Journal of Information Processing
Online ISSN : 1882-6652
ISSN-L : 1882-6652
Data Augmentation to Build High Performance DNN for In-bed Posture Classification
Yu EnokiboriKenji Mase
Author information
JOURNAL FREE ACCESS

2018 Volume 26 Pages 718-727

Details
Abstract

E-textiles have come to be used instead of several types of common equipment, such as bed-sheets, in some cases. An application using body pressure data collected through such bed-sheet type sensors is the in-bed posture classification expected for pressure ulcer prevention. Since such body pressure data is a kind of low-resolution image, Deep Neural Network (DNN) based algorithms seem suitable. However, it is difficult to collect enough data to use for DNN in this domain because the number of sleep postures obtained from one experiment is small. For an example, the number of postures collected from 19 subjects with four hours of sleep each is only 224. To solve such a small data-size problem in DNN, data augmentation techniques have been proposed. However, random augmentations are not so suitable. Therefore, we investigated appropriate augmentation parameters for this domain. As a result, the combination of the up to ±20% and ±40% random shifts along short and long sides of a bed, the up to ±10degree rotation, and non-use of other transformations showed the best performance. With the parameters, the built DNN showed 99.7% accuracy and 0.997 Weighted F1-score for three posture classifications: supine, left and right lateral positions, and 97.1% accuracy and 0.970 Weighted F1-score for four posture classifications: supine, prone, left and right lateral positions.

Content from these authors
© 2018 by the Information Processing Society of Japan
Previous article Next article
feedback
Top