一种应用于婴儿监护的改进YOLOv5算法

An improved YOLOv5 algorithm applied to infant monitoring

婴儿在婴儿床、椅子或床上睡觉时可能会有一些危险情况发生,比如婴儿因力量不足无法翻身、一直维持趴着状态而导致窒息。该研究专注于婴儿特征,包括行为特征(如趴着、爬动、躺着、坐着、站着)和神情特征(如平静、开心、伤心),基于YOLOv5目标检测算法并对其进行改进使之更适合于快速准确检测出婴儿状况。传统YOLOv5采用的原始NMS对重合度很高的预测框不友好,改进的YOLOv5则在传统YOLOv5基础上添加了Merge-NMS,使预测框的位置更为准确。经过验证,改进后的YOLOv5相比传统的YOLOv5的特征平均mAP@0.5∶0.95提升了2.7个百分点,其中动作特征爬动提升最高,提升了4.0个百分点。

There may be some dangerous situations when a baby is sleeping in the crib,chair or bed,such as the baby’s lack of strength to roll over and stay on his stomach,leading to suffocation.This study focuses on infant characteristics,including behavioral characteristics(such as lying on the stomach,crawling,lying down,sitting and standing)and facial characteristics(such as calm,happy and sad).Based on YOLOv5 object detection algorithm,it is improved to make it more suitable for rapid and accurate detection of infant status.The original NMS used by traditional YOLOv5 is not friendly to prediction boxes with high coincidence degree.The improved YOLOv5 adds Merge-NMS on the basis of traditional YOLOv5 to make the position of prediction boxes more accurate.After verification,the average feature of the improved YOLOv5 has increased by 2.7 percentage points compared with the traditional YOLOv5 at mAP@0.5∶0.95,in which the action feature crawling has the highest improvement 4.0 percentage points.