0% found this document useful (0 votes)
34 views6 pages

Stereotyped Gesture Recognition An Analysis Between HMM and SVM 2017

Uploaded by

nabeel hasan
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
34 views6 pages

Stereotyped Gesture Recognition An Analysis Between HMM and SVM 2017

Uploaded by

nabeel hasan
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 6

Stereotyped Gesture Recognition:

An Analysis between HMM and SVM


Marcos Y. O. Camada Jés J. F. Cerqueira Antonio Marcus N. Lima
Instituto Federal Baiano Eletrical Engineering Department Eletrical Engineering Department
IFBaiano – Campus Catu Federal University of Bahia Federal University of Campina Grande
Bahia, Brazil Bahia, Brazil Paraı́ba, Brazil
Email: [email protected] Email: [email protected] Email: [email protected]

Abstract—Stereotypic behaviours are present in both human capture their attention and encourage their social and cognitive
and nonhuman primates. Usually, these behaviours are a welfare development.
indicator. However, the stereotypic behaviours may be also a Computer softwares capable of recognizing or detecting
symptom of some mental disorder in the humans. A specific
case is Autism Spectrum Disorder (ASD). The individuals with autistic’s stereotyped gestures have been useful to reveal
ASD may exhibit stereotypic behaviours through some gestures. either the level of autism severity or their engagement with
The classic stereotyped gestures of autism are: (i) Body Rocking; therapeutic activity. Goncalves et al. [7] have proposed the use
(ii) Hand Flapping; and (iii) Top Spinning. This paper study of both the Kinect R
device and the Dynamic Time Warping
the performance between two machine learning algorithms to [DTW] algorithm to detect stereotyped gesture of the hand
recognition the stereotyped gestures typical of autism: (i) Hidden
Markov Model [HMM]; and (ii) Support Vector Machine [SVM]. flapping. Shamsuddin et al. [8] have analyzed the interaction
Sequence of orientations data from some joints obtained through between children with autism and humanoid robot aiming
a RGB-D (Red Green Blue -Depth) camera [Kinect R
] are used for the reduction of stereotyped behaviour percentage during
for analysis. The results of these two machine learning algo- the interactive process. Albanali et al. [9] have addressed
rithms are compared with state-of-the-art. The HMM approach recognition of stereotyped gesture type Body Rocking and
proposed in this paper have shown 98.89% average recognition
rate and 98.9% recall. This value is higher compared to the SVM Hand Flapping using wireless accelerometer sensors and used
approach and the others of art method presented. decision tree algorithm to classify it. They have compared
Index Terms—Stereotypic Behaviours, HMM, SVM, autism. also results from both laboratory and classroom environment.
Rad et al. [10] have used also accelerometer on limbs
I. I NTRODUCTION of autistic, but they applied Convolutional Neural Network
[CNN] to classify the stereotyped gestures [body rocking and
Stereotypic behaviours are abnormal gesture patterns of hand flapping].
body that have no obvious function [1]. These behaviours may In this paper, it is presented a study about performance
be present in both humans and nonhuman primates [2]. There of machine learning algorithms for recognition of stereotyped
are different reasons for stereotypic behaviours. In general, gestures typical of autistic (hand flapping, body rocking and
they may be related to the conditions of poor welfare due top spinning). Sequence of orientations data from some joints
to either frustration, stress or fear experiences [1], [3]. In obtained through a RGB-D [Red Green Blue - Depth] camera
particular for humans, the stereotypic behaviours may be also [Kinect R
] are used for analysis. The machine learning chosen
related to neural development disorders as the autism. are: (i) Hidden Markov Model [HMM]; and (ii) Support Vector
Medical researches about autism have pointed out that the Machine [SVM].
stereotypic behaviours may be related to either defense [3] or This paper is organized as follows. In Section II, the autism
self-stimulation [4] mechanism due to their hypersensitivity to and the machine learning algorithms [HMM and SVM] are
the environment. These behaviours increase the activation level introduce. The system architecture is Summarized in Section
of the autistic. An autistic can present some classic stereotyped III. The experiments with the machine learning algorithms and
gestures, such as: (i) Body Rocking; (ii) Hand Flapping; (iii) their results are discussed in Section IV. Finally, discussions
Top Spinning and (iv) Head-Banging [3]. about the contributions of this paper and future works are
Stereotypic behaviours may interfere directly with learn- addressed in Section V.
ing and socialization of children with autism. In general,
traditional treatments rely on techniques of positive/negative II. BACKGROUND
reinforcement to mitigate these stereotypies [4]. Recently, Autism Spectrum Disorder [ASD] is a neuropsychiatric
both computer softwares [5] and robots [6] have been disorder characterized by severe damage for the socialization
explored for therapeutic purposes and to helping the autistic. and communication processes. An autistic may also have either
These technologies are part of intervention strategies to both a unusual behaviour pattern or stereotypic behaviours [11].
978-1-5090-5795-5/17/$31.00
c 2017 IEEE Novel researches have been indicated that several factors can
be associated with the autism. Some of them are genetic,
neurological anomalies and psychosocial risks [11].
Stereotypic behaviours in autistic may be related to some
defense mechanism due to their hypersensitivity [3]. This type
of behaviour increases the level activation of autistic [1]. The
stereotyped gestures may occur independently or in pairs.
Here, the gestures are considered apart in the experiments.
Some stereotyped gestures that can be noted are: (i) Body
Rocking [BR] – repetitive movement to forward and backward
of the upper torso; (ii) Top Spinning [TS] – walk in a circle;
(iii) Hand Flapping [HF] – swing motion of the hands up and
down; (iv) Head Banging [HB] – hitting head on the floor
or wall. The Head Banging was not considered in this papers
specially because the trajectory of their movement is similar
to Body Rocking.
A. Hidden Markov Model
Hidden Markov Models (HMM) are stochastic models
which present two types of chains: (i) a underlying Markov
chain [invisible]; and (ii) an stochastic states symbols [visible]. Fig. 1. Stereotyped gestures typical of autism obtained by Kinect
R
camera:
Body Rocking (A.1 and A.2), Hand Flapping (B.1 and B.2) and Top Spinning
Since the symbol output probability distribution of a contin- (C.1 and C.2).
uous HMM is given by a mixture of Gaussian, a HMM can
be expressed as λ = (A, c, µ, U ), where: A is a matrix of
transition probabilities; c is a set of coefficients [weights for as infrared [IR] emitter, IR depth receptor, color sensor,
each Gaussian]; µ represents the averages of each Gaussian microphone array, and tilt motor. The first three resources were
in the mixture; and U represents the covariance matrix of the used here to get frames RGB [Red, Green and Blue] with depth
Gaussian [12]. information [RGB-D] about a scenery and elements contained
The HMM can be applied for supervised learning pattern into it [people and objects].
recognition tasks. The training process of the HMM consists of In Figure 1, it is shown some RGB-D frames of the stereo-
the presentation of sequences of outputs [training sequences] typed gestures: Body Rocking [A.1 – A.2)], Hand Flapping
from a particular system. [B.1 – B.2], and Top Spinning [C.1 – C.2)].
A training algorithm can adjust the HMM’s parameters such In order to extract from these frames, the information about
that when a new observation sequence [from the system being joints of one subject [actor] the OpenNI/NITE package was
modeled] is given as input to the HMM, the resolving of output used. The orientation of joints is quaternion notation and after
is based on the probability of the model generated. it is converted to euler notation [right side of (1), where T
B. Support Vector Machines (defined in [14, eq.(6.3)]) is a transformation matrix]. In this
paper, the joints considered are head, shoulders, elbows, trunk,
Support Vector Machines [SVMs] are based on principles
hips, and knees. Hence, ten joints were used [j = 10].
of statistical learning theory and of convex optimization, such
that two-class can be classified from a set of training examples.  1
c12 c13 c14
 1
y1 z1
 
Hence, the algorithm aims to find a suitable boundary into a c1 x
c21 c22 c23 c24  x2 y2 z2
data space and to separate two classes of elements [13].
 .. .. .. ..  T =  .. (1)
   
.. .. 
Extensions of SVM have include methods for regression, . . . . . . .
clustering, factor analysis. The use of linear SVM for clas- cj1 cj2 cj3 c4j
xj yj zj
sification is Computationally simpler and to search a ‘good’
[based on some measure of performance] separating hyper- IV. E XPERIMENTS AND R ESULTS
plane in a high-dimensional feature space. The optimal hy-
perplane is the one with the maximal margin of separation The Waikato Environment for Knowledge Analysis
between the two classes. Nonlinear SVMs are able to represent [WEKA] was used as a support tool for evaluate the per-
more complex functions to deal with real-world data. The formance of different machine learning algorithms from joint
SVMs have been used in various domains of application such orientation data [15]. The WEKA provides different interfaces
as, text categorization, handwriting recognition, face detection, for the users to test and evaluate machine learning algorithms
and bioinformatics [13]. available by itself or third-party. It allows also to build
workbench for the main data mining problems: regression,
III. S YSTEM OVERVIEW classification, clustering, association rule mining, and attribute

R
A Kinect camera is used to capture frames with the selection. The files for test and training are usually either
stereotyped gestures. This device has a set of resources such Attribute-Relation File Format [ARFF] or Comma-Separated
Values [CSV]. The performance of machine learning algo-
rithms to recognize stereotyped gestures typically autistic are
compared through experiments.
A. Methodology
Several RGB-D frames of each scenario of the stereotyped
gestures were recorded repeatedly. The joint [orientation] data
of an actor were extracted using the OpenNI/NITE frameworks
and stored with their corresponding RGB-D frames. After (a)
that, the samples were manually extract. For each stereotyped
gesture, two scenarios were simulated: high and low activation.
Each scenario had 120 records with varying amounts of joint
samples and also varying amounts of joint samples represented
by (2), where eji is euler notation from (1), j is index of a
joint and n is size of a sample. Therefore, each stereotyped
gesture has 240 records. A stereotyped gesture is recognized
regardless of the activation level.
e1 e21 · · · ej1
 1 
(b)
 e1 e2 · · · ej  Fig. 2. Result of radial–SVM experiments, CCI (2a) and recall (2b), for each
 2 2 2
. .. .. ..  (2) stereotyped gestures BR, HF and TS
 .. . . .
e1n e2n ··· ejn TABLE I
C ONFIGURATIONS OF SVM FOR S TEREOTYPED G ESTURES WITH
The 10-folds cross validation technique was used. In this G AMMA /D EGREE T YPE K ERNEL AND M AXIMUM CCI/R ECALL
one, the training is randomly split into 10 sets, such that
nine sets are used in training and the reaming set is used for Stereotyped Kernel Gamma(G)/ Max
validations then another nine sets is picked and so forth. Gesture Type Degree(D) CCI/Recall
In this paper, for each configuration of machine learning BR Radial G=1 94.52%/94.5%
algorithms [Diagonal–HMM, Spherical–HMM, Radial–SVM, HF Polynomial G=1;D=2 93.11%/93.1%
and Polynomial–SVM] is applied 25 experiments. In order to TS Radial G=12 100%
compare the performance of the machine learning, it uses two General Radial G=1 94.52%/94.5%
relevant statistical qualifiers: Correctly Classified Instances
[CCI], and recall (measures the classifiers completeness).
in variation of gamma (Figure 3c and 3d). For TS, the high
B. Experiments with SVM
recognition rate (96.84% CCI and 96.8% recall) is 6 gamma
The experiments with SVM have had 1437 instances. Each (and 2 degree).
instance is a row of (2). The definition these experiments was The experiment results of SVM with the best performance
organized in relation to two kernel type: (i) radial [depends on for each stereotyped gesture are summarized in Table I. This
gamma variable]; and (ii) polynomial [depends on degree and table categorizes these results in relation to stereotyped ges-
gamma variables]. ture, kernel type, (gamma/degree), and maximum CCI/recall.
The results of SVM experiments using radial kernel are
observed in Figure 2. The high recognition rate for BR In the SVM experiments, the general and BR had the best
[94.52% CCI and 94.5% recall] occurs with gamma equal 1. recognition rate with radial type kernel [94.52% CCI and
The high recognition rate for HF [85.34% CCI and 85.3% 94.5% recall], and 1 gamma. The HF had high recognition rate
recall] occurs with 1 gamma. The best recognition rate for TS [93.11% CCI and 93.1% recall] with polynomial [1 gamma
[100% CCI] occurs with value of the gamma from 12 to 24. and 2 degree] type kernel. The TS had the best performance
The highest average recognition rate for SVM radial [94.52% for recognition rate [100%] with radial type kernel and 12
CCI and 94.5% recall] has gamma equal 1. gamma.
In order to find the degree of polynomial SVM with the best
CCI, the gamma value was set to 1 (Fig. 3a and 3b). After C. Experiments with HMM
that, it is fixed the value of degree with the value defined in the Experiments were accomplished with HMM of the type
previous step. Thus, the value of gamma is varied to determine right-left and 0.01 interaction cut off. In these experiments,
its best value (Fig. 3c and 3d). 720 instances have been used, which were several sequences
In the Figure 3 (3a and 3b), it can observe that the of joint orientation extracted manually from eq. (2). The
best general performance (93.11% CCI and 93.1% recall) is performance of HMM was analyzed for each gesture over
achieved with 2 degree (and 1 gamma). These values of degree number of HMM states and covariance type [diagonal and
and gamma are also the best values for general, BR and HF spherical matrices].
(a) (a)

(b)
(b)

(c)

(c)

(d)
Fig. 3. Result of polynomial–SVM experiments varying in the degree,
CCI (3a) and recall (3b), and in the gamma, CCI (3c) and recall (3d), for
(d)
each stereotyped gestures BR, HF and TS.
Fig. 4. Results of diagonal–HMM experiments, CCI (4a) and recall (4b), and
spherical covariance (CCI (4c) and recall (4d)) types in relation to number of
states for each stereotyped gestures BR, HF and TS.
Results of experiments (CCI and recall) of HMMs with
diagonal and spherical type covariance varying the amount
of the states are shown in Figure 4. 15 to 17 states, 19 and 20 states, and from 24 to 26 states. For
In experiments with diagonal covariance HMM (see Fig- HF, 12 and 18 states had the highest average recognition rate
ure 4a and 4b), the BR had 100% recognition rate (both CCI [95.83%CCI and 95.8% recall]. The TS had 100% recognition
and recall) with 12 and 13, from 15 to 20, and from 22 to 26 rate except for 2, 4 and 5 states. The best general average
states. The highest recognition rate for HF [96.67% CCI and recognition rate [98.61% CCI and 9.6% recall] was with 16
96.7% recall] occurred with 20 states and TS [both CCI and states.
recall were 100%] except for 3 states. The highest average The configurations of HMM for each stereotyped gesture
recognition rate for HMM diagonal type covariance [98.89%] with both the minimum number of states and the maximum
had 20 states. CCI/recall are summarized in Table II. Thus, the best con-
Experiments with spherical covariance HMM present CCI figuration for BR is HMM diagonal type covariance with 12
(Fig. 4c) and recall (Fig. 4d) results. The highest recognition states that reaches 100% recognition rate. The HF has the
rate for BR [both CCI and recall with 100%] occurred from highest recognition rate [96.67% CCI and 96.7% recall] with
TABLE II A comparative study about performance between RGB-D
C ONFIGURATIONS OF HMM FOR S TEREOTYPED G ESTURES WITH [Kinect R
] and accelerometer to detect stereotypical motor
M AXIMUM CCI/R ECALL AND M INIMUM N UMBER OF S TATES
movements [hand flapping] in real time was proposed in
Stereotyped Covariance Min. Number Max. [16] with two approaches. In their first approach, the authors
Gesture Type of States CCI/Recall have used Dynamic Time Warping [DTW] to recognize the
BR Diagonal 12 100% gesture through joint coordinates data. In the second one,
HF Diagonal 20 96.67%/96.7% they have used a statistical methods [standard variation, root
TS Diagonal 2 100% mean square and peaks and valleys]. Four autistic children are
General Diagonal 20 98.89%/98.9% used in both approaches. The accuracy of the first and second
approaches were respectively 51% and 76%. An extension
TABLE III
of this previous works was also presented [7]. In this one,
P ERFORMANCE FOR R ECOGNITION OF S TEREOTYPED G ESTURES AMONG the work have aimed to determining ground truth of the first
M ACHINE L EARNING . approach using Kinect R
with regard to video analysis. The
average performance is 65% ([disregarding the experiments
Stereotyped HMM SVM
with false-positive values].
Gesture
An approach using Convolutional Neural Network [CNN]
BR 100% CCI 94.52% CCI
and 94.5% recall
was proposed in [10]. In this one, data from accelerometer
HF 96.67% and 96.7% recall 93.11% CCI
were used to classify stereotyped gesture [BR and TS] of
93.1% recall
6 autistic. Two approaches were proposed: (i) to use cross
TS 100% CCI 100% CCI
trained dataset [transferred]; and (ii) to use only trained
General 98.89% CCI and 98.9% recall 94.52% CCI
dataset. The accuracy of first and second approach were
94.5% recall
respectively 78% and 74%.
In this present paper, from joint orientation data obtained
using the Kinect R
device, HMM have presented better result
the diagonal–HMM with 20 states. The diagonal–HMM with for stereotyped gestures recognition [BR, HF, TS] than SVM
2 states reaches the best recognition rate [100%] for TS. [see Table III]. However, the others papers considered in Table
The general average recognition rate [98.89% CCI and 98.9% IV took in account only two stereotyped gestures: BR and HF.
recall] is with the HMM spherical type covariance with 16 Hence, only these two gestures were considered in the analysis
states. between the approaches of this paper [HMM and SVM] and
D. Discussions the approaches from others papers presented in Table IV.
HMM reaches high recognition rate 100% and 96.67%
The best results for recognition of the stereotyped gestures
for BR and HF, respectively [see Table III]. The average
[BR, HF, and TS] using machine learning algorithms [HMM
recognition rate with these gesture is 98.89%. SVM reaches
and SVM)] are summarized in Table III. The HMM have
also high recognition rates for BR and HF gestures [94.52%
presented better or equal (for TS) results for all stereotyped
and 93.11%, see Table III], and the average recognition rate
gestures than the SVM. The highest difference between them
is 94.52%. The performance of the approach using HMM
have occurred with BR [5.48% CCI]. The general performance
presented here is greater than those presented in the other
of HMM [98.89% CCI and 98.9% recall] was also higher
papers in Table IV.
than SVM [94.52% CCI and 94.5% recall], with 4.37% CCI
difference. The use of accelerometer sensors in autistic can be consid-
Currently, the works related to recognition of stereotyped ered as the disadvantage of [9], [16], and [10]. The use of de-
gestures for autism are still rare and, Table IV summarize vices embedded in the body may be intrusive or inconvenient
some these papers comparing all of them the target stereotyped in hypersensitive autistic. Hence, a less intrusive approach [e.g.
gestures, type of sensor, machine learning algorithm used, the camera or infrared sensors] may be recommend for most cases
body feature considered and the format for data generation of ASD (Autism Spectrum Disorder).
[actors or non-actors]. However, the approach presents some restrictions. The use
A study on the performance of stereotyped gesture [body of single non-autistic subject to test and validate the machine
rocking, hand flapping] recognition into different environments learning algorithms from simulated gestures has two disad-
[between classroom and laboratory] was presented by [9]. vantage: (i) subtle gestural features of the autistic may not
In this paper, accelerometers were used in wrists and torso be accomplished or perceived; and (ii) the influence of both
of 6 autistic children, jointly with Decision Tree for classi- the idiosyncrasy and the artistic quality of actor become more
fication. The final results were not organized by stereotyped relevant. Anyhow, the use of actors can be justified because
gesture, but by environment. The performance for recognition allows to obtaining a large amount of samples for training and
of each stereotyped gesture class was not determined. The testing, which in the real conditions of a medical environment
average performance stereotyped gesture recognition for lab- would be difficult. In addition, invariant body information
oratory [89.5%] was better than classroom [88.6%]. [orientation joints] were used for all experiments.
TABLE IV
AUTOMATIC S TEREOTYPED G ESTURE R ECOGNITION S YSTEMS . ML = M ACHINE L EARNING , (L)L ABORATORY AND (C)C LASSROOM E NVIRONMENT.
ACC = ACCELERATION , COORD = C OORDINATES , DTW = DYNAMIC T IME WARPING , CNN = C ONVOLUTIONAL N EURAL N ETWORK , N =
N ON - TRANSFERRED - CNN , T = T RANSFERRED - CNN

References Stereotyped Sensor ML Body Actors Accuracy


Gesture Algorithm Feature (%)
[9] BR, HF accelerometer Decision Tree wrists and 6 non-actors 89.5(L),
torso acc. 88.6(C)
[16] HF RGB-D DTW joint coord. 4 non-actors 51
accelerometer Statistical wrists and 76
torso acc.
[7] HF RGB-D DTW joint coord. 5 non-actors 65
[10] BR, HF accelerometer CNN wrists and 6 non-actors 74(N)
torso acc. 78(T)

V. C ONCLUSIONS AND F UTURE WORKS [6] M. A. Goodrich, M. Colton, B. Brinton, M. Fujiki, J. A. Atherton,
L. Robinson, D. Ricks, M. H. Maxfield, and A. Acerson, “Incorporating
In this paper, a study about the performance of some a robot into an autism therapy team,” IEEE Intelligent Systems, vol. 27,
machine learning algorithms [HMM and SVM] applied to no. 2, pp. 52–59, mar 2012.
[7] N. Gonçalves, S. Costa, J. Rodrigues, and F. Soares, “Detection of
the recognition of stereotyped gestures typical of autism from stereotyped hand flapping movements in autistic children using the
joint orientation data was presented. The approach using HMM kinect sensor: A case study,” in IEEE International Conference on
reaches a general average recognition rate of 98.89% CCI and Autonomous Robot Systems and Competitions, ICARSC 2014, 2014, pp.
212–216.
98.9% recall and recognition rate of 100% and 96.67% for [8] S. Shamsuddin, H. Yussof, S. Mohamed, F. A. Hanapiah, and L. I.
BR and HF, respectively. For TS, the HMM has recognition Ismail, “Stereotyped behavior of autistic children with lower iq level
rate 100%. These values are higher than all others approaches in hri with a humanoid robot,” in Proceedings of IEEE Workshop on
Advanced Robotics and its Social Impacts, ARSO, 2013, pp. 175–180.
either presented or cited here. The high values of CCI and [9] F. Albinali, M. S. Goodwin, and S. S. Intille, “Recognizing stereotypical
recall presented by classifiers studied in this paper (HMM and motor movements in the laboratory and classroom,” in Proceedings of
SVM) reveal a satisfactory performance from the considered the 11th international conference on Ubiquitous computing - Ubicomp
’09. New York, New York, USA: ACM Press, 2009, p. 71.
samples. [10] N. M. Rad, A. Bizzego, S. M. Kia, G. Jurman, P. Venuti, and
Of course the idiosyncrasy of a single non-autistic subject C. Furlanello, “Convolutional neural network for stereotypical motor
[actor] to obtaining of samples for training and testing may movement detection in autism,” in 5th NIPS Workshop on Machine
Learning and Interpretation in Neuroimaging (MLINI), 2015, pp. 1–9.
have influenced the results. However, the use of actor is accept- [11] S. E. Levy, D. S. Mandell, and R. T. Schultz, “Autism,” The Lancet,
able in the prototype phase of some projects. For instance, this vol. 374, no. 9701, pp. 1627–1638, nov 2009.
allowed to collecting a greater amount of stereotyped gesture [12] L. R. Rabiner and B. H. Juang, “An introduction to hidden markov
models,” IEEE ASSP Magazine, vol. 3, no. January, pp. 4–16, 1986.
samples [with invariant feature] than an artistic. [13] A. Mammone, M. Turchi, and N. Cristianini, “Support vector machines,”
This research is one part of HiBot Project [17], [18]. Wiley Interdisciplinary Reviews: Computational Statistics, vol. 1, no. 3,
HiBot is a robotic platform for experiments of Human-Robot pp. 283–289, 2009.
[14] K. Großekatthöfer and Z. Yoon, “Introduction into quaternions for
Interaction [HRI] in development at Robotic Laboratory of spacecraft attitude representation,” TU Berlin, pp. 1–16, 2012.
Federal University of Bahia. This research is a prototype for a [15] E. Frank, M. A. Hall, and I. H. Witten, “The weka workbench.
HiBot’s interface. In the context of the HiBot Project, the robot online appendix for ”data mining: Practical machine learning tools and
techniques”,” Morgan Kaufmann, Fourth Edition, 2016.
will acts as an assistive device to aid the treatment of autistic [16] N. Gonçalves, J. Rodrigues, S. Costa, and F. Soares, “Automatic detec-
patients as well as others types of interaction with humans. Of tion of stereotyped hand flapping movements: Two different approaches,”
course the ground truth needs to be improved for better results in Proceedings - IEEE International Workshop on Robot and Human
Interactive Communication, 2012, pp. 392–397.
in the real world. Performances from either different actors or [17] C. M. Salcedo, C. A. Peña, J. J. F. Cerqueira, and A. M. N. Lima,
autistic can be used to minimize the influence of idiosyncrasy. “Designing a real time artificial vision system for human interaction
with an omnidirectional mobile platform,” in 2012 Brazilian Robotics
R EFERENCES Symposium and Latin American Robotics Symposium, Oct 2012, pp.
21–26.
[1] G. J. Mason, “Stereotypies: a critical review,” Animal Behaviour, vol. 41, [18] M. Y. O. Camada, D. Stéfano, J. J. F. Cerqueira, A. M. N. Lima,
no. 6, pp. 1015–1037, jun 1991. A. G. S. Conceição, and da A. C. P. L. Costa, “Recognition of affective
[2] C. K. Lutz, “Stereotypic behavior in nonhuman primates as a model for state for austist from stereotyped gestures,” in Proceedings of the 13th
the human condition,” ILAR Journal, vol. 55, no. 2, pp. 284–296, 2014. International Conference on Informatics in Control, Automation and
[3] J. Kuhn, “Stereotypic behavior as a defense mechanism in autism,” Robotics, vol. 1, no. Icinco, 2016, pp. 197–204.
Harvard Brain Special, vol. 6, no. 1, pp. 11–15, 1999.
[4] A. B. Cunningham and L. Schreibman, “Stereotypy in autism: The
importance of function,” Research in Autism Spectrum Disorders, vol. 2,
no. 3, pp. 469–479, jul 2008.
[5] S. Parsons, P. Mitchell, and A. Leonard, “The use and understanding of
virtual environments by adolescents with autistic spectrum disorders,”
Journal of Autism and Developmental Disorders, vol. 34, no. 4, pp.
449–466, aug 2004.

You might also like