An Improved Merge-YOLOv5 Algorithm for Infant Monitoring

When a baby is sleeping in a crib, chair or bed, there may be some dangerous situations, such as choking because the baby is not strong enough to turn over and remains on his or her stomach. This study focuses on infant characteristics including behavioral characteristics (such as lying down on their stomachs, crawling, lying down, sitting, standing) and facial features (such as calm, happy, sad), based on the YOLOv5 target detection algorithm and improved it to make it more suitable for rapid and accurate detection of infant conditions. The original NMS used by original YOLOv5 is not friendly to prediction boxes with high coincidence degree. The Merge-YOLOv5 adds merge-NMS on the basis of original YOLOv5 to make the position of prediction boxes more accurate. After verification, the merge-YOLOv5 features improved by 2.5 percentage points compared with the original YOLOv5 features mAP@0.5:0.95 on average, in which the motion features of lying on the stomach get an improvement up to by 4.4 percentage points. In this way, it is proved that our MergeYOLOv5 is promising for infant monitoring.