Classification of Electrooculography Signals Using Convolutional Neural Networks for Interaction with a Manipulator Robot

Electrooculography (EOG) has been widely applied in human–machine interfaces (HMI) because it provides a reliable communication channel to assist people with disabilities. However, signal behavior under different conditions hinders eye movement classification when algorithms based on voltage threshold detection are used. Therefore, recalibration of the system is required for the classification algorithm to work correctly. Based on the above, a classification algorithm was developed to analyze the data vector corresponding to the entire EOG waveform, instead of just one characteristic value of the signal, thus avoiding the system recalibration process. A convolutional neural network (CNN) was implemented to classify six targets corresponding to different eye movements. The proposed model was compared with a feedforward neural network (FNN) to evaluate its performance. The results were implemented in an HMI for interaction with a manipulator robot.