Research On Video-based Sleep Evaluation Algorithm For Infants

Sleep accounts for most of the baby’s life.Good quality sleep not only helps the development of the baby’s bones,tissues and immune system,but also contributes to the maturity and intellectual growth of the central nervous system.At home and abroad,research on sleep evaluation is mainly aimed at patients or adults,and there are few studies on infants.This thesis mainly studies video-based baby sleep assessment algorithm.Because the videos used in this article are all from the Internet,considering the uniqueness of video sources and the unique sleep habits of infants at different ages.In this article,three aspects have been studied including sleep-awake identification,sleep pose classification and sleep quality evaluation.The main work of this thesis is as follows:First,infant sleep-awake identification.This thesis proposes an SVM(Support Vector Machine)-based infant sleep detection algorithm.Because sleep is a continuous process,to determine whether a certain moment is in sleep,it should be combined with a certain interval including this time.The algorithm uses the data within one minute before and after the current time to evaluate and assigns different weights according to the distance from the target time.Finally,through training with SVM classifier and under the condition of this baby video,the actual sleep detection accuracy is up to 85.5%.Second,the classification of infant sleep postures.This thesis proposes an XGBoostbased baby sleep posture classification algorithm.The current popular sleep posture detection is based on pressure sensors and depth cameras.The classification of infant sleep posture based on ordinary video is missing in the literature,and the sleep posture of the baby is very different from that of adults,and it is sometimes difficult to determine which sleep posture belongs to.In order to simplify the work and facilitate the subsequent sleep quality assessment,the algorithm divides the baby’s sleep posture into four postures: left side,right side,supine,and prone.The algorithm combines the characteristics of infant sleep posture,and extracts the features based on HOG(Histogram of Oriented Gradients),Hu moment feature and difference feature.Finally,PCA(Principle Component Analysis)is used to reduce the feature dimension.Through the integrated training model XGBoost,better training results are obtained.Finally,Under the condition of data set adopted in this thesis,the accuracy rate of classification in infant sleep posture is up to 84.7%.Third,infant sleep quality assessment.This thesis presents a video-based infant sleep quality assessment algorithm.At present,there are many ways to evaluate sleep quality,such as smart bracelets,sleep detectors,pressure sensors and so on,but these are not suitable for infants.This algorithm takes into account the sleep habits of infants at different ages,and combines the influence of different sleep postures on sleep quality.It achieves objective mathematical evaluation independent of sleep length and infant age.The results obtained from the algorithm is almost consistent with the professional doctors’ evaluation of video infants.Under the condition of sleep evaluation in this thesis,the proposed algorithm in this thesis can be used to determine sleep quality.