Classification and Features Visualization of Gait Using Convolutional Neural Network for Gait Feedback Training Optimized for Individuals

Abstract

In this research, to develop a gait feedback training system optimized for individuals where trainees can efficiently train features that do not satisfy ideal walking using a deep learning, we examine classification and features visualization of gait using Convolutional Neural Network (CNN) and Grad-CAM. In the experiment, the thumb-floor distance of right foot was measured when young people walked normally and when they walked with a brace, to limit their movement. Further, these data were clustered to 3 clusters using k-shape method. And these data were learned and classified as input data and using these cluster as the label. As the result, the accuracy was 86.07%. In addition, the part where the feature in thumb-floor distance appears were visualized as heat map using Grad-CAM and it is confirmed that usefulness for gait training.