全球高齡化導致長者照護壓力日增，傳統穿戴式健康監控設備雖常用於活動追蹤與跌倒預警，但因佩戴不便、依從性低及資料不連續等問題，限制其實務應用效益。為解決此挑戰，本研究提出一套結合邊緣運算、深度學習與人機迴圈機制的非接觸式長者活動監控系統，採用ESP32-CAM進行影像擷取，透過Google MediaPipe Pose萃取骨架關鍵點，並將關鍵點序列輸入預訓練LSTM模型進行行為分類，同時導入DDS協議實現模組間低延遲、高可靠性資料傳輸。系統亦結合React前端與MongoDB後端，提供即時活動紀錄與日誌審查介面，並設計Human-in-the-Loop機制允許照護者參與低信心樣本的分類校正與客製化建議。實驗結果顯示，本系統於ETRI資料集中達成平均分類準確率80%以上，端對端處理延遲控制於400ms內，具良好即時性與分類效能。相較傳統穿戴式系統，本研究所提方案在依從性、資料連續性與部署成本方面具明顯優勢，且人機協作設計更能兼顧彈性與準確性，展現其在智慧健康監控與銀髮族照護領域的高度應用潛力;The growing global aging population has significantly increased the burden on elderly care. Although traditional wearable health monitoring devices are commonly used for activity tracking and fall detection, their practical effectiveness is often limited by wearing discomfort, low user compliance, and discontinuous data. To address these issues, this study proposes a non-contact elderly activity monitoring system that integrates edge computing, deep learning, and a human-in-the-loop mechanism. The system utilizes ESP32-CAM modules to capture real-time images, from which 33 skeletal keypoints are extracted using Google MediaPipe Pose. These keypoint sequences are then fed into a pre-trained Long Short-Term Memory (LSTM) model to classify daily activities. A Data Distribution Service (DDS) protocol is employed to achieve low-latency, high-reliability data transmission among modules. The system also incorporates a React-based frontend and a MongoDB backend to provide real-time activity records and a log review interface. Furthermore, a human-in-the-loop mechanism is designed to allow caregivers to participate in correcting low-confidence classifications and providing personalized recommendations. Experimental results on the ETRI dataset demonstrate that the system achieves an average classification accuracy exceeding 80%, with end-to-end latency controlled under 400 milliseconds, indicating strong real-time performance. Compared with conventional wearable systems, the proposed approach offers clear advantages in compliance, data continuity, and deployment cost. Its human-in-the-loop design further enhances flexibility and accuracy, demonstrating high potential for practical application in smart healthcare monitoring and elderly care.
