EEG Based Eye Movements Multi-Classification Using Convolutional Neural Network

Several studies stated that more natural and direct interactive control methods are needed in pilot training especially in case of emergency. The expression of intention based on biological signal is a natural way. Eye movement plays an important role in human interaction with the environment which is suitable as a control instruction. Eye movement signals have certain temporal feature in Electroencephalogram (EEG) signals. We collected the eye movement signals by an EEG headset with dry electrodes with high performance. At the same time, deep learning has shown great success in EEG signal processing. In this work, a shallow 2D-Convolutional Neural Network (CNN) structure is proposed to classify the eye movements in 4 directions of upward, downward, leftward and rightward in EEG signals, which are used as 4 kinds of commands to control virtual aircraft. The data are fed to the network through the simulation training experiment designed by us. The average accuracy of the 4 eye movement directions was 88.13%, 86.59%, 87.78% and 87.85% respectively. Among the 10 subjects, the highest accuracy was 92.71%, the lowest was 84.88%, and the total average accuracy was 87.59%, which proves the effectiveness of the scheme.