Head nod detection in dyadic conversations

Abstract

In face-to-face interactions, head gestures play an important role as one of the back-channel signals. As one of them, head nods can be used to display the approval or interest of listeners as a feedback in dyadic conversations. Hence detection of head nods is expected to improve understanding of the given feedback and to improve human-computer interaction. This study targets to detect head nods in the purpose of making human-computer interaction more human like. In the process, 3D head model is obtained by the Microsoft Kinect and the Openface application. Binary classification is performed on spectral features, which are extracted from 3D head motion, with the Support Vector Machine (SVM) classifier. Consequently, upon the classification, `head nod' or `not head nod' outputs are obtained. In the experimental studies, head nod detection accuracy is obtained as 92% for Microsoft Kinect and 91% for Openface over the Joker dataset.