situations such as waiting for a ight at the airport gate, hanging around before an exam, or while in an hospital waiting room. In this work we devise a decision system for a robotic aroma di usion device designed to lessen stress and anxiety-related behaviors. The robot is intended as designed for deployments in closed environments that resembles the aspect and structure of a waiting room with di erent chairs where people sit and wait. The robot can be remotely driven by means of an arti cial intelligence based on Radial Basis Function Neural Networks classi ers. The latter is responsible to recognize when stress or anxiety levels are arising so that the di usion of speci c aromas could relax the bystanders. We make use of thermal images to infer the level of stress by means of an ad hoc feature extraction approach. The system is prone to future improvements such as the re nement of the classi cation process also by means of accurate psychometric studies that could be based on standardized tests or derivatives.
In our society we are nowadays experiencing a constant increase in stress and
anxiety levels as never seen before. Since such a phenomenon is a typically
obCopyright © 2020 for this paper by its authors. Use permitted under Creative
Commons License Attribution 4.0 International (CC BY 4.0).
served within industrial and hyper-urbanized areas, common belief relates stress
and anxiety to an overwhelming amount of work and occupation. On the
contrary stress and anxiety are a twofold part of the human mental process which
is often related to boredom [
While generally considered a detrimental habit, overthinking is often also an
unavoidably yield by circumstances and situations such as waiting for a ight at
the airport gate, hanging around before an exam, or while in an hospital waiting
room. It has been proven that the quality of the time spent in a waiting room
also a ect the perceived quality of the awaited service, especially in
healthcarerelated contests [
Many techniques has been developed to relieve anxiety and lessen stress,
among such techniques several of them focuses on the psychological impact of
the surrounding environment and its perception. Aromatherapy has been proven
to be an e ective remedy for stress and anxiety-related behaviors due to the
apparent interaction between sense of smell and pituitary gland's physiology [
In this work we devised a robotic aroma di usion device for deployments
in waiting rooms, airports, hallways and similar environments. Such a robot as
been designed to be remotely driven by means of an arti cial intelligence that
is responsible to recognize when the stress or anxiety levels are arising and the
di usion of speci c aromas could relax the bystanders. The devised robot could
easily blend with other similar device which are nowadays di used for di erent
use ranging from cleaning to healthcare applications [
Our approach provides aromatherapy in closed environments that resembles the aspect and design of a waiting room with several chairs where people sit and THERMAL IMAGE
FEATURES
VECTOR t ∑ ∑ ∑ ∑
TIMFEEA-DTEULRAEYSED
VECTOR
HIDDEN LAYERS INPUT LAYER
PATTERN LAYER
SUM LAYER
OUTPUT
LAYER
wait. It follows that we can imagine people's locations as xed points in space
and suppose that each person will remain on the same position during all his
staying. In this kind of setting we use a thermal camera to acquire images during
time. Such images are at the basis of our classi cation task to decide whether or
not aromatherapy is needed. If we can dispose of several cameras that produce
frames, then those frames can be geometrically segmented and normalized in
order to depict only one person at the time. It has been shown the existence
of strong correlations between stress and thermal e ects on several body parts
(e.g. see [
M X pikj ( ) j=1 8 k; i; From each feature vector, we can compute the variation at each time step as ik( ) = aik( ) aik( 1). Starting from such variations we can de ne a 4N 1 time-delayed features vector xk( ) as a concatenation of 4 time steps so that: xk( ) = k i ( 3)j ik( 2)j ik( 1)j ik( ) 8 k; i; 4 The vector xk( ) can be used as input u for a classi cation system since it stores values proportional to an horizontal orthogonal projection of the time-dependent variations of temperature in the k-th person body. In facts, due to the body symmetry, and the top-bottom mechanism of the physiological actions on blood (1) (2) pressure and supply from the sympathetic and parasympathetic nervous system, we are interested to the distribution of temperature along the vertical axis. In order to classify these features we can use a radial basis function neural network (RBFNN). 2.2
RBF-based classi er
For the purpose of stress level classi cation from thermal images, it is possible to
implement a dedicated system based on Radial Basis Function Neural Networks
(RBFNN). The said network has a topology similar to common four-layer Feed
Forward Neural Networks (FFNN), with a Radial Basis Function (RBF) used
for its rst hidden layer and a purely linear activation function on his its second
hidden layer. This kind of neural architecture can generate a model of the latent
features [
C(1) ! j jj ; where u is the input vector, C(1) the RBF distribution centroid for the j-th j neuron, and is a distribution shape control parameter.
The second hidden layer computes a weighted sums of values received from
the preceding neurons. This second hidden layer is called summation layer with
the output of the k-esime summation unit
x(k2) = X Wjkx(j1);
j
(3)
(4)
where Wjk represents the weight matrix elements. Such a weight matrix is
composed by a weight value for each connection from the j-esime pattern units to
the k-esime summation unit. The output layer ful ls the nonlinear mapping for
classi cation. In fact, the rst hidden layer of the RBPNN is responsible for
the fundamental task expected (see [
Robotic device In our scenario each person sit on a xed spot, therefore it is possible to assign a geometrical location L(k) to each person k in the set of people K. It follows that we can obtain a statistical distribution of the estimated level of stress both as a global distribution S( ) for each person in the room, and as many as possible local distributions Sl( ) so that
S( ) = 1 jKj
X sk( ) jKj k=1
Sl( ) =
X sk( ) 1 j lj k2 l
K = [
l l (5) where l is the set of people k that belongs to the l-th location. It follows that, given two threshold m (to identify a mild stress level) and h (to identify high stress levels) we can drive the robot to di use di erent aromas in di erent locations l of the room when Sl( ) > m or Sl( ) > h. 3
The devised decision system, based on thermal imagery and Radial Basis
Function Nerual Networks, seems promising for the development of a robotic aroma
di usion devices to be deployed in closed environments that resembles the
aspect and design of a waiting room with several chairs where people sit and wait.
Since the robot can be remotely driven by means of an arti cial intelligence it
could release aromas in the air only when and when it is needed, therefore only
when the stress or anxiety levels are arising in the room or in a smaller portion
of it. The system is prone to future improvements such as the re nement of
the classi cation process also by means of accurate psychometric studies that
could be based on well known tests (e.g. MASLAC [
This work has been supported by \Piano della Ricerca 2016/2018 - linea di intervento 2, University of Catania".