실시간 적용을 위한 CNN 모델 기반 보행 이벤트 검출 알고리즘

Abstract

It is important to apply the control algorithm for gait assistance according to the appropriate gait cycle. Therefore, a gait event detection algorithm that can be applied in real time is required. When an inertial measurement unit (IMU) is attached to the waist to reduce the number of sensors, it becomes difficult to detect a gait event compared to the lower body. We intend to utilize a convolutional neural network (CNN) that can automatically select and learn distinguishable signals. In this study, we pro posed a gait event detection algorithm based on a convolutional artificial neural network model using a moving window to enable real time application. Eight participants walked at three walking speeds on a treadmill to collect data for algorithm developmen t and validation. Acceleration and angular velocity signals were measured from the IMU attached to the waist, and the ground reaction force was measured using force plates. Time points of heel strike and toe off were detected using the gait events classifi ed by the CNN model. 92.3 % of heel strikes and 89.9 % of toe offs were detected among the total gait event time points, and the heel strike and toe off were estimated with mean absolute errors of 14.5 ms and 12.5 ms, respectively. These results imply that the CNN model based gait event detection algorithm can be applied in real time.