Journal of Biomechanical Science and Engineering
Online ISSN : 1880-9863
ISSN-L : 1880-9863
Generation of gait data less prone to stumbling considering the physical differences among trainees
ジャーナル フリー 早期公開

論文ID: 20-00478


Several training methods have been developed to obtain motion information during real-time walking and feed it back to trainees who adjust their gait to ensure that the measured gait parameters approach target value, which may not always be suitable for every trainee owing to physical differences between individuals. This paper proposes a method of setting this target value considering these physical differences and discusses the usefulness of the gait training method, wherein a multichannel deep convolutional neural network (MC-DCNN) gait classification model constructed by learning ideal or non-ideal gait features beforehand is used for trainee gait classification. Activation maximization is applied to the MC-DCNN model; data wherein the ideal walking features are activated are generated based on trainee gait data. However, the amounts of features to be activated to generate a possible and natural gait are restricted. The original trainee gait, beyond individual physical differences, and gait data generated based on the original gait data seem to yield the target value considering the physical differences among individuals. This study focused on gait related to stumbling. To verify its usefulness, a multivariate gait dataset consisting of kinematic and kinetic indices labeled as “gait rarely associated with stumbling” or “gait frequently associated with stumbling” was divided into a training set, validation set, and test set. The MC-DCNN model learned gait features for multivariate gait data classification in the training set. It classified the gait with 96.04±0.12% accuracy against the validation set. Finally, by applying the proposed method to the multivariate gait data contained in the test set, we generated multivariate gait data classified as “gait rarely associated with stumbling” based on the input data. In addition, the generated multivariate gait data include motion that increases the thumb-to-ground distance and describe possible and natural gait considering the physical differences among individuals.

© 2021 by The Japan Society of Mechanical Engineers