An Engagement Evaluation Algorithm Based on Dynamic Bayesian Network for Human-Robot Interaction

In the process of human-robot interaction, an engagement evaluation algorithm is proposed based on dynamic Bayesian network and domain expert knowledge, which focuses on solving the problem of insufficiency of training samples. Firstly, the social signals during the human-robot interaction and the reasoning process of dynamic Bayesian network are described, and the topology of the evaluation model of the engagement is developed. Then, limitations of the parameterization of the data-driven model are considered. Experts' opinions are collected by designing the linguistic variable set, and are processed by fuzzification and defuzzification. Thus, the construction of evaluation model is completed. Finally, the evaluation model is verified through the robot platform NAO and the design of interactive scenario. The experiment shows that the results of engagement evaluation are consistent with the actual behaviors of the interactive object, so that the robot can correctly identify the engagement intention of the interactive object in the process of human-robot interaction. Compared with the existing methods of engagement evaluation, the proposed method is not limited by training samples, and has high practical value and strong generalization ability.