臨床靜脈輸液治療仰賴護理人員定期巡視確認點滴袋液位，在護理人 力不足的情況下，延誤換袋可能導致血液回流或空氣栓塞等醫療風險。現 有自動化方案多採用重量感測器或光電液滴計數，無法直接辨識刻度標 記；基於深度學習的影像方案則多依賴雲端推論，存在傳輸延遲與隱私疑 慮。為此，本研究提出一套部署於 ESP32-S3 微控制器之邊緣端智慧點滴 監控系統，採用 YOLOv12n 同時偵測刻度標記（ scale_mark ）與液位 liquid_level ），並透過 ESP-PPQ 進行 INT8 後訓練量化，搭配 ESP-DL 推論框架完成嵌入式部署。實驗結果顯示， YOLOv12n 於 224×224 輸入解 析度下之部署重測結果達 mAP@0.5 97.8%，量化後模型大小為 2.9 MB ，於 ESP32-S3 雙核心模式下之 純推論時間為 1，691 ms 。另以純卷積神經網路 架構之 ESPDet-Pico 作為對比基線，其雙核心推論時間為 103 ms ，部署 重測之 mAP@0.5 為 59.2% 。分析結果顯示， Attention-Centric YOLO 雖具 較高偵測精度，但其運算效率受 Transpose 與 MatMul 等非卷積運算子缺 乏 SIMD 最佳化支援所限制。為提升液位定位能力，本研究進一步提出兩 階段液位估算方法，以 scale_mark 作為幾何錨點裁切感興趣區域，並利 用輕量化 TinyRegressor 直接回歸液面相對位置。實驗結果於驗證集上達 到 100% 液位偵測率與 Success@0.05 = 97.4% 。整體而言，本研究驗證了 Attention-Centric YOLO 模型於微控制器平台部署之可行性，並建立一 套兼具系統化設計與實機驗證之智慧輸液監控方案，可作為未來醫療物聯 網與邊緣智慧感知系統發展之參考。;Clinical intravenous infusion therapy requires continuous monitoring of drip bag fluid levels， yet delayed replacement may cause blood backflow or air embolism. Existing automated solutions either rely on sensors that cannot directly identify scale marks or depend on cloud inference， introducing latency and privacy concerns. This study proposes an edge-based IV drip monitoring system deployed on an ESP32-S3 microcontroller. YOLOv12n was quantized to INT8 using ESP-PPQ and deployed through ESP-DL to simultaneously detect scale marks and liquid levels. Deployment re-evaluation achieved 97.8% mAP@0.5 with a 2.9 MB model and a dual-core inference time of 1，691 ms. A lightweight CNN baseline， ESPDet-Pico， achieved 59.2% mAP@0.5 and 103 ms inference time. Analysis indicates that non-convolutional operators such as Transpose and MatMul become the primary performance bottleneck on microcontrollers due to limited SIMD acceleration. To improve liquid-level localization， a lightweight TinyRegressor (115K parameters) was introduced to estimate the relative liquid-surface position， achieving a 100% detection rate and 97.4% Success@0.05 on validation set. The results demonstrate the feasibility of deploying Attention-centric YOLO models on microcontroller-based medical IoT systems.
