Anomalous gait feature detection using three-dimensional walking data with dimension reduction

Abstract

Human gait is a database that contains information from the subject. During the past decades, there have been many studies to find the characteristics of the subject like gender, identity, and disease through gait analysis. Most of the existing methods for gait analysis have lost much information because they used parameters obtained through only some part of kinematics. In this paper, a method to extract the features with substantial information from gait data using an autoencoder was proposed, one of machine learning techniques. The autoencoder is learned for movement of each segment of the human body and mapped to a low-dimensional feature parameter while minimizing the loss of information in gait data. The extracted feature parameters were applied for two applications. Gait predictions in five different environments were implemented through feature transfer between feature parameters, and person identification was performed. Finally, the relationship between the extracted feature parameter and the orthopedic anomalous gait feature was implemented as a regression network, and the presence or absence of anomalous gait feature was predicted through gait data.