Estimating Whole Body Motion from Partially Observed Body Motion During Walking

Abstract

Measurement of whole body motion using motion capture play important role in fields of rehabilitation, imitation learning, and human-robot interaction. However, the measurement requires a large number of markers and sensors to be worn and there are many limitation on the measurement environment. In this paper, we propose a novel method to estimate whole body motion from partially observed body motion during walking. We create a deep neural network similar to the Convolutional Neural Network (CNN) auto-encoder to extract features of walking motion from partial motion data and estimate whole body motion from the features. In the experiment, we compare the estimated joint angles and motion data to verify the usefulness of the proposed method.