Content uploaded by Jonathan Aigrain
Author content
All content in this area was uploaded by Jonathan Aigrain on May 21, 2015
Content may be subject to copyright.
Person-specific behavioural features for automatic stress detection
Jonathan Aigrain1, S´
everine Dubuisson1, Marcin Detyniecki2,3and Mohamed Chetouani1
1Sorbonne Universit´
es, UPMC Univ Paris 06, UMR 7222, ISIR, Paris, France
2CNRS, UMR 7606, LIP6, Paris, France
3Polish Academy of Sciences, IBS, Warsaw, Poland
Abstract— This paper introduces behavioural features for
automatic stress detection, and a person-specific normalization
to enhance the performance of our system. The presented
features are all visual cues automatically extracted using video
processing and depth data. In order to collect the necessary
data, we conducted a lab study for stress elicitation using a
time constrained arithmetic mental test. Then, we propose a
set of body language features for stress detection. Experimental
results using a SVM show that our model can detect stress with
high accuracy (77%). Moreover, person specific normalization
significantly improves classification results (from 67% to 77%).
Also, the performance of each of the presented features is
discussed.
I. INTRODUCTION
One of the key elements in human-human interaction is
to understand the nonverbal cues and the social signals of
the person we are communicating with. It has been shown
that nonverbal communication conveys an important part of
the meaning of a message [8], [9], [32]. Designing machine
sensing and understanding social signals is a difficult task,
since the meaning of these signals is person and context
dependent. In this paper, we focus on the specific context of
stressful situations, and we aim at automatically detect the
presence of stress by studying the body language.
Automatic detection of stress is an emerging domain.
It faces issues related to the difficulty to obtain a nat-
ural database and to ethical problems. Current detection
techniques are based mainly on speech signals [11], [23]
and/or from physiological signals, using either traditional
sensors (ECG, skin conductance, . . . ) [1], [12], [17], [29]
or hyperspectral images [5].
In [11] and [17], stress is detected though voice and
physiological signals during a real-world driving-task. How-
ever, in this context, voice and physiological signals are
not necessarily the most suited sources in order to detect
stress. Someone who is driving does not always talk enough,
and physiological sensors are obstructive. In this case, cues
extracted from the body language are likely to be the most
convenient sources of information to assess stress since they
can be captured by camera sensors. To our knowledge,
there are few papers adressing stress detection using body
language. Giakoumis et al. [13] showed that behavioural
features enhanced the performance of automatic stress de-
tection systems using physiological features. Soury [30] used
This work was partially supported by the Labex SMART (ANR-11-
LABX-65) under French state funds managed by the ANR within the
Investissements d’Avenir programme under reference ANR-11-IDEX-0004-
02.
postural features in a multimodal fusion model, but in this
work, this modality turned out to be the least efficient.
However, the correlation between behavioural cues and
affective states has been extensively studied, mainly for au-
tomatic emotion recognition [15], [18], [20] and expression
synthesis [4], [18], [20], as well as in the psychology [8],
[9], [32]. However, though most of the affective information
studied can be used to effectively detect a large number of
emotions, there likely are some visual cues that are specific
to the expression of stress.
In this paper, we present the first results of a preliminary
study on stress detection by extracting and analyzing human
body features. This paper is organized as follows: Section
2 reports the stress concept by reviewing the different ap-
proaches in the psychology literature. Section 3 presents our
lab study for stress elicitation, in particular the protocol, the
setup and the annotation of the subjects’ stress level. Section
4 describes our feature extraction framework. Section 5
presents the person-specific normalization and its motivation.
Section 6 outlines and discusses the results obtained by our
learning approach based on SVM. We present the impact of
kernel functions, person-specific normalization and features
used. Finally, Section 7 concludes by discussing future
research directions.
II. THE STRESS CONCEPT
The term stress appeared in the medical literature in 1950s
[25]. Since then, it has been widely used [21], [25]. Holmes
and Rahe [19], when developping a tool called the Social
Readjustment Rating Scale (SRRS), described stress as a
set of events that require a fixed amount of adjustement.
In this case, stress is used as the stimulus that triggers the
adaptation, but it can also be used to describe the response
to a stimulus. Hans Selye viewed stress as the “nonspecific
response of the body to noxious stimuli” [28]. In his work,
he described a noxious stimulus as a stimulus that threatens
homeostasis, the body “preferred” set-point for physiological
variables.
One of the main problems in both Holmes’ and Selye’s
conceptions of stress is the absence of impact from indi-
vidual perception of the stimulus. Ursin et al. [31] later
demonstrated that coping behaviour reduced physiological
activity. Thus, we think that transactional models of stress
such as those presented in [22], [24] fits better our intuitive
vision of stress. These models present stress as a process
that includes the stimulus, the appraisal of the situation and
the physiological, behavioural and psychological reactions.
In addition, Koolhaas et al. [21] stated that the stimulus had
to be uncontrollable and/or unpredictable in order to define
this process as stress.
The protocol of our stress-induction experiment is mainly
based on the transactional definition of stress.
III. DATA COLLECTION
In this section, we present how we collected the data
through a stress-induction experiment.
A. Stress induction procedure
According to Dickerson and Kemeny[7], the best way to
increase a subject’s cortisol level, which is largely considered
as being correlated to psychological stress [6], [7], [25],
[30] is to make her perform a task that combines public
speaking and cognitive challenges. Indeed, tasks that con-
tain uncontrollable traits and/or social evaluation are the
most stressful ones. Thus, we designed an evaluated time-
constrained mental arithmetic test.
Fig. 1. Screenshot of the test software used for the study
Figure 1 shows a screenshot of the test. The question
asked is shown in the middle of the screen. There are two
possible answers below it. At the bottom, the progress bar
represents the time limit for the current question. On the
right is the score bar, which gives the subject a feedback of
her performance. The color of the score bar has a meaning:
green means “above average”, yellow means “average” and
red means “below average”.
The test is composed of 6 steps of increasing difficulty.
The three first steps contain 20 questions, the others contain
30 questions. The timing of each step is about 1 minute
and there is a break period of 5 seconds between them. The
subjects are told that both quickness and correctness of their
answers are taken into account for their score. It gives the
subject the impression of being evaluated, but in reality the
values the score bar shows are set in advance. It displays an
“above average” score at the beginning. Thus, the participant
finds the test easy enough and feels like she should suceed.
Then, the score drops to “average” and “below average”
levels, giving the participant the impression she is failing
the test.
B. Stress level annotation
At the end of the test, a questionnaire including stress
self-assessment is conducted for each step of the test, using
a Likert-scale (1-5). Since the quality of the annotation is
greatly dependent of subjects’ memory, they watch their own
videos of the different steps before giving a stress level.
Self-assessment has been chosen because it is the most
faithful annotation to the used definition of stress, which
gives great importance to the personal perception. In ad-
dition, annotation based on judgement of human experts
can induce a bias in the performance of some features, and
there were some cases where annotators did not achieve a
good agreement score [30]. Regarding the use of hormone
levels for annotation, [33] has shown that the difference in
dynamics between behaviour and hormone levels makes it
complex, but it will be investigated in future works.
TABLE I
STR ESS LE VEL DI STRI BUTI ON
Stress level 1 2 3 4 5
Number of videos 17 20 20 17 10
C. Experimental setup
Participants are standing at a distance of around 3 meters
from the screen where the test is displayed. For each ques-
tion, they have to clearly express to a staff member which
answer they choose among the two propositions. Video and
skeleton data were collected with a Microsoft Kinect placed
under the screen. A high definition camera located above the
screen provides an optic zoom of the particpant’s face.
D. Participants
Fourteen subjects participated in the experiment (3
women, 11 men, mean age = 24.8 ±2.8 years old). Each
subject is aware of the purpose of the experiment, but do not
know about the biased behaviour of the score bar, and gives
written consent.
E. Acquired data
For each of the 14 partipants, for each of the 6 steps, the
acquired data are:
•video of the whole body in 640 ×480 from the Kinect
•skeleton from the Kinect
•video of the face in 1440 ×1080 from the HD camera
•self-assessed stress level
The average duration of a video is 57.8 seconds ±12.6.
IV. FEATURE EXTRACTION
In this section, extracted features and procedures to extract
them are described. First, the features extracted from the
Kinect video are presented, followed by those extracted
from the video of the face. All features are computed over
the whole video for each step of the protocol. They are
summarized in Table II.
A. Body activity features
There is few research of how stress can affect our body
language. However, in order to recognize or to regenerate
someone’s affective state, there are several sets of features
extracted from the body that are usually used: the body
activity [4], [13], [15], [20], posture information such as
symmetry [13], [15], center of gravity displacements [14],
or the spatial extent [4], [15], kinematics information such
as smoothness [4], [15] and detection of specific gestures
[13], [15].
Using the skeleton and the video captured by the Kinect,
3 categories of features are extracted: the Quantity of Move-
ment (QoM) and its derivatives, the detection of periods of
high activity and posture changes and the detection of self-
touching in the region of the head.
1) Quantity of Movement: The QoM is the activity of the
body or of some parts of it. For a given frame, we compute
the QoM in two different ways. The first one (SQoM) uses
the skeleton extracted by the Kinect and is the sum of the
displacements of each joint between two frames.
QoMskel(i) = ∑
j∈joints q(vji−vji−1)2
with ithe frame index and vjithe 3D position vector of the
joint jin the ith frame. We also compute the QoM using RGB
videos, and call it IQoM. This corresponds to the number of
pixels that have changed between successive frames.
QoMvideo(i) = Card({pi|abs(pi−pi−1)>t})
with pithe RGB vector of pixel pfor the ith frame and ta
threshold. Once the QoM has been computed on all frames,
the mean value is used as the feature. In order not to be
biaised by the size of a person, it is also divided by the
surface of the bounding box containing the skeleton.
The SQoM is computed separately for the whole body
and the head (HeM). For the head only, we also compute
the SQoM following Z-axis only (HeMZ). For the body only,
the Fourier transform of the signal is calculated, divided into
10 bins (FFT1 - FFT10), which gives frequency information
about the subject’s activity.
2) Detection of high activity periods and posture changes:
Using the peaks of the QoM computed for each frame, the
periods of high activity are extracted (see Fig. 2) . The
number of periods (HAPC), their mean duration (HAPMD)
and their mean intensity (HAPMV) are used as features. For
each peak, the frame at the beginning and the one at the end
are compared to determine whether the person has changed
her posture. The number of posture changes (PCC) is used
as a feature. We make the hypothesis that posture changes
may reflect an increasing uncomfortability.
3) Detection of self-touching: It has been shown that self-
touching can be an indicator of negative affect [16]. We
detect self-touching in two parts of the body: the head and
the hands. To detect whether a person is self-touching in
the region of the head, we compute the distance between
hand, head and neck joints. If one of the distances is below a
Fig. 2. Example of detection of a posture change. The left image is the
frame at the start of the peak, the middle image is the frame at the end, and
the right one is the difference between the two others
threshold, the person is considered as self-touching the head
(see Fig. 3). The number of times we detect a single hand
as self-touching the head (STHC) and the mean duration
(STHMD) of it are used as features. We extract the sames
features in cases of two hands face-touching (ST2HC and
ST2HMD).
Fig. 3. Examples of detection of self-touching in the region of the head.
On the left is an example of self-touching with one hand. On the right is
an example with two hands
The detection of hand self-touching is more complex.
Indeed, the skeleton given by the kinect cannot be used alone
in order to detect movements such as fingers rubbing. Thus,
we compute the extraction of this feature as follows:
•Detection of the hands: because the skeleton provided
by the Kinect can be imprecise in terms of joint location,
we also use skin detection to find a more reliable
position of the hands. Thus, starting from the original
hand joint position, we look for the closest pixel where
skin has been detected. This becomes the new hand
position.
•Computation of the QoM: using the position we de-
tected, we extract the sub-image of the hands. Then,
we compute the video QoM between following hands
frames.
The QoM for each hand separately (LHM for the left hand
and RHM for the right one) and the QoM of both hands
(HM) are used as features.
B. Facial features
Using the method described in [27], we extract the ac-
tivation level of 12 Action Units (AU). Action Units are
presented by Ekman in the Facial Action Coding System
(FACS) [10], which is a common standard to systematically
categorize the physical expression of emotions. The 12 AU
extracted are described in Table II. Since the activation level
is given by frame, we compute the mean over all frames and
the standard deviation for each AU as features.
TABLE II
EXT RACTE D FEATUR ES
Feature Description
SQoM QoM computed with the skeleton
IQoM QoM computed with the RGB frames
FFTi i-th bin of the Fourier transform of the IQoM (from
1 to 10)
HAPC Number of periods of high activity
HAPMD Mean duration of periods of high activity
HAPMV Mean highest value of periods of high activity
PCC Number of posture changes
STHC Number of times self-touching with one hand in the
region of the head occured
STHMD Mean duration of self-touching with one hand in the
region of the head
ST2HC Number of times self-touching with two hands in the
region of the head occured
ST2HMD Mean duration of self-touching with two hands in
the region of the head
LHM QoM for the left hand
RHM QoM for the right hand
HM QoM for both hands
HeM QoM for the head
HeMZ QoM for the head only along Z-axis
AU1 Inner Brow Raiser
AU2 Outer Brow Raiser
AU4 Brow Lowerer
AU5 Upper Lid Raiser
AU6 Cheek Raiser
AU9 Nose Wrinkler
AU12 Lip Corner Puller
AU15 Lip Corner Depressor
AU17 Chin Raiser
AU20 Lip Stretcher
AU25 Lips Part
AU26 Jaw Drop
V. PERSON-SPECIFIC NORMALIZATION
In addition to testing the features, the impact of a person-
specific normalization is evaluated. As said in [29] regarding
physiological signals or in [2] regarding body motions, the
evolution of the signal compared to a personal baseline is
likely to be more meaningful than its absolute value. Thus,
for each person, we compare the value of each feature to its
value during the first step of the experiment, considered as
non-stressful.
˜
fp j =fp j −fp1
fp1
where pa person, jthe step the video has been extracted
from, fp j the vector of features for the person pon step jand
˜
fp j the normalized vector. Thus, the features are expressed
as an evolution percentage rather than a raw value.
VI. STRESS DETECTION
This section presents the results obtained by our model.
First, the evaluation process is described. Then, the impact of
several parameters - the SVM kernel function, the person-
specific normalization and the different features - are dis-
cussed.
A. Evaluation process
In order to test whether our features are predictive or
not, we use Support Vector Machines (SVMs) to build a
classification model. Each video is associated with a label:
stress (S) for video associated with a stress level above or
equal to 4, non-stress (NS) for the remaining ones. For each
evaluation, we use a leave-one-subject-out cross validation:
videos from all people but one person are used as the training
set, and the videos of the remaining person is used as the
testing set. This cross validation is repeated 10 times in order
to reduce the impact of randomness in the SVM parameters
optimization. The mean accuracy and its standard deviation
are presented.
We discard videos that are labelled as non-stress if they
followed a stress labelled video. These videos represent peo-
ple acknowledging that they psychologically gave up because
of the difficulty. Thus, the accuracy is computed on 76 out of
the 84 videos for raw features, 49 labelled as non-stress and
27 labelled as stress. Since the person-specific normalization
uses the features extracted from the first step, the accuracy is
computed on 62 videos for normalized features, 35 labelled
as non-stress and 27 labelled as stress.
B. Results
TABLE III
AVERA GE A ND S TAND AR D DE VI ATIO N OF T HE A CC UR ACY O BTA IN ED
WITH RAW AND PERSONNALIZED FEATURES AND WITH THREE KERNEL
FUNCTIONS.
kernel type raw normalized
Poly 5 0.64 ±0.04 0.77 ±0.02
RBF 0.65 ±0.03 0.76 ±0.02
Linear 0.67 ±0.01 0.77 ±0.01
TABLE IV
IMPACT O F S EV ER AL S ET O F FE ATUR ES W IT H AN D W IT HO UT
NO RM AL IZ ATIO N ON A CC UR ACY W IT H A LI NE A R SVM
features set raw normalized
All 0.67 ±0.01 0.77 ±0.01
Face 0.68 ±0.01 0.65 ±0.03
Body 0.63 ±0.03 0.80 ±0.01
1) Impact of the kernel function: Table III shows the
results obtained by 3 different kernel functions: the Radial
Basis Function (RBF), the polynomial of degree 5 (Poly
5) and the linear kernel. There is no significant difference
between the 3 kernels concerning the mean accuracy. Since
the linear kernel is known to be faster to train, we have
chosen to keep it for the following experiments.
2) Impact of the person-specific normalization: It clearly
appears that using the person-specific normalization signifi-
cantly improves classification results, by 13%, 11% and 10%
for the polynomial, RBF and linear kernels, respectively as
shown in Table III. However, as we can see with Table IV, the
normalization is not effective on facial features. This might
be explained by the much smaller range of values the facial
features can take compared to body features. Indeed, facial
features values vary from 0 to 5, and there is little difference
in scale among the participants. In comparaison, the IQoM
can be 30 times bigger from a subject to another.
TABLE V
AVERA GE C ON FU SI ON M ATRI X WI TH P ER S ON -SP EC IFI C FE ATUR ES F OR
10 RU NS
classified as Stress classified as Non-Stress
Stress 19.8 ±0.42 7.2 ±0.42
Non-Stress 7.1 ±0.31 27.9 ±0.31
TABLE VI
AVERA GE C ON FU SI ON M ATRI X WI TH R AW FE ATUR ES F OR 1 0 RU NS
classified as Stress classified as Non-Stress
Stress 14 ±013 ±0
Non-Stress 12 ±0.66 37 ±0.66
3) Performances of facial versus body features: Table IV
shows the impact of the choice of features on the classi-
fication accuracy. When using raw features, facial features
give better results than body features. This is probably due
to smaller inter-individual differences with facial features.
However, information on the evolution of body features
are much more discriminative than information on the evo-
lution of facial expressions. This confirms the idea that
body language can bring valuable information to determine
someone’s emotion state.
TABLE VII
FIV E BE ST A ND FI VE W OR ST R AW FE ATUR ES AC CO RD I NG TO T HE IR
CLASSIFICATION ACCURACY IF USED ALONE USING THE RBF KERN EL .
feature accuracy
PCC 0.73 ±0.02
AU9 - std 0.72 ±0.01
FFT2 0.72 ±0.01
AU4 - std 0.72 ±0.01
AU4 - mean 0.72 ±0.01
AU25 - mean 0.60 ±0.03
RHM 0.58 ±0.01
AU5 - std 0.58 ±0.02
AU9 - mean 0.56 ±0.01
HAPMD 0.55 ±0.04
4) Performances of individual features: In order to better
understand the relevance of each feature, we calculated their
classification accuracy if used alone. We use the RBF kernel
to compute these results, since a linear kernel would allow
only a single “split value” along the feature axis. The results
for raw features are presented in Table VII, those for person-
specific features in Table VIII. Concerning raw features,
we can see that several features achieve better average
classification accuracy when used alone than when combined
with others. FFT1 for person-specific features also achieve
alone a similar score to what we obtain with all normalized
TABLE VIII
FIV E BE ST A ND FI VE W OR ST P ER SO N-S PE CI FIC F EAT UR ES AC CO RD IN G
TO TH EI R CL A SS IFI CATI ON AC C UR ACY I F US ED A LO NE U SI N G TH E RBF
KE RN EL .
feature accuracy
FFT1 0.76 ±0.02
HAPC 0.74 ±0.01
FFT7 0.73 ±0.02
HAPMV 0.73 ±0.02
PCC 0.73 ±0.02
AU1 - std 0.49 ±0.03
AU25 - mean 0.46 ±0.02
AU26 - mean 0.45 ±0.02
AU15 - mean 0.45 ±0.03
AU17 - std 0.43 ±0.04
features. Thus, we can see that for our experiment, a single
feature can be enough to detect stress with good accuracy.
Among these features, it appears that brows activity as
raw features is a valuable information for our study. Indeed,
AU4 and AU9 are correlated since AU4 is often activated
when AU9 is. The number of times a subject changes its
posture also achieve good classification results, both for raw
and person-specific features. Features on body activity level
such as FFT1 or information on periods of high ativity
also achieve good accuracy as person-specific features. It is
however difficult to interpret what FFT7 really means in an
intuitive way.
Fig. 4. Distribution of FFT 1 raw and normalized values
Finally, we can see that person-specific normalization
works well for body features, but not for facial ones. Figure 4
shows how normalization on FFT 1 improved the seperation
between the two classes. The 5 best person-specific features
are body features, and the 5 worst are all facial features. This
observation can not be made for raw features, where body
and facial information are both represented in the best and
worst features.
VII. CONCLUSIONS AND FUTURE WORKS
In this paper, we have presented behavioural features
for automatic stress detection, and a person-specific nor-
malization to enhance the performance of our system. We
have discussed the stress concept and presented a set of
49 automatically extracted behavioural features. We also
observed that a person-specific normalization significantly
improved the performance of our model, and that, for our
study, the body language of the subjects provided more
valuable information than their facial expressions. Finally, we
saw that it is possible to have an unobtrusive stress detection
system using only video and depth data.
However, the results obtained can still be improved. The
results presented in Table VII and VIII clearly show that
some features do not bring valuable information and thus
should not be included training the SVM. Preliminary tests
on feature selection showed promising results for classi-
fication accuracy. Moreover, we saw that person-specific
normalization is effective only for body features. Trying to
include both raw facial and person-specific body features in
our model may give interesting results.
Finally, more subjects will participate in our experiment.
However, they will not be aware of the purpose of the study.
Thus, we will be able to observe the impact of individual
perception of a situation on stress level.
REFERENCES
[1] A. Barreto, J. Zhai, and M. Adjouadi. Non-intrusive Physiological
Monitoring for Automated Stress Detection in Human-Computer In-
teraction. In Human-Computer Interaction, pages 29–38, 2007.
[2] D. Bernhardt and P. Robinson. Detecting affect from non-stylised body
motions. In Affective Computing and Intelligent Interaction, pages 59–
70, 2007.
[3] N. Bianchi-Berthouze, P. Cairns, A. Cox, C. Jennett, and W. Kim. On
Posture as a Modality for Expressing and Recognizing Emotions. In
Emotion in HCI workshop at BCS HCI, pages 74–80, 2006.
[4] G. Caridakis, A. Raouzaiou, K. Karpouzis, and S. Kollias. Synthe-
sizing Gesture Expressivity Based on Real Sequences. ”Multimodal
Corpora. From Multimodal Behaviour Theories to Usable Models” In
: Iinternational Conference on Language Resources and Evaluation,
pages 19–23, 2006.
[5] T. Chen, P. Yuen, M. Richardson, G. Liu, Z. She, and S. Member.
Detection of Psychological Stress Using a Hyperspectral Imaging
Technique. IEEE Transactions on Affective Computing, 5(4):391–405,
2014.
[6] S. Cohen, D. Janicki-deverts, and G. E. Miller. Psychological Stress
and Disease. Journal of American Medical Association, 298(14):1685–
1687, 2007.
[7] S. S. Dickerson and M. E. Kemeny. Acute stressors and cortisol re-
sponses: a theoretical integration and synthesis of laboratory research.
Psychological bulletin, 130(3):355–91, May 2004.
[8] P. Ekman and W. V. Friesen. Hand Movements. The journal of
communication, 22(4):353–374, 1973.
[9] P. Ekman and W. V. Friesen. Detecting Deception From The Body Or
Face. Journal of personality and Social Psychology, 29(3):288–298,
1974.
[10] P. Ekman and W. V. Friesen. Facial action coding system. 1977.
[11] R. Fernandez and R. W. Picard. Modeling drivers’ speech under stress.
Speech communication, 40(1):145–159, 2003.
[12] A. Gaggioli, G. Pioggia, G. Tartarisco, G. Baldus, M. Ferro, P. Ci-
presso, S. Serino, A. Popleteev, S. Gabrielli, R. Maimone, and G. Riva.
A system for automatic detection of momentary stress in naturalistic
settings. Studies in health technology and informatics, 181:182–6, Jan.
2012.
[13] D. Giakoumis, A. Drosou, P. Cipresso, D. Tzovaras, G. Hassapis,
A. Gaggioli, and G. Riva. Using activity-related behavioural fea-
tures towards more effective automatic stress detection. PloS one,
7(9):e43571, Jan. 2012.
[14] T. Giraud, D. A. G. J´
auregui, J. Hua, B. Isableu, E. Filaire, C. L.
Scanff, and J. C. Martin. Assessing postural control for affect
recognition using video and force plates. In Humaine Association
Conference on Affective Computing and Intelligent Interaction, pages
109–115, 2013.
[15] D. Glowinski, N. Dael, a. Camurri, G. Volpe, M. Mortillaro, and
K. Scherer. Toward a Minimal Representation of Affective Gestures.
IEEE Transactions on Affective Computing, 2(2):106–118, Apr. 2011.
[16] J. Harrigan. Self-touching as an indicator of underlying affect and
language processes. Social Science and medicine, 20(11):1161–1168,
1985.
[17] J. Healey and R. Picard. Detecting Stress During Real-World Driving
Tasks Using Physiological Sensors. IEEE Transactions on Intelligent
Transportation Systems, 6(2):156–166, June 2005.
[18] J. Hoey, D. Kuli, M. Karg, A.-a. Samadani, R. Gorbet, and K. Kolja.
Body Movements for Affective Expression : A Survey of Automatic
Recognition and Generation. IEEE Transactions on Affective Comput-
ing, 4(4):341–359, 2013.
[19] T. H. Holmes and R. H. Rahe. The social readjustement rating scale.
Journal of psychosomatic research, 11(5):213–218, 1967.
[20] A. Kleinsmith and N. Bianchi-Berthouze. Affective Body Expression
Perception and Recognition: A Survey. IEEE Transactions on Affective
Computing, 4(1):15–33, Jan. 2013.
[21] J. M. Koolhaas, a. Bartolomucci, B. Buwalda, S. F. de Boer, G. Fl¨
ugge,
S. M. Korte, P. Meerlo, R. Murison, B. Olivier, P. Palanza, G. Richter-
Levin, a. Sgoifo, T. Steimer, O. Stiedl, G. van Dijk, M. W¨
ohr, and
E. Fuchs. Stress revisited: a critical evaluation of the stress concept.
Neuroscience and biobehavioral reviews, 35(5):1291–301, Apr. 2011.
[22] R. S. Lazarus. From psychological stress to the emotions: a history of
changing outlooks. Annual review of psychology, 44:1–21, Jan. 1993.
[23] I. Lefter, L. J. Rothkrantz, D. A. van Leeuwen, and P. Wiggers.
automatic stress dectection in emergency calls. International Journal
of Intelligent Defence Support Systems, 4(2):148–168, 2011.
[24] S. Levine. Developmental determinants of sensitivity and resistance
to stress. Psychoneuroendocrinology, 30(10):939–46, Nov. 2005.
[25] B. L. Lyon. Stress, Coping, and Health. In Handbook of Stress,
Coping, and Health, pages 2–20. 2012.
[26] P. Mundy, M. Sigman, J. Ungerer, and T. Sherman. Defining the social
deficits of autism: The contribution of non-verbal communication
measures. Journal of child psychology and psychiatry, 27(5):657–669,
1986.
[27] J. Nicolle, K. Bailly, and M. Chetouani. Facial Action Unit In-
tensity Predicition via Hard Multi-task Metric Learning for Kernel
Regression. ”Facial Expression Recognition and Analysis Challenge”
In : IEEE International Conference on Automatic Face and Festure
Recognition, 2015.
[28] H. Selye. The stress of life. McGraw-Hill, New York, NY, US, 1956.
[29] Y. Shi, M. H. Nguyen, P. Blitz, B. French, S. Fisk, F. D. Torre,
A. Smailagic, D. P. Siewiorek, M. Absi, E. Ertin, T. Kamarck, and
S. Kumar. Personalized Stress Detection from Physiological Measure-
ments. In International Symposium on Quality of Life Technology,
2010.
[30] M. Soury. D´
etection multimodale du stress pour la conception de
logiciels de rem´
ediation. PhD thesis, Universit´
e Paris-sud, 2014.
[31] H. Ursin, E. Baade, and S. Levine. Psychobiology of stress - a study
of coping men. Academic Press, 1978.
[32] H. G. Wallbott. Bodily expression of emotion. European journal of
social psychology, 28(1):879–896, 1998.
[33] O. Weisman, E. Delaherche, M. Rondeau, M. Chetouani, D. Cohen,
and R. Feldman. Oxytocin shapes parental motion during father–infant
interaction. Biology letters, 9(6):20130828, 2013.
[34] M. Zuckerman, B. M. DePaulo, and R. Rosenthal. Verbal and
nonverbal communication of deception. Advances in experimental
social psychology, 14:1–59, 1981.
