Comparison of Head and Body Movement Patterns in Naturalistic Human-Machine Interaction

This paper aims on the investigation and recognition of upper-body movements during a naturalistic Human-Machine Interaction, in which humans interact with a technical system while sitting in front of it. Therefore, we focus on the Last Minute Corpus, that provides such a naturalistic scenario in combination with multimodal recordings. For feature extraction an approach called Probabilistic Breadth Features was used, allowing a condensed investigation of movement patterns. Finally, the classification was based on Extreme Learning Machines, comparing features obtained in three different conditions: the Kinect's spine point, head point, and a combination of both. In context of this naturalistic interaction setting, a mean accuracy of 86.1% was achieved.