本論文以駝背老人、坐輪椅的殘疾人、拄著盲杖的盲人為對象，識別這些需要幫助的人群。本研究主要基於谷歌Teachable Machine平台的姿勢識別來訓練模型，該模型適用於識別老年人、殘疾人和盲人的個體姿勢。 並且，我們基於識別結果的信心分數(Confidence Score)分析了模型背後的數據，並用它來探索每一人群姿勢的差異。
所提出的系統對於姿態辨識模型訓練與姿態數據分析探討步驟有四步：
第一步：先從谷歌Teachable Machine平台中訓練週期次數(Epochs)、批量(Batch Size)、學習率(Learning Rate)可調參數來最佳化模型。
第二步：將老年人、殘疾人、盲人姿態透過Coral PoseNet數據化。
第三步：各別針對老年人、殘疾人、盲人數據化的資料加以統計與歸類。
第四步：則是以數據化來比較各族群差異與相異之處，並做小結來改善模型。
未來，通過整合17個關鍵點數據，每族群的識別率可以高於68%。如果該場域只需要識別老年人、殘疾人和盲人等有需要的群體與其他群體的二分法，識別率可以高達94%。

This thesis focuses on the elderly with hunchbacks, the disabled in wheelchairs, and the blind with a guide stick, to recognize the groups who need help. This research is mainly based on the posture recognition of the Google Teachable Machine platform to train the model, which acknowledges individual postures of the elderly, the disabled, and the blind. Moreover, we analyzed the data behind the model based on the confidence score of the recognized result and used it to explore the differences in the posture of each group.
   The proposed system has four steps for training the posture recognition model and analyzing the posture data: The first is to optimize the model from the Epochs, Batch Size, and Learning Rate adjustable parameters in Teachable Machine. Secondly, the step is to digitize the posture of the elderly, the disabled, and the blind through Coral PoseNet. The third is to collect statistics and categorize the elderly, the disabled, and the blind. The final step is to compare the differences and dissimilarities of each group based on 17 keypoint data and make a summary to improve the model.
In the future, the recognition rate of each group can be higher than 68% by incorporating 17 keypoint data. If the field only needs to identify the dichotomy between the group in need like the elderly, disabled, and blind persons and another group, the recognition rate can reach 94% high.

致謝	   I
中文摘要	  II
英文摘要	 III
目錄	  IV
圖目錄	  VI
表目錄	 VII

第一章 緒論	1
1.1 前言	1
1.2 研究動機	1
第二章 文獻探討	2
2.1 老年人、肢體障礙者(殘疾人)與視覺障礙者(盲人)	2
2.1.1 中華民國人口推估	2
2.1.2 視覺障礙與肢體障礙人口比例	3
2.2 姿態辨識先前相關技術	5
2.2.1 BlazePose	5
2.2.2 MediaPipe Pose	6
第三章 軟體與硬體介紹	8
3.1 硬體介紹	8
3.1.1 谷歌 Coral Dev Board 4GB	8
3.1.2 攝影機模組	9
3.2 軟體介紹	10
3.2.1 Teachable Machine	10
3.2.2 Coral PoseNet	11
第四章 系統架構設計及模型訓練與結果	13
4.1 系統架構設計	13
4.2 模型訓練簡介	16
4.3 Teachable Machine Pose Project	17
4.4 透過PoseNet算出姿態分數	25
4.5 比較各族群關鍵點姿態得分與小結	27
第五章 結論與未來展望	34
5.1 結論	34
5.2 未來展望	35
參考文獻	36
 
圖目錄
圖2.1 三階段年齡人口變動趨勢	2
圖2.2 高齡化時程	3
圖2.3 2004~2020年盲人與殘疾人數量趨勢圖	5
圖2.4 以虛構的人物透過BlazePose去做兩個關鍵點的對齊	6
圖2.5 BlazePose各關鍵點標記	6
圖2.6 MediaPipe Pose	7
圖2.7 MediaPipe Pose真實世界的 3D 座標	7
圖3.1 谷歌Coral Dev Board 4GB	8
圖3.2 E-books W10 Webcam	9
圖3.3 谷歌Teachable Machine頁面	10
圖3.4 Coral PoseNet	11
圖3.5 Coral PoseNet 單姿態偵測演算法架構	12
圖3.6 利用Coral PoseNet來將姿態數值化	12
圖4.1 姿態辨識系統架構	13
圖4.2 模型訓練系統架構	14
圖4.3 經由PoseNet數值化姿態數據	15
圖4.4 模型辨識流程圖	15
圖4.5 訓練過程中觀察各參數與訓練狀況	18
圖4.6 使用額外測試集驗證該次訓練好的模型	19
圖4.7 各種每週期準確度曲線圖	19
圖4.8 各種每週期損失曲線圖	21
圖4.9 利用額外50張圖片去驗證各週期次數模型	22
圖4.10 只需二分法的場域模型訓練圖	23
圖4.11 各種週期次數模型的二分辨識正確率	24
圖4.12 訓練集中老年人姿態得分與平均得分	25
圖4.13 訓練集中視障者姿態得分與平均得分	26
圖4.14 訓練集中殘疾人姿態得分與平均得分	26
圖4.15 拄拐杖的老人[21]在Coral PoseNet的關鍵點圖	28
圖4.16 拿導盲杖的盲人[22]在Coral PoseNet的關鍵點圖	28
圖4.17 老年人與盲人各關鍵點得分	29
圖4.18 殘疾人[23]在Coral PoseNet的關鍵點圖	30
圖4.19 拄拐杖的老年人[24]在Coral PoseNet的關鍵點圖	30
圖4.20 殘疾人與老年人各關鍵點得分	31
圖4.21 左右手及左右手腕關鍵點得分	31
圖4.22 頭部關鍵點得分	32
圖4.23 左右膝關鍵點得分	32

表目錄
表2.1 2004~2020年視覺障礙者與肢體障礙者統計表	4
表4.1 各族群姿態平均得分與標準差	26
表4.2 各關鍵點對照表	27

[1]國家發展委員會，三階段年齡人口變化趨勢。取自https://www.ndc.gov.tw/Content_List.aspx?n=D527207EEEF59B9B
[2]國家發展委員會，高齡化時程。取自https://www.ndc.gov.tw/Content_List.aspx?n=695E69E28C6AC7F3
[3]衛生福利部統計處，2.3.1身心障礙者人數按類別及縣市別分。取自
https://dep.mohw.gov.tw/dos/cp-2976-61106-113.html
[4]V. Bazarevsky, I. Grishchenko, K. Raveendran, T. Zhu, F. Zhang, and M. Grundmann. (2020). BlazePose: On-device Real-time Body Pose tracking. Cornell University. Retrieved from https://arxiv.org/abs/2006.10204
[5]Google AI blog, On-device, Real-time Body Pose Tracking with MediaPipe BlazePose. Retrieved from https://ai.googleblog.com/2020/08/
on-device-real-time-body-pose-tracking.html
[6]Coral, Product＞Prototyping products＞Dev Board. Retrieved from https://coral.ai/products/dev-board/
[7]中景科技股份有限公司，產品分類-視訊類-E-PCC072。取自http://www.ebooks-cct.com.tw/product_show.php?id=72
[8]Teachable Machine, About. Retrieved from https://teachablemachine.withgoogle.com/
[9]GitHub, google-coral/project-posenet. Retrieved from https://github.com/google-coral/project-posenet
[10]ITREAD01.COM，深度學習: 學習率 (Learning Rate)。取自https://www.itread01.com/content/1541934373.html
[11]G. Papandreou, T. Zhu, L.-C. Chen, S. Gidaris, J. Tompson, and K. Murphy. (2018). PersonLab: Person Pose Estimation and Instance Segmentation with a Bottom-Up, Part-Based, Geometric Embedding Model. Cornell University. Retrieved from https://arxiv.org/abs/1803.08225
[12]S. Ioffe and C. Szegedy. (2015). Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. Cornell University. Retrieved from https://arxiv.org/abs/1502.03167
[13]H. Jeong. (2020). Feasibility Study of Google’s Teachable Machine in Diagnosis of Tooth-Marked Tongue. Department of Dental Hygiene, Daegu Health College, Daegu 41453, Korea. Retrieved from https://www.koreascience.or.kr/article/JAKO202009135419402.pdf
[14]Y. Bengio. (2012). Practical recommendations for gradient-based training of deep architectures. Cornell University. Retrieved from https://arxiv.org/ abs/1206.5533
[15]GitHub, teachablemachine-community/libraries/pose/. Retrieved from https://github.com/googlecreativelab/teachablemachine-community/tree/master/libraries/pose
[16]Q. Chen, C. Zhang, W. Liu, and D. Wang. SHPD: Surveillance Human Pose Dataset and Performance Evaluation for Coarse-Grained Pose Estimation. IEEE. doi: 10.1109/ICIP.2018.8451116
[17]Y. Chen, R. Du, K. Luo, and Y. Xiao. (2021). Fall detection system based on real-time pose estimation and SVM. IEEE. doi: 10.1109/ICBAIE52039.2021.9390068
[18]Yeh James，第3.4講：支援向量機(Support Vector Machine)介紹。取自https://medium.com/jameslearningnote/資料分析-機器學習-第3-4講-支援向量機-support-vector-machine-介紹-9c6c6925856b
[19]Y. Jiao, H. Yao, and C. Xu. (2020). PEN: Pose-Embedding Network for Pedestrian Detection. IEEE. doi:10.1109/TCSVT.2020.3000223
[20]X. Chen and G. Yang. (2018). Multi-Person Pose Estimation with LIMB Detection Heatmaps. IEEE. doi:10.1109/ICIP.2018.8451055
[21]Park Plaza, May 2020: Volume 12 Issue 5. Retrieved from https://www.park-plaza.org/wp-content/uploads/2020/05/may-2020-newsletter.pdf
[22]自由時報，社福人劉志枰績效獎金全捐 買200支導盲杖助視障。取自https://news.ltn.com.tw/news/life/breakingnews/2351882
[23]yahoo！拍賣，電動輪椅車折疊輕便老人殘疾人代步車四輪越野可坐便全自動多功能。取自https://tw.bid.yahoo.com/item/電動輪椅車折疊輕便老人殘疾人代步車四輪越野可坐便全-100562437745
[24]每日頭條，老了會駝背？跟著骨科醫生做，身姿筆挺活到老。取自https://kknews.cc/health/jn2x8q.html