Prediction of Intentions Behind a Single Human Action: An Application of Convolutional Neural Network

Abstract

Due to the rapidly increasing need of human robot interaction (HRI), more intelligent robots are in demand. However, the vast majority of robots can only follow strict instructions, which seriously restricts their flexibility and versatility. A critical fact that strongly negates the experience of HRI is that robots cannot understand human intentions. This study aims at improving the robotic intelligence by training it to understand human intentions. Different from previous studies that estimated human intentions from distinctive actions, this paper introduces a method, CNN-vote, that applies convolutional neural networks (CNN) to predict human intentions behind a single action. Experiments of predicting the intention of human throwing a ball towards designated targets are conducted to verify the effectiveness of the method. The proposed deep learning based method proves the feasibility of applying CNN under a novel circumstance. Experiment results show that the CNN-vote method requires less tuning work and out competes three traditional machine learning techniques. In current context, the CNN-vote predictor achieves the highest testing accuracy with relatively less data needed.