Action recognition for educational proposals applying concepts of Social Assistive Robotics

Abstract

Action recognition has been gaining interest in research due to its great number of applications. So, the main contribution of this manuscript is a human–robot interaction framework, relying on dimension reduction of the system’s inputs in order to require a smaller dataset for training of an Artificial Neural Network. Our motivation is the development of a Social Assistive Robotics application. In summary, we choose nine standard actions to guide a robot, and two neutral ones to represent stand-by or resting cases. The dataset is created by people with different body shape, for robustness purposes, using only 5 to 10 samples of each class per person. Offline and online tests validate the method’s accuracy and confusion matrices clarify the results. A TicTacToe game using a ground robot exemplify a real world application, where each action represents a desired spot in the game. The results confirm a high accuracy, above 96.7%, in all the tests. Based on this, we can conclude our preprocessing strategy and classifier are capable of identifying the action patterns, even for a tiny dataset; thus, it is recommended for educational proposals due to its simplicity.