基于双模型的婴儿异常睡姿识别方法

Recognition Method for Abnormal Sleeping Posture of Infants Based on Dual Models

摘要: 在父母对婴儿的监护中，婴儿的异常睡姿会引起婴儿睡眠猝死，尤其在一两岁经常发生，研究表明，婴儿的俯卧极容易引发这种现象，及时、准确地对婴儿进行目标检测和睡姿识别，对其健康安全有着至关重要的作用。为了综合满足婴儿监护的准确度、误报率和检测速度的要求，提出了一种基于双模型结构婴儿异常睡姿识别方法。采用改进后YOLOv5与OpenPose相结合的方法，结合YOLOv5的目标检测实时性和OpenPose的人体姿态估计的准确性，通过引入感知机与门作为二分类器，将两个模型的输出结果融合，对婴儿异常睡姿进行准确的二元分类判定，实验表明，本文模型在婴儿睡姿检测中准确率达到94.7%，误报率为2.0%，能够有效处理婴儿异常睡姿检测任务，在提高准确率的同时，还具有较低的误报率，可提高系统的可靠性和监护人的舒畅度。

Abstract: In the parents’ monitoring of the baby, the baby’s abnormal sleeping position can cause sudden in-fant sleep death, especially in one or two years old, research shows that the baby’s prone lying is very easy to cause this phenomenon, timely and accurate detection of the baby’s target and sleeping position identification, has a vital role in its health and safety. In order to meet the requirements of accuracy, false alarm rate and detection speed of infant monitoring comprehensively, a method of infant abnormal sleeping position recognition based on dual model structure was proposed. By us-ing the improved method of combining YOLOv5 and OpenPose, combining the real- time target de-tection of YOLOv5 and the accuracy of human pose estimation of OpenPose, the perceptron and door were introduced as binary classifier, and the output results of the two models were fused to make accurate binary classification judgment on the abnormal sleeping position of infants. The experi-ment showed that, the accuracy rate of this model in infant sleeping position detection reached 94.7%, and the false alarm rate was 2.0%, which can effectively handle the task of abnormal sleep-ing position detection of infants. While improving accuracy, the model also has a low false alarm rate, which improves the reliability of the system and the comfort of guardians.