表面筋電位の時間周波数解析に基づくCNNを用いた手書き動作の識別

抄録

Many studies on the development of HMI using surface electromyogram (sEMG) have been reported. However, challenges remain in the reproduction of skillful motions. This study, focusing on the handwriting of Japanese characters as an example of such a skillful motion, proposes and examines a new method for motion classification as follows. At first, sEMGs are measured at 4 points on the forearm of the dominant hand during the handwriting motion. Secondary, time-frequency analysis is performed to extract the features from the sEMGs. Finally, the result is supplied as an input image to a CNN which is trained for classification of the written characters. By comparing the classification rate with a conventional 3-layered neural network, it is confirmed that the proposed method improves the classification accuracy by about 11.2% for the training data and 10.3% for the verification data.