사용자 환경에 따라 적용 가능한 EMG 기반 손동작 인식 기법

Abstract

Recognition and prediction of hand gestures using electromyography(EMG) has various applications, such as in devices for human robot interactions, rehabilitation of paralytic patients, and prostheses for amputees. These application can be broadly divided into specific user-only environment and universal user environment. However, in multi-channel EMG used in many recent studies, there are numerous individual differences in the acquired data depending on the user, which cause serious performance degradation in the universal user environment. In this study, EMG data were collected from five subjects for 15 hand gestures using an 8 channel armband EMG sensor at a sampling frequency of 1000 Hz, and the individual differences in the EMG intensity (voltage) and activation channels were analyzed. In addition, the hand gestures were recognized using a convolutional recurrent neural network (CRNN) deep- learning architecture as a classifier, and an average accuracy of 95% was obtained in the experiments considering the user-only environment. In the experiments considering the universal user environment, 15 hand gestures were recognized with an average accuracy of 87.7% by applying the transfer learning technique. Thus, for hand gestures used in everyday life, we propose a recognition method that can be applied according to the user environment, from user-only environments to universal user environments.