2017 年 83 巻 851 号 p. 17-00129
This paper presents a method for detecting and cancelling motion artifact related to standing and walking in a functional near-infrared spectroscopy (fNIRS) signal. Our fNIRS device has 22 channels. The motionless fNIRS signal from each channel is represented by a fourth-order autoregressive model (AR model), and the related parameters are estimated based on the motionless fNIRS signal using Yule Walker equation. The motion artifacts included in the fNIRS signal are cancelled using the Kalman filter constructed from the AR model. However, the cancellation may be insufficient when the motion artifacts are strong. To determine in which fNIRS channels the motion artifacts are cancelled insufficiently, we apply an observation prediction error related to the Kalman filter and a discrete Fourier transform. The brain activity of the user is then recognized from those fNIRS channels in which the motion artifacts are cancelled sufficiently. To evaluate the proposed method, a mobile robot is controlled using an fNIRS devise as worn by 10 subjects while standing, walking, or sitting. The success rate of brain-activity recognition by the proposed method was 64.2%, whereas that without the proposed method was 54.0%.