| 摘要: | 照護者在照顧嬰兒時，可能發生無法隨時關注其狀態的情形，使嬰兒因溢奶、翻身、趴睡等情形，致使呼吸不順而發生憾事。又因現有產品多為感測器式嬰兒偵測系統，功能單一且易干擾孩童；而既有的視覺式嬰兒偵測研究中，又多僅關注於呼吸頻率、面部特徵及單一動作，尚有許多值得探討之處。
因此，本論文提出基於深度學習技術，專注於嬰兒影像畫面之危險監測系統，包含兩大功能之偵測：（1）臉部遮擋辨識：判斷嬰兒臉部是否遭非奶嘴之異物遮蔽，進而可能發生窒息危險、及（2）姿勢辨識：分析嬰兒正躺、爬躺、坐姿及站立四種基礎姿勢，若為趴躺或站立之姿，則有可能發生呼吸不順或跌落床面等危險。綜上功能，當本系統讀取一段嬰兒影片後，可藉模型判斷嬰兒是否處於警示狀態，以提醒照護者。
本研究中，嬰兒臉部偵測部分，使用速度較快的 SSD 演算法，以及準確率較高的 RetinaFace 演算法，使整體系統在執行速度及準確度間達到平衡。而由於目前未有公開之嬰兒資料集，故本文收集網路真實嬰兒之不同視角圖片及影片，自製嬰兒臉部與姿勢資料集各 3475 張及 15416張影像，再以 ResNet50 進行臉部遮擋辨識及姿勢辨識兩模型之訓練，其訓練及測試準確度皆達 99%。由此證明，本研究對於嬰兒危險監測系統具有良好的可用性及獨特性。;The babysitter may not focus on the status of the infant at any time. When unpredictable things happen to the baby, such as spitting up, rolling over, or sleeping on his stomach, the babysitter won’t notice immediately. Most of the existing products are sensor-based infant detection systems, which are single-function and may disturb the movement of the baby. However, the existing vision-based infant detection studies only focus on breathing rate, facial features, and individual movements.
Therefore, this paper proposes a danger monitoring system based on deep learning technology. The system focuses on baby images and includes two major functions: (1) Facial Occlusion Recognition: Determine whether the infant’s face is occluded by foreign objects, which may cause suffocation. (2) Posture Recognition: The four basic postures of infants are analyzed: lying on the back, lying on the stomach, sitting and standing. If the baby is lying on his stomach or standing, he may be at risk of breathing difficulties or falling off the bed. In summary, while monitoring the baby’s video, the system can alert the babysitter when the infant is in an alarm state.
In this study, infant face detection uses the faster execution time SSD algorithm and the higher performance RetinaFace algorithm. With these algorithms, the system strikes a balance between execution speed and accuracy. There is currently no open source infant dataset. Therefore, this paper collects real baby images and videos from different perspectives from the Internet to create an infant face dataset with 3475 images and an infant posture dataset with 15416 images. Then, two models of face occlusion recognition and posture recognition are trained using ResNet50, and the training and testing accuracy are 99%. This proves that this study has good utility and uniqueness for infant danger monitoring system. |
