A Basic Study on Gait Event Detection by Deep Learning Using Inertial Sensor

Abstract

Evaluation of motor function is important in rehabilitation. In our previous studies, a method for estimating the timing of gait events (Heel-off, Toe-off, Initial-contact, Foot-flat) using an inertial measurement unit attached to the foot was developed. This algorithm detects gait events based on the zero-cross point or threshold of the angular velocity of the foot. However, the threshold value might need to be adjusted for each subject, which was not practical. In this study, the method using semantic segmentation model by convolutional neural network was proposed in order to develop an automatic detection method for gait events. In the walking of healthy subjects, the model was trained using the gait events measured using pressure sensors attached to a shoe as teacher data. It was shown that the gait events could be detected with generally good accuracy. In the future, it is necessary to verify the method with the hemiplegic walking.