隨著人口結構逐漸邁向高齡化，跌倒已成為年長者常見且高風險的意外事件之一，若未能即時發現與通報，往往可能造成嚴重後果。因此，本研究旨在建立一套能即時監測並辨識年長者異常行為的跌倒偵測系統，結合深度學習與模糊邏輯技術，以提升居家安全與照護效率。系統設計透過 LabelImg 工具標註資料，建立模型訓練所需的標記資訊。接著應用 YOLO 模型快速鎖定人物位置，並整合 Mediapipe 技術，擷取人體關鍵點座標，包括鼻子、髖部與腳踝。進一步推算鼻子的下降速度、關鍵點距離變化與偵測框的寬高比例等動作參數，作為後續分析依據。
在行為辨識階段，系統將上述參數輸入模糊邏輯系統，結合推理機制與自定義規則進行判斷。當參數組合達到特定條件時，系統能即時識別跌倒事件並觸發警示與通報功能，實現自動化安全監控的目標，進一步強化高齡者的防護。實驗結果顯示，系統在多種測試條件下皆具備穩定且高準確率的辨識能力，能有效降低誤判與漏判。本研究展示了影像式跌倒偵測系統在實際應用中的潛力，未來亦可擴展至其他行為辨識或居家安全監控場景。

As the global population continues to age, falls have become one of the most common and high-risk accidents among the elderly. If not detected and reported in a timely manner, such incidents can lead to serious consequences. This study aims to develop a fall detection system capable of real-time monitoring and identification of abnormal behaviors in older adults, integrating deep learning and fuzzy logic techniques to enhance home safety and caregiving efficiency.The system begins with data annotation using the LabelImg tool to generate labeled datasets for model training. YOLO is then employed to rapidly locate individuals in video frames, while Mediapipe is used to extract key human body landmarks, including the nose, hips, and ankles. These data points are further used to compute action-related parameters, such as the downward velocity of the nose, changes in distances between key points, and the aspect ratio of the bounding box.
In the behavior recognition stage, these parameters are input into a fuzzy logic system, which uses a rule-based reasoning mechanism to evaluate and classify actions. When specific threshold conditions are met, the system can accurately detect fall events and trigger instant alerts and notifications, aiming to automate safety monitoring and improve protection for the elderly.
Experimental results demonstrate that the system maintains high accuracy and stability under various testing conditions, effectively minimizing false alarms and missed detections. This research highlights the potential of vision-based fall detection systems in real-world applications and suggests that the approach can be extended to other forms of behavior recognition or home safety monitoring scenarios in the future.

目錄
目錄	IV
圖目錄	VI
表目錄	IX
第一章 緒論	1
1.1 研究背景與動機	1
1.2 研究目的	2
第二章 背景技術介紹	3
2.1 機器學習	3
2.2 深度學習	5
2.3 卷積神經網路	8
2.4  YOLO的物件偵測	10
2.4.1 YOLOv2	13
2.4.2 YOLOv3	15
2.4.3 YOLOv4	17
2.4.4 YOLOv7	18
2.5  MEDIAPIPE POSE	20
2.6 模糊理論概述	21
第三章 研究方法	23
3.1 資料標註與預處理	23
3.2 人體姿勢辨識	25
3.2.1 鼻子下降速度	25
3.2.2 髖部與腳踝距離	26
3.2.3 Bounding Box 寬高比	27
3.3 跌倒辨識	28
3.3.1 模糊化-隸屬度函數建立	30
3.3.2 模糊推論規則庫建立	34
3.3.3 解模糊化	35
3.3.4 跌倒閥值判斷	35
3.3.5 警報通知與報表生成	37
第四章 實驗設計與結果	39
4.1 實驗資料	39
4.2 實驗數據分析	39
4.3 實驗結果	41
第五章 結論	50
參考文獻	51

圖目錄
圖 1. 監督式學習示意	3
圖 2. 非監督式學習示意圖	4
圖 3. 強化學習示意圖	5
圖 4. AI、機器學習和深度學習關係	5
圖 5. 類神經網路示意圖	6
圖 6. 機器學習與深度學習的差異	6
圖 7. 卷積神經網路CNN基本架構	10
圖 8. YOLO卷積神經網路架構圖	11
圖 9. YOLO分割影像及非極大值抑制NMS框選	12
圖 10. IOU計算公式	12
圖 11. 誤差平方和（SUM-SQUARED ERROR）損失函數公式	13
圖 12. YOLOV2性能精確度	14
圖 13. 物件偵測演算法準確度與效率分佈圖	15
圖 14. 各物件偵測模型準確率比較表	16
圖 15. 各物件偵測模型準確率與推理速度比較表	17
圖 16. YOLOV7 效能與其他模型比較（MS COCO）	19
圖 17. 33 POSE LANDMARKS.	21
圖 18. 模糊控制系統基本架構	22
圖 19. 基於深度學習之老人異常活動偵測報告系統架構圖	23
圖 20. 人像資料集共計700張	24
圖 21. 人體姿勢辨識流程圖	25
圖 22. 鼻子下降速度參數	26
圖 23. 髖部-腳踝距離	27
圖 24. BOUNDING BOX 寬高比	28
圖 25. 跌倒辨識流程圖	29
圖 26. NOSE SPEED 模糊隸屬函數分佈圖	32
圖 27. HIP-ANKLE DISTANCE 模糊隸屬函數分佈圖	33
圖 28. BOUNDING BOX RATI 模糊隸屬函數分佈圖	33
圖 29. ROC 曲線與最佳閾值位置（AUC = 0.93）	36
圖 30. LINE通知	37
圖 31. 介面警示更新	38
圖 32. 各影像樣本之跌倒辨識成功與否分布圖	40
圖 33. 動作參數與信心值變化趨勢圖	40
圖 34. 姿勢辨識結果示意圖(FRAME 10與FRAME 20)	42
圖 35. 姿勢辨識結果示意圖(FRAME 30與FRAME 40)	43
圖 36. 姿勢辨識結果示意圖(FRAME 50與FRAME 60)	44
圖 37. 姿勢辨識結果示意圖(FRAME 70與FRAME 80)	45
圖 38. 跌倒偵測模型混淆矩陣	49














表目錄
表 1. 模糊規則表	34
表 2. 資料記錄	38
表 3. 動作型態表	39
表 4. 跌倒各閥值之準確率	48

參考文獻
[1] United Nations, “Number of People Over 60 Years Set to Double by 2050,” 2022.
[2] S. Gates, J. D. Fisher, M. W. Cooke, Y. H. Carter, and S. E. Lamb, “Multifactorial Assessment and Targeted Intervention for Preventing Falls and Injuries among Older People in Community and Emergency Care Settings: Systematic Review and Meta-analysis,” BMJ, vol. 336, no. 7636, pp. 130–133, Jan. 2008.
[3] M. Mubashir, L. Shao, and L. Seed, “A Survey on Fall Detection: Principles and Approaches,” Neurocomputing, vol. 100, pp. 144–152, Jan. 2013.
[4] D. Dar, R. Alam, M. E. El Najjar, F. Aparicio, and M. A. Khalil, “Elder Tracking in Healthcare Environments: State of the Art and Trends,” IEEE Access, vol. 7, pp. 62114–62124, May 2019. 
[5] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi,“You Only Look Once: Unified, Real-Time Object Detection,”in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Las Vegas, NV, USA, 2016, pp. 779–788. 
[6] Google, “Mediapipe: A Cross-Platform Framework for Building Multimodal Applied Machine Learning Pipelines,” Online. Available: https://mediapipe.dev/, accessed on: Jan. 2020. 
[7] S. Shah, H. Nguyen, A. Bagavathi, V. Pallipuram, and S. Chatterjee, “Video Based Fall Detection Using Human Poses,” arXiv preprint, arXiv:2107.14633, 2021. 
[8] L. A. Zadeh, “Fuzzy Sets,” Information and Control, vol. 8, pp. 338–353, 1965. 
[9] T. M. Mitchell, “Machine Learning,” New York, NY, USA: McGraw-Hill, 1997. 
[10] I. Goodfellow, Y. Bengio, and A. Courville, “Deep Learning,” Cambridge, MA, USA: MIT Press, 2016. 
[11] C. M. Bishop, “Pattern Recognition and Machine Learning,” New York, NY, USA: Springer, 2006. 

[12] J. MacQueen,“Some Methods for Classification and Analysis of Multivariate Observations,”in Proc. 5th Berkeley Symp. Math. Stat. Probab., Berkeley, CA, USA, 1967, pp. 281–297.
[13] D. M. Blei, A. Y. Ng, and M. I. Jordan, “Latent Dirichlet Allocation,” J. Mach. Learn. Res., vol. 3, pp. 993–1022, 2003. 
[14] I. T. Jolliffe, “Principal Component Analysis,” 2nd ed., New York, NY, USA: Springer, 2002. 
[15] O. Chapelle, B. Schölkopf, and A. Zien, “Semi-Supervised Learning,” Cambridge, MA, USA: MIT Press, 2006. 
[16] R. S. Sutton and A. G. Barto, “Reinforcement Learning: An Introduction,” 2nd ed., Cambridge, MA, USA: MIT Press, 2018. 
[17] Y. LeCun, Y. Bengio, and G. Hinton, “Deep Learning,” Nature, vol. 521, pp. 436–444, 2015.
[18] J. Redmon and A. Farhadi,“YOLO9000: Better, Faster, Stronger,”in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Honolulu, HI, USA, 2017, pp. 6517–6525. 
[19] J. Redmon and A. Farhadi, “YOLOv3: An Incremental Improvement,” arXiv preprint, arXiv:1804.02767, 2018. 
[20] A. Bochkovskiy, C. Y. Wang, and H. Y. M. Liao, “YOLOv4: Optimal Speed and Accuracy of Object Detection,” arXiv preprint, arXiv:2004.10934, 2020.
[21] C.-Y. Wang, A. Bochkovskiy, and H.-Y. M. Liao, “YOLOv7: Trainable Bag-of-Freebies Sets New State-of-the-Art for Real-Time Object Detectors,” arXiv preprint, arXiv:2207.02696, 2022.