Drowsiness Detection From EOG Using Convolutional Neural Networks

Fatigue reduces human’s ability to react to and handle the sudden changes. If not react in time and response properly, serious acident would happen. Consider the situations like driving, fatigue is known for one of the major causes for acidents and leads to significant property losses and casualties.To solve this serious problem, we proposed a new drowsiness detection model based on convolutional neural network. This model extracts features from the electrooculogram(EOG) signals, predicts the body states of the person and warn the person if he or she gets tired in order to reduce the possiblity of accidents.Using deep neural networks, the model proposed in our paper uses a completely new way to extract features from EOG signals automatically against manual ad-hoc feature extractions. The manual ad-hoc feature extraction procedures are usually complicated and result in poor performance and ineffectiveness as a result of individual body differences. The convolutional neural network relyed on backpropagation algorithmn can detect the patterns in EOG signals generated by different types of eye movements and extracts features from these patterns and fuse them into more complicated feature in latter stage. The automatically extracted features always result in higher reliability and accuaracy. This paper also proposed a preprocessing scheme for EOG signals which are fed to the network and try to extert as few in?uence on the raw signal as possible. The final prediction of degrees of fatigue is obtained from smoothed result by linear dynamic system which further reduce the in?uence of fatigue-unrelated EOG signals.The EOG signals collected in 22 experiments are processed. Comparison between convolutional neural network and traditional manual feature extraction shows that convolutional neural network yields models of significantly higher correlation coe?cients and results indicate that our model processes an equivalent or even better abilities than the corresponding models built on commonly used ad-hoc statistical features on drowsiness detection.