Multimodal emotion recognition based human-robot interaction enhancement
Date
2018-05-03
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Üks afektiivse arvutiteaduse peamistest huviobjektidest on mitmemodaalne emotsioonituvastus, mis leiab rakendust peamiselt inimese-arvuti interaktsioonis. Emotsiooni äratundmiseks uuritakse nendes süsteemides nii inimese näoilmeid kui kakõnet. Käesolevas töös uuritakse inimese emotsioonide ja nende avaldumise visuaalseid ja akustilisi tunnuseid, et töötada välja automaatne multimodaalne emotsioonituvastussüsteem. Kõnest arvutatakse mel-sageduse kepstri kordajad, helisignaali erinevate komponentide energiad ja prosoodilised näitajad. Näoilmeteanalüüsimiseks kasutatakse kahte erinevat strateegiat. Esiteks arvutatakse inimesenäo tähtsamate punktide vahelised erinevad geomeetrilised suhted. Teiseks võetakse emotsionaalse sisuga video kokku vähendatud hulgaks põhikaadriteks, misantakse sisendiks konvolutsioonilisele tehisnärvivõrgule emotsioonide visuaalsekseristamiseks. Kolme klassifitseerija väljunditest (1 akustiline, 2 visuaalset) koostatakse uus kogum tunnuseid, mida kasutatakse õppimiseks süsteemi viimasesetapis. Loodud süsteemi katsetati SAVEE, Poola ja Serbia emotsionaalse kõneandmebaaside, eNTERFACE’05 ja RML andmebaaside peal. Saadud tulemusednäitavad, et võrreldes olemasolevatega võimaldab käesoleva töö raames loodudsüsteem suuremat täpsust emotsioonide äratundmisel. Lisaks anname käesolevastöös ülevaate kirjanduses väljapakutud süsteemidest, millel on võimekus tunda äraemotsiooniga seotud ̆zeste. Selle ülevaate eesmärgiks on hõlbustada uute uurimissuundade leidmist, mis aitaksid lisada töö raames loodud süsteemile ̆zestipõhiseemotsioonituvastuse võimekuse, et veelgi enam tõsta süsteemi emotsioonide äratundmise täpsust.
Automatic multimodal emotion recognition is a fundamental subject of interest in affective computing. Its main applications are in human-computer interaction. The systems developed for the foregoing purpose consider combinations of different modalities, based on vocal and visual cues. This thesis takes the foregoing modalities into account, in order to develop an automatic multimodal emotion recognition system. More specifically, it takes advantage of the information extracted from speech and face signals. From speech signals, Mel-frequency cepstral coefficients, filter-bank energies and prosodic features are extracted. Moreover, two different strategies are considered for analyzing the facial data. First, facial landmarks' geometric relations, i.e. distances and angles, are computed. Second, we summarize each emotional video into a reduced set of key-frames. Then they are taught to visually discriminate between the emotions. In order to do so, a convolutional neural network is applied to the key-frames summarizing the videos. Afterward, the output confidence values of all the classifiers from both of the modalities are used to define a new feature space. Lastly, the latter values are learned for the final emotion label prediction, in a late fusion. The experiments are conducted on the SAVEE, Polish, Serbian, eNTERFACE'05 and RML datasets. The results show significant performance improvements by the proposed system in comparison to the existing alternatives, defining the current state-of-the-art on all the datasets. Additionally, we provide a review of emotional body gesture recognition systems proposed in the literature. The aim of the foregoing part is to help figure out possible future research directions for enhancing the performance of the proposed system. More clearly, we imply that incorporating data representing gestures, which constitute another major component of the visual modality, can result in a more efficient framework.
Automatic multimodal emotion recognition is a fundamental subject of interest in affective computing. Its main applications are in human-computer interaction. The systems developed for the foregoing purpose consider combinations of different modalities, based on vocal and visual cues. This thesis takes the foregoing modalities into account, in order to develop an automatic multimodal emotion recognition system. More specifically, it takes advantage of the information extracted from speech and face signals. From speech signals, Mel-frequency cepstral coefficients, filter-bank energies and prosodic features are extracted. Moreover, two different strategies are considered for analyzing the facial data. First, facial landmarks' geometric relations, i.e. distances and angles, are computed. Second, we summarize each emotional video into a reduced set of key-frames. Then they are taught to visually discriminate between the emotions. In order to do so, a convolutional neural network is applied to the key-frames summarizing the videos. Afterward, the output confidence values of all the classifiers from both of the modalities are used to define a new feature space. Lastly, the latter values are learned for the final emotion label prediction, in a late fusion. The experiments are conducted on the SAVEE, Polish, Serbian, eNTERFACE'05 and RML datasets. The results show significant performance improvements by the proposed system in comparison to the existing alternatives, defining the current state-of-the-art on all the datasets. Additionally, we provide a review of emotional body gesture recognition systems proposed in the literature. The aim of the foregoing part is to help figure out possible future research directions for enhancing the performance of the proposed system. More clearly, we imply that incorporating data representing gestures, which constitute another major component of the visual modality, can result in a more efficient framework.
Description
Väitekirja elektrooniline versioon ei sisalda publikatsioone