Correlated space formation for human whole-body motion primitives and descriptive word labels

Abstract

The motion capture technology has been improved, and widely used for motion analysis and synthesis in various fields, such as robotics, animation, rehabilitation, and sports engineering. A massive amount of captured human data has already been collected. These prerecorded motion data should be reused in order to make the motion analysis and synthesis more efficient. The retrieval of a specified motion data is a fundamental technique for the reuse. Imitation learning frameworks have been developed in robotics, where motion primitive data is encoded into parameters in stochastic models or dynamical systems. We have also been making research on encoding motion primitive data into Hidden Markov Models, which are referred to as "motion symbol", and aiming at integrating the motion symbols with language. The relations between motions and words in natural language will be versatile and powerful to provide a useful interface for reusing motion data. In this paper, we construct a space of motion symbols for human whole body movements and a space of word labels assigned to those movements. Through canonical correlation analysis, these spaces are reconstructed such that a strong correlation is formed between movements and word labels. These spaces lead to a method for searching for movement data from a query of word labels. We tested our proposed approach on captured human whole body motion data, and its validity was demonstrated. Our approach serves as a fundamental technique for extracting the necessary movements from a database and reusing them.