Continuous multi-modal human interest detection for a domestic companion humanoid robot

This paper addresses the scenario of how a humanoid robot can detect its human companion's intention of initiating interaction in a domestic environment. Previous research was mostly focused on the robot's understanding of human intentions during the course of the interaction, but not the initiation of it. The performance of the several unimodal and multi-modal algorithms have been evaluated using over 100 hours of real life experiments conducted in a city apartment living room. A key evaluation criterion is the ability of the algorithm to continuously monitor its environments to be responsive to human's infrequent requests, while at the same time having an extremely low number of false positives so as not to fire off interactions at random. The multi-modal approach incorporating speech recognition, audio source localisation and face detection features is proven to be superior, but further improvements are desired for real life applications.