Human-Machine Interaction, Communication Initiation Probability Estimation - Communication initiation using Computer Vision, Machine Learning and Artificial Intelligence

Abstract

Robots and digital assistants today typically require the use of key words or phrases to activate them. In this paper we propose and implement a prototype for initiating conversations with robots in a more natural way. Measurements of three key indicators of conversation interest – eye contact, head pose and distance – are captured from a video stream and, using various techniques such as convolutional neural networks and 2d/3d projections, analyzed to estimate who, if any, is most interested in engaging in conversation. Through user tests, we gather and compile data to evaluate the validity of the approach, and the accuracy and performance of the implementation. It is our conclusion that this is a viable approach, with the potential to feel very natural to the users, although the prototype has a number of problems with regards to accuracy and performance.