GRADUATION THESIS BACHELOR OF ENGINEERING IN INFORMATION TECHNOLOGY

Abstract

Nowadays, in preschools, nursery teachers are usually responsible for overseeing the preschoolers to ensure their safe. They also install surveillance cameras to further guarantee the safe of children when they are not around. In lunch-break, the nursery teachers often use this opportunity to take a rest or have lunch. In addition, there is normally only one supervisor responsible for inspecting the surveillance cameras. So, it’s difficult for the supervisor know when there are some waking up, affecting others or getting out of class in lunch break. Therefore, the model of recognizing children behaviors to identify their sleeping state is performed in this thesis. The system will send a warning text to preschool teachers mobile phone when there is a kid who wakes up in lunck-break. The model uses YOLOv4 to locate and label the children actions on the frame. The training dataset are collected from a preschool and consist of 34 videos with the total length around 1 hour. The system can process a video at the speed more than 36 frames per second on a GPU. With the testing set of 50 images with around 20 children in each image, the mean average precision (mAP) of the predict children’s actions system is 80.18% with IOU=0.5