Validating Attention Classifiers for Multi-Party Human-Robot Interaction

Foster, M. E. (2014) Validating Attention Classifiers for Multi-Party Human-Robot Interaction. Proceedings of the HRI 2014 Workshop on Attention Models in Robotics, Bielefeld, Germany, 03 Mar 2014.

Full text not currently available from Enlighten.

Publisher's URL:


A critical task for a robot designed for interaction in a dynamic public space is estimating whether each of the people in its vicinity is currently seeking the robot’s attention. In previous work, we implemented two strategies for estimating the attention-seeking state of customers for a robot bartender—a rule-based classifier derived from the analysis of natural human behaviour, and a set of classifiers trained using supervised learning on a labelled multimodal corpus—and compared the classifiers through cross-validation and in the context of a full-system evaluation. However, because the ground-truth user behaviour was not available, the user study did not fully assess the classifier performance. We therefore carried out a new study validating the performance of all classifiers on a newly recorded, fully labelled test corpus. The highest-scoring trained classifier from the cross-validation study performed very badly on this new test data, while the hand-coded rule and other trained classifiers did much better. We also explored the impact of including information from previous frames in the classifier state: including previous sensor data had a mixed effect, while including the previous attention estimates greatly diminished the performance of all classifiers.

Item Type:Conference or Workshop Item
Glasgow Author(s) Enlighten ID:Foster, Dr Mary Ellen
Authors: Foster, M. E.
College/School:College of Science and Engineering > School of Computing Science

University Staff: Request a correction | Enlighten Editors: Update this record