helpfn

Emotion Detection and Recognition using HRV Features Derived from Photoplethysmogram Signals

Raj Rakshit TCS Research & Innovation, India India [email protected]

V Ramu Reddy TCS Research & Innovation India [email protected]

Parijat Deshpande TCS Research & Innovation India [email protected]

ABSTRACT Detection of true human emotions has attracted a lot of interest in the recent years. The applications range from e- retail to health-care for developing effective companion sys- tems with reliable emotion recognition. This paper proposes heart rate variability (HRV) features extracted from photo- plethysmogram (PPG) signal obtained from a cost-effective PPG device such as Pulse Oximeter for detecting and recog- nizing the emotions on the basis of the physiological signals. The HRV features obtained from both time and frequency domain are used as features for classification of emotions. These features are extracted from the entire PPG signal ob- tained during emotion elicitation and baseline neutral phase. For analyzing emotion recognition, using the proposed HRV features, standard video stimuli are used. We have consid- ered three emotions namely, happy, sad and neutral or null emotions. Support vector machines are used for develop- ing the models and features are explored to achieve average emotion recognition of 83.8% for the above model and listed features.

Keywords HRV features; Emotion detection; Emotion recognition; Video stimuli, Physiological signal; PPG; SVM

1. INTRODUCTION In the past few years, automatic recognition of human

emotions has been extensively carried out using knowledge from various fields like biomedical engineering, psycho-physiology, computer science, and artificial intelligence. Emotion recog- nition is very useful in number of scenarios, for instance, forensic scenarios using audio and/or video, medical applica- tions such as determination of level of pain when the person is unconscious or unable to convey pain, retail applications such as to identify whether an individual is really interested in buying a certain item, scholastic applications such as per- son specific design of syllabus, based on subjective analysis of cognitive load, stress levels etc. Till date there are many

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full ci- tation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy other- wise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected].

ERM4CT’16, November 16 2016, Tokyo, Japan c© 2016 ACM. ISBN 978-1-4503-4558-3/16/11. . . $15.00

DOI: http://dx.doi.org/10.1145/3009960.3009962

emotion recognition systems that are developed across dif- ferent domains like speech [11] [18] [17], facial expression [28], body movements [30], textual [8] etc. The main draw- back of these systems is that they can be easily faked by the subject’s intentional gestures, maneuver and/or expressions. In recent times focus has been on physiological signal based emotion recognition system which are involuntary in nature and therefore researchers have extensively used the signals obtained from electrocardiogram (ECG) for analyzing the emotions [9] [16] [2] [13]. Though ECG is the gold standard for obtaining heart-rate and heart rate variability, using such system in real time scenario for recognizing emotions is not feasible and moreover is a very costly equipment. This has motivated our work and we have conducted the experiments using a cost-effective fingertip pulse oximeter device which is medically approved [15]. The PPG signal can even be derived from wearable such as watch, mobile phone camera etc and prove to be a useful tool for further value add. We have extracted features from the PPG signal obtained from pulse oximeter device during emotion elicitation phase. We have used standard video stimuli to discriminate the positive and negative emotions like happy and sad emotion and for neutral emotion we have used black screen with white plus (+) sign in the middle of the screen with a soft background music [10]. We have proposed heart rate variability (HRV) features extracted from the time and frequency domains for automatic classification of the emotions using support vector machines (SVM).

Rest of the paper is organized as follows: In section 2 re- lated work on emotion classification based on physiological signals is presented. Different blocks involved in the pro- posed methodology such as data collection, pre-processing, time and frequency domain HRV features, SVM based clas- sification are explained in section 3. Results of the proposed system are given in section 4. Finally, summary and future scope of the work is laid out in section 5.

2. STATE OF THE ART Researchers in the field of emotion recognition domain

generally use some form of elicitation trigger (audio, video, pictures, games etc.), record the subject’s physiological sig- nals, extract important features and classify these data using appropriate classifiers and report accuracy. The pioneering work in this domain was done by Picard [14] using personal imagery as an elicitation trigger. Since then several stan- dard elicitation databases have been proposed such as IAPS, IADS, DEAP, LIRIS-ACCEDE [13] etc.

Kim [9] and Lisetti [12] used audio, video, cognitive task,

movie clip and tough math questions to detect emotions such as sadness, anger, stress, surprise, fear, frustration and amusement whereas Rani [16] and Balienson [2] used com- puter based cognitive task and video to trigger emotions and used ECG, PPG and face tracking techniques to ex- tract affective features and classify emotions like frustration, anger, engagement, boredom, anxiety, amusement, sadness and neutral. Recently Sung-Nien [29], Valderas [25], Nardelli [13] and Zhang [31] have used visual-auditory, videos, IADS and IAPS respectively to extract HRV related features from ECG and other psychobiological signals for emotion classifi- cation such as neutral, happiness, stress, sadness, relax joy and fear.

Our novelty lies in using only one cost-effective physiologi- cal sensor and proposing a set of features which are sufficient to classify the listed emotions.

3. EMOTION DETECTION AND RECOG- NITION SYSTEM (EDRS)

Fingertip pulse oximeter data (sampling rate 60 Hz) from 33 healthy subjects (13 Females, 20 Males) with average age 27, ranging from 24 to 33 was collected. Only one emotion elicitation video was shown to a subject on a single day to prevent unwanted biases.

Figure 1: Architecture of Emotion Detection and Recognition System (EDRS)

The basic architecture of the proposed emotion detection and recognition system (EDRS) developed using SVM is shown in Figure 1. The system contains different blocks such as i) Data collection block ii) pre-processing block iii) Feature Extraction block iv) SVM model block. The func- tionality of each block is explained briefly below.

3.1 Data Collection A Pulse Oximeter is used to detect and record the PPG

signal. A photoplethysmogram (PPG) signal is an optically obtained plethysmogram, a volumetric measurement of an organ. A PPG is often obtained by using a fingertip pulse- oximeter which illuminates the skin and measures changes in

light absorption [5]. These data are collected and processed as discussed later.

During the data collection two computers work in paral- lel (Com1 and Com2, 4 GB primary memory, 64 bit Win 7 operating system) which are time-synced using Network Time Protocol. A web-cam (1080p 30 fps, placed on top of the Com1, where the subject is made to sit) is attached to Com2, see Figure 2. The subject cannot see his/her facial expressions during the experiment. This rigorous experi- mentation procedure and meticulous data-collection ensures that the time synchronization error between the stimuli and recorded physiological data is always less than 1 sec.

Figure 2: Experimental set-up

The video is played as well as the PPG signal is recorded in Com1. Using a video editor a digital clock gadget is placed on the web-cam recording window to keep track of the synced system-time of the collected physiological data (Com1) and subject’s facial expression (Com2). Using the video editor a time synced video is created to be used for data analysis which consists of the video stimuli, the facial expression of the subject and recorded PPG signals.

3.1.1 Validation of affective data We validated the collected data by generating ground truth

emotion. A ground truth emotion is “an emotion that is perceived by people, agreed upon by most receivers and the same time, the experienced true emotion as felt by the per- son her/himself ” [24]. Ground truth generation in this study is a two fold process. The first one being video stimuli [6] we used in our experiment while the other one is questionnaire analysis. Video Stimuli: Jonathan [19] and very recently Smitha et al. [23] found that film as an emotion elicitation method requires more cognitive participation and can induce strong emotional feeling and Lisetti [12] determined the most likely time slots of intended emotion by film plots. Hazer et al. [7] studied the influence of age on film-clip choice, the correla- tion between age and valence/arousal rating for the chosen clips and the differences in valence and arousal ratings in the different age groups. Gross and Levenson [6] suggested two best films among the set of gathered films which strongly elicit only one target emotion, after evaluation of over 250 films and experimenting them on 494 English speaking sub- jects. We used two movie clips as suggested in [6] for happy (Film: When Harry Met Sally, Scene: Discussion of orgasm

in cafe) and sad (Film: The Champ, Scene: Boy cries at father’s death) emotion elicitation. A black screen window with a white plus (+) sign in the middle of the screen with soft background music was used to get the neutral emotion. All the stimuli are 75 sec long. These clips are selected as they are agnostic to various demographies and elicit the same emotions universally. Online Questionnaire: We have created an online ques- tionnaire (to ensure anonymity and privacy) to collect sub- jects feedback before and after they take part in the experi- ment. Before the data collection is started we make the sub- jects take the pre-experiment questionnaire that records the subject’s age, gender, the frequency at which he/she watches movies and their genres, language, places they watch them (eg theatre, Computer, TV etc), how watching movies affect them emotionally in general on a scale of 1 to 5 and most importantly their emotional state at that point of time. Af- ter the data collection is over the subject is asked to fill up another online questionnaire that asks whether they have seen the clip ever before in their life and on a scale of 0 to 4 they are asked to mark a set of emotions that the clip they just saw made them feel, where 0 meant they did not feel even the slightest bit of the emotion and 4 meant the most they have ever felt in their life. If the answer to the first question is yes then his/her data is discarded from the study to avoid unnecessary biases.

3.2 Pre-processing The collected PPG data is fed as an input to the pre-

processing block to obtain accurate RR intervals as an out- put. This is achieved by following the sequence of steps like baseline detection, baseline-removal and peak detection. The signals obtained across each stage are also depicted in Figure 3, where Figure 3(a) represents raw PPG signal ob- tained during emotion elicitation and Figure 3(b) is the cor- responding baseline of the recorded PPG signal which is ba- sically the trace of the troughs of the PPG signal, which oc- curs mainly due to breathing and unintentional hand move- ments of the subject. Figure 3(c) is the signal obtained after beseline removal and 3(d) shows the detected peaks in the PPG signal to extract RR intervals which is shown in Figure 3(e). The RR intervals are the time difference between sub- sequent peaks of the processed PPG signal which is shown in Figure 4.

3.3 Feature extraction We have explored heart rate variability (HRV) features

which are extracted from RR intervals which are in turn ob- tained from raw PPG waveforms to discriminate the emo- tions. In time domain, 10 features such as meanRR, me- dianRR, SDNN, SDANN, pNN50, NN50, RMSSD, SDNNi, meanHR, stdHR are computed in each measurement cycle (See Table 1 for explanations of the abbreviations. In the frequency domain analysis, Welch’s algorithm is carried out for spectral analysis. Welch algorithm estimates the power spectrum by using an averaging modified periodogram [27]. The power spectrum of the HRV signal is divided into 3 bands, namely, very low frequency-VLF (0-0.04 Hz), low frequency-LF (0.04-0.15 Hz) and high frequency-HF (0.15- 0.5 Hz) where LF and HF bands are related to sympa- thetic and parasympathetic activities. We have extracted 13 features from these frequency bands, namely, peakVLF, peakLF, peakHF, aVLF, aLF, aHF, aTotal, pVLF, pLF,

5 10 15 20 25 30 35 40

20 40

60 80

R a

w P

P G

0 5 10 15 20 25 30 35 40 −30

−20

−10

B a

se lin

e

5 10 15 20 25 30 35 40 0

20

40

60

B a

se lin

e R

e m

o ve

d P

P G

5 10 15 20 25 30 35 −20

0 20 40 60 80

P e

a k

D e

te ct

e d

P

P G

5 10 15 20 25 30 35 40

500

1000

time (s)

R R

( m

s)

(a)

(b)

(c)

(d)

(e)

Figure 3: (a) Raw PPG signal (b) Baseline (c) Baseline-corrected PPG signal (d) Peak detected on Baseline removed PPG signal and (e) RR inter- vals in millisecond

0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

−20

0

20

40

60

80

Time(s)

P

P G

RR 2RR1

RR 3

RR 4 RR5

RR 6

RR 7

Figure 4: RR intervals on processed PPG signal

pHF, nLF, nHF and LFHF, respectively. The power spec- trum of RR intervals for happy and sad emotions is repre- sented in Figure 5. From Figure 5, it is observed that the relative powers across different frequency bands are different for two different emotions. Therefore, the features extracted from these bands will serve as a good discrimination factor to classify different emotions. The description of each fea- ture in a nutshell is given in Table 1.

3.4 Classification Model In this work, we have considered three emotions namely

happy, sad and neutral for studying the role of time domain and frequency domain HRV features extracted from PPG signal in recognizing the emotions. Support Vector Ma- chines (SVMs) are used to develop emotion recognition mod- els. SVM classification is an example of supervised learning. SVMs are useful due to their wide applicability for classi- fication tasks in many signal processing applications such as baby cry classification [26], crowd noise detection [22], binarization selection [4], activity recognition [20] and per- son identification [21]. The Libsvm [3] package is used for this purpose. Non-linear SVM, employing the C SVC algo- rithm and having a Radial Basis Function (RBF) kernel is used. Each SVM is trained with positive and negative ex- amples. Positive feature vectors are derived from the PPG signal of intended emotion, and negative feature vectors are derived from the PPG signals of other emotion. Therefore, two SVMs are developed to represent two emotions.

4. RESULTS

Table 1: List of time domain and frequency domain HRV features Domain # Features Feature Description

1 meanNN mean value of NN intervals

1 medianNN median value of NN intervals

Time domain

1 SDNN standard deviation of NN intervals

1 SDANN standard deviation of mean NN intervals

1 pNN50 percentage of successive NN intervals differing

by greater than or equal to 50 ms

1 NN50 total number of successive NN intervals differing

by greater than or equal to 50 ms

1 RMSSD root mean square of successive NN differences

1 SDNNi mean of all the standard deviations of

NN intervals for all windows

1 meanHR mean value of heart rates

1 stdHR standard deviation of heart rates

3 peakVLF highest power in VLF band

peakLF highest power in LF band

Frequency domain

peakHF highest power in HF band

4

aVLF raw area of VLF band (ms2)

aLF raw area of LF band (ms2)

aHF raw area of HF band (ms2)

aTotal total raw area of VLF, LF and HF bands

3 pVLF relative VLF area w.r.t to total area

pLF relative LF area w.r.t to total area

pHF relative HF area w.r.t to total area

2 nLF normalized LF area (i.e. LF/(LF+HF))

nHF normalized HF area (i.e. HF/(LF+HF))

LFHF ratio of LF and HF areas

0 0.1 0.2 0.3 0.4 0.5 0.6 0

5

10

15

x 10 4

P S

D (

s2 /H

z)

0 0.1 0.2 0.3 0.4 0.5 0.6 0

5

10

15

x 10 4

P S

D (

s2 /H

z)

0 0.1 0.2 0.3 0.4 0.5 0.6 0

5

10

15

x 10 4

P S

D (

s2 /H

z)

Freq (Hz)

Neutral

Sad

Happy

Figure 5: Power spectral density obtained from HRV signals for Happy, Sad and Neutral emotions.

For evaluating the performance of the EDRS, the feature vectors, derived from the test PPG signal are given as inputs to the trained emotion models. The output of each model is given to the decision block as shown in Figure 1, where the category of the emotion is hypothesized based on the highest evidence among the 3 emotion models. In this study we have used Leave One Subject Out cross validation technique. EDRS performance using HRV features is given in Table 2. From results it is observed that the happy emotion is recognized with 78.7%, sadness with 81.8% and neutral with 90.9% where as overall average model accuracy is 83.8%. The key contribution has been in the form of meticulous data collection and identifying features for classification.

It is later verified from the online questionnaire that out of the 5 subjects which are misclassified as having neutral emo- tion with happy stimuli, 3 subjects actually reported that they felt disgusted, whereas out of 4 subjects misclassified as having neutral emotion with sad stimuli, 3 subjects felt boring. If we remove these subjects the accuracy obtained in classifying happy and sad are 86.7% and 90.0% respectively and overall average accuracy of the system is further im- proved to 89.2%. The improvement in precision, sensitivity and specificity measurements are also given in Table 3 inside brackets in italics. The prominent increment in the preci- sion value for the neutral emotion is because of the decrease in the false positive counts based on online questionnaire analysis.

Table 2: Performance of the emotion recognition system. ``````````̀Predicted

Actual Happy Sad Neutral

Happy 26 2 1

Sad 2 27 2

Neutral 5 4 30

Table 3: Experimental result

Happy Sad Neutral

True Positive 26 (26 ) 27 (27 ) 30 (30 )

False Positive 3 (3 ) 4 (4 ) 9 (3 )

False Negative 7 (4 ) 6 (3 ) 3 (3 )

True Negative 63 (60 ) 62 (59 ) 57 (57 )

Precision 0.90 (0.90 ) 0.87 (0.87 ) 0.77 (0.91 )

Sensitivity 0.79 (0.87 ) 0.82 (0.90 ) 0.91 (0.91 )

Specificity 0.95 (0.95 ) 0.94 (0.94 ) 0.86 (0.95 )

Model Accuracy 0.84 (0.89 )

5. SUMMARY AND FUTURE WORK In this study we have explored HRV features to facilitate

development of a companion system with quick and reliable emotion recognition via affordable sensing. It is found that the average accuracy obtained for classification of emotions

on 33 subjects is 83.8%. To improve the system performance and scope of study further, we have planned to execute the following steps in future

• We look forward to considering all the other six dis- crete emotional states in addition to the ones that we have already considered

• We also plan to detect the degree of emotion felt by a subject in one emotion space, i.e. If a subject is being shown a happiness elicitation stimuli (say), it is to be reported that what is the level of happiness felt by the subject at any particular instant of time.

• Galvanic skin response (GSR) data capture, analysis and inference has also been planned in addition to PPG. We would also like to fuse the features extracted from both PPG and GSR signals for improving the ac- curacy of detecting whether the particular target emo- tion has been fulfilled.

• We can perform analysis like Poincaré, non-linear, Frac- tal analysis on PPG and/or GSR signals in addition to time and frequency domain feature analysis and deter- mine contribution of individual features in classifying the emotions.

• Agrafioti [1] suggested that too many emotion mea- surements would cause interference to the subjects and might not be feasible for practical problems. Hence in future we plan to collect data unobtrusively and rec- ognize affective physiological patterns in real time to aid retail business industry.

6. ACKNOWLEDGMENTS The authors are grateful to Arnab Guha for helping dur-

ing the experimental set-up, data acquisition, and recruit- ment of eligible subjects and to the subjects for their active participation.

7. REFERENCES [1] F. Agrafioti, D. Hatzinakos, and A. K. Anderson. Ecg

pattern analysis for emotion detection. Affective Com- puting, IEEE Transactions on, 3(1):102–115, 2012.

[2] J. N. Bailenson, E. D. Pontikakis, I. B. Mauss, J. J. Gross, M. E. Jabon, C. A. Hutcherson, C. Nass, and O. John. Real-time classification of evoked emotions using facial feature tracking and physiological re- sponses. International journal of human-computer studies, 66(5):303–317, 2008.

[3] C.-C. Chang and C.-J. Lin. Libsvm: a library for sup- port vector machines. ACM Transactions on Intelli- gent Systems and Technology (TIST), 2(3):27, 2011.

[4] T. Chattopadhyay, V. R. Reddy, and U. Garain. Au- tomatic selection of binarization method for robust ocr. In 2013 12th International Conference on Doc- ument Analysis and Recognition, pages 1170–1174. IEEE, 2013.

[5] M. Elgendi. On the analysis of fingertip photoplethys- mogram signals. Current cardiology reviews, 8(1):14– 25, 2012.

[6] J. J. Gross and R. W. Levenson. Emotion elicitation using films. Cognition & emotion, 9(1):87–108, 1995.

[7] D. Hazer, X. Ma, S. Rukavina, S. Gruss, S. Walter, and H. C. Traue. Emotion elicitation using film clips: Effect of age groups on movie choice and emotion rat- ing. In International Conference on Human-Computer Interaction, pages 110–116. Springer, 2015.

[8] E.-C. Kao, C.-C. Liu, T.-H. Yang, C.-T. Hsieh, and V.-W. Soo. Towards text-based emotion detection a survey and possible improvements. In Information Management and Engineering, 2009. ICIME’09. In- ternational Conference on, pages 70–74. IEEE, 2009.

[9] K. H. Kim, S. Bang, and S. Kim. Emotion recognition system using short-term monitoring of physiological signals. Medical and biological engineering and com- puting, 42(3):419–427, 2004.

[10] A. Konar and A. Chakraborty. Emotion Recognition: A Pattern Analysis Approach. John Wiley & Sons, 2014.

[11] S. G. Koolagudi, V. R. Reddy, and K. S. Rao. Emo- tion recognition from speech signal using epoch pa- rameters. In Proc. of SPCOM 2010 6th Biennial Con- ference, pages 1–5, Bangalore, India, July 2010.

[12] C. L. Lisetti and F. Nasoz. Using noninvasive wear- able computers to recognize human emotions from physiological signals. EURASIP Journal on Advances in Signal Processing, 2004(11):1–16, 2004.

[13] M. Nardelli, G. Valenza, A. Greco, A. Lanata, and E. P. Scilingo. Recognizing emotions induced by af- fective sounds through heart rate variability. Affec- tive Computing, IEEE Transactions on, 6(4):385–394, 2015.

[14] R. W. Picard, E. Vyzas, and J. Healey. Toward ma- chine emotional intelligence: Analysis of affective physiological state. Pattern Analysis and Machine In- telligence, IEEE Transactions on, 23(10):1175–1191, 2001.

[15] K. Pilt, K. Meigas, R. Ferenets, K. Temitski, and M. Viigimaa. Photoplethysmographic signal wave- form index for detection of increased arterial stiffness. Physiological measurement, 35(10):2027, 2014.

[16] P. Rani, C. Liu, N. Sarkar, and E. Vanman. An em- pirical study of machine learning techniques for affect recognition in human–robot interaction. Pattern Anal- ysis and Applications, 9(1):58–69, 2006.

[17] K. S. Rao, S. G. Koolagudi, and R. R. Vempada. Emotion recognition from speech using global and local prosodic features. International journal of speech technology, 16(2):143–160, 2013.

[18] K. S. Rao, V. R. Reddy, S. Maity, and S. G. Koolagudi. Characterization of emotions using dy- namics of prosodic features. In Speech Prosody, Chicago, USA, May 2010.

[19] R. D. Ray. Emotion elicitation using films. Handbook of emotion elicitation and assessment, pages 9–28, 2007.

[20] V. R. Reddy and T. Chattopadhyay. Human ac- tivity recognition from kinect captured data using stick model. In International Conference on Human- Computer Interaction, pages 305–315. Springer Inter- national Publishing, 2014.

[21] V. R. Reddy, T. Chattopadhyay, K. Chakravarty, and A. Sinha. Person identification from arbitrary posi- tion and posture using kinect. In Proceedings of the

12th ACM Conference on Embedded Network Sensor Systems, pages 350–351. ACM, 2014.

[22] V. R. Reddy, A. Sinha, and G. Seshadri. Fusion of spectral and time domain features for crowd noise classification system. In 2013 13th International Con- ference on Intellient Systems Design and Applications, pages 1–6. IEEE, 2013.

[23] K. Smitha, N. Robinson, and A. Vinod. A study on the effect of emotion evoking videos on physiological signals. In Devices, Circuits and Systems (ICDCS), 2014 2nd International Conference on, pages 1–5. IEEE, 2014.

[24] K. P. Truong, D. A. van Leeuwen, and M. A. Neer- incx. Unobtrusive multimodal emotion detection in adaptive interfaces: speech and facial expressions. Springer, 2007.

[25] M. T. Valderas, J. Bolea, P. Laguna, M. Vallverdu, and R. Bailon. Human emotion recognition using heart rate variability analysis with spectral bands based on respiration. In Engineering in Medicine and Biology Society (EMBC), 2015 37th Annual Inter- national Conference of the IEEE, pages 6134–6137. IEEE, 2015.

[26] R. R. Vempada, B. S. A. Kumar, and K. S. Rao. Characterization of infant cries using spectral and prosodic features. In Communications (NCC), 2012 National Conference on, pages 1–5. IEEE, 2012.

[27] P. D. Welch. The use of fast fourier transform for the estimation of power spectra: A method based on time averaging over short, modified periodograms. IEEE Transactions on audio and electroacoustics, 15(2):70– 73, 1967.

[28] C. Yu, J. Zhou, and J. Riekki. Expression and anal- ysis of emotions: Survey and experiment. In Ubiqui- tous, Autonomic and Trusted Computing, 2009. UIC- ATC’09. Symposia and Workshops on, pages 428–433. IEEE, 2009.

[29] S.-N. Yu and S.-F. Chen. Emotion state identifica- tion based on heart rate variability and genetic algo- rithm. In Engineering in Medicine and Biology Society (EMBC), 2015 37th Annual International Conference of the IEEE, pages 538–541. IEEE, 2015.

[30] H. Zacharatos, C. Gatzoulis, and Y. L. Chrysanthou. Automatic emotion recognition based on body move- ment analysis: A survey. Computer Graphics and Ap- plications, IEEE, 34(6):35–45, 2014.

[31] L. Zhang, S. Rukavina, S. Gruss, H. C. Traue, and D. Hazer. Classification analysis for the emotion recognition from psychobiological data.