Detection of social signals for recognizing engagement in human-robot interaction
This work addresses engagement recognition for improving robot dialogue strategies, but it is incremental as it builds on existing methods for social signal detection.
The paper tackled the problem of detecting engagement in human-robot interaction by recognizing social signals like nodding and laughter, achieving reasonable performance with automatic detection compared to manual annotation.
Detection of engagement during a conversation is an important function of human-robot interaction. The level of user engagement can influence the dialogue strategy of the robot. Our motivation in this work is to detect several behaviors which will be used as social signal inputs for a real-time engagement recognition model. These behaviors are nodding, laughter, verbal backchannels and eye gaze. We describe models of these behaviors which have been learned from a large corpus of human-robot interactions with the android robot ERICA. Input data to the models comes from a Kinect sensor and a microphone array. Using our engagement recognition model, we can achieve reasonable performance using the inputs from automatic social signal detection, compared to using manual annotation as input.