Neonatal Sleeping State Estimation by Body Movement Detection Using Optical Flow

Abstract

We proposed an automatic classification method for estimating sleep-wake state in neonates using body movements. We used Optical Flow to detect body movements, and then applied block separation to the frame images to create a histogram of the amount of movement in each direction. The histogram of each block was combined to create feature vectors, and classified into 6 classes using hierarchical SVM. We also proposed two methods to mitigate the differences in the direction of video capture in the dataset. The experimental results showed that the highest value of macro-F1 was 0.675, which was classified by the proposed method using the data expansion method. Furthermore, the percentage of correct answers for each State exceeded 50% by unifying the orientation of the newborns.