依照體育專家的研究而知，一個人的四肢靈活度與肌力發展，和幼年的協調性訓練有很大的關係，可以藉由觀察立定跳遠的動作窺知一二。但是一個教師要觀看全部小朋友的姿勢是否正確需要花費很多時間，並且可能無法完全客觀的做判斷。因此我們希望發展出一套系統，可以自動地判斷受測者的姿勢哪裡不正確，進而根據不正確的姿勢給予受測者意見，使受測者可以了解自己動作的缺點而不需要專業的老師在側協助。

    本篇論文大致可以分為四大部分: 第一部份是將取出影片中的受測者。我們將此階段分為四個細項來處理：第一個項目是根據輸入的圖片序列重新建造背景；第二個項目是取出每個影格中的前景物件(亦即受測者輪廓)；第三個項目是去除雜訊；第四個項目是修補去除雜訊後的輪廓，使其破損程度降低。

    第二部分是預測骨架。我們採用細化演算法搭配圖學基礎運算，對細化後的初步結果做更進一步的修飾，刪除雜訊，保留骨架。這部份也是分為四個細項：第一個項目是對輸入的物件輪廓做細化，取得初步的骨架；第二個項目將初步的骨架轉換成圖(Graph)儲存；第三個項目是去除迴圈；第四個項目是去除不正確的骨架雜訊。

    第三部份是姿勢判斷。透過動態貝氏網路結構的建立與訓練，利用訓練好的網路對前面第二部份得到的骨架做識別，判斷出最可能的姿勢。這部份將分為三個細項：第一個項目是將骨架轉換成特徵向量；第二個項目是送入貝氏網路當中做訓練；第三個項目是使用訓練好的貝氏網路去找出使受測的特徵向量機率為最大的姿勢。
第四部份是動作改正建議。當判斷出來的姿勢中，有不合理或是屬於我們定義的「不正確」姿勢，我們可以將專家對這些錯誤姿勢的修正建議顯示給受測者，讓受測者可以即時知道該如何修正姿勢。這樣一來就如同教練就在身邊一樣，受測者可常常做自我練習而不用擔心練習了很久卻是錯誤的姿勢。

    這篇論文主要的貢獻是在於受測者真的可以透過這套系統，知道某幾種姿勢的修正方式。雖然目前可以辨識出的動作數量不是很多，但是相信在增加更多的參考資訊(如：更多的參考點，更準確的參考點計算方式以及更多的區塊)後，必定可以得到更佳的結果。

From the research of sport experts, body controlling and development of a person’s muscle are related to coordinative training when he/she was a child. If there is any pose which is not good enough, it may influence his body controlling attributes in the future. We can observe this problem from the progress of standing long jump. But it takes a lot of time to see if each child’s pose is good enough or not when there are many children. So we want to develop a system to recognize the defined wrong poses while the human is doing standing long jump, and to give him suggestions about the wrong poses.

  This thesis can be divided into four parts. First of all, extract the silhouette of the user from each frame. There are 4 steps in the first part. Firstly, we have to re-build the background from the input image sequence. Secondly, the foreground object is extracted from each frame by subtracting the background from each frame. Thirdly, to remove noise from the result that was generated in the second step. The last step is to fill the holes of the extracted silhouette.

  The second part is to find out the skeleton of the extracted silhouette. We use the thinning algorithm and the graph theorem to complete this work. Using the basic graph operations based on the graph theorem to refine the raw skeleton generated by the thinning algorithm to remove noise. There are also four steps of this part. First step is to thin the input silhouette to obtain raw skeleton. Secondly, raw skeleton should be converted into a graph structure, and the structure of the graph is refined by removing the adjacent junction vertices. Then, remove the loops of the graph to make sure that a simple path between any two vertices is the only single path always. Fourthly, the redundant branches are pruned to obtain the final skeleton.

  The third part is judging the pose of the skeleton. By constructing and training the Dynamic Bayesian Network. We can then use this network to recognize the skeleton, finding out the most possible pose of each frame. There are three steps of this part. First, convert each skeleton into a feature vector. Then, train the Bayesian network with the feature vectors obtained in the previous step. Finally, we test the feature vectors converted from the test data by finding out the maximum probability of each feature vector in the trained network.

  The last part is to output suggestions for detected poses that are incorrect. If there is any pose recognized in previous part defined as “incorrect”, we will show the experts’ suggestion for this pose. This actually helps users to know how to adjust the pose just like there is a coach standing aside.

  The contribution of this research is the user actually can improve his poses by using this system. Although, we define only four different “incorrect” poses to give suggestions. In the future, we will add more information (e.g. more reference nodes, more precise equation of finding out the reference nodes, and more partitions) to this system to obtain better results.

文目錄
第 1 章 緒論 -------------------------------------------------- 1 
1.1 研究動機 ------------------------------------------------- 1 
1.2 論文組織簡介 ------------------------------------------ 2 
第 2 章 文獻分析 ----------------------------------------------- 4 
2.1 動作分析 ----------------------------------------------- 4 
2.2 取出物件 ----------------------------------------------- 5 
2.3 骨架偵測 ----------------------------------------------- 6 
2.4 貝氏網路 -------------------------------------------------- 9 
第 3 章 系統架構 --------------------------------------------- 14 
3.1 簡介 ----------------------------------------------------- 14 
3.2 輪廓分割 ------------------------------------------------- 14 
3.2.1 背景建立 ----------------------------------------------- 14 
3.2.2 物件取出 ----------------------------------------------- 15 
3.2.3 修補破損 ----------------------------------------------- 17 
3.3 骨架建立 ------------------------------------------------- 17 
3.3.1 細化 --------------------------------------------------- 18 
3.3.2 去除鄰近連結點 ----------------------------------------- 21 
3.3.3 去除迴圈 ----------------------------------------------- 24 
3.3.4 去除雜訊 ----------------------------------------------- 25 
3.4 動作辨識 ------------------------------------------------- 29 
3.4.1 特徵向量之表示 ----------------------------------------- 29 
3.4.2 貝氏網路之建構 ----------------------------------------- 30 
3.4.3 訓練方式 ----------------------------------------------- 33 
3.4.4 測試方式 ----------------------------------------------- 35 
第 4 章 系統實作 --------------------------------------------- 38 
4.1 開發環境 ------------------------------------------------- 38 
4.2 實驗資料與方法 ------------------------------------------- 38 
4.3 實驗結果與討論 ------------------------------------------- 42 
第 5 章 結論與未來展望 --------------------------------------- 47 
附錄一   ----------------------------------------------------- 48 
附錄二   ----------------------------------------------------- 50 
參考文獻 ----------------------------------------------------- 80 
英文論文 ----------------------------------------------------- 84 


圖目錄
圖一、	動作分析系統架構圖 ------------------------------  4
圖二、	貝氏網路模型範例 -------------------------------- 10
圖三、	貝氏網路機率計算圖示 ---------------------------- 11
圖四、	(a)第一個影格 ----------------------------------- 14
圖四、	(b)最後一個影格 --------------------------------- 14
圖四、	(c)建立的背景 ----------------------------------- 14
圖五、	步驟四的結果 ------------------------------------ 16
圖六、	步驟五的結果 ------------------------------------ 16
圖七、	(a) 新演算法(TH=20) ----------------------------- 16
圖七、 (b) 原來的方法 (TH=0) ---------------------------- 16
圖八、	破損修補後 -------------------------------------- 17
圖九、	連結數示意圖 ------------------------------------ 19
圖十、	(a) 取出的物件 ---------------------------------- 19
圖十、	(b) 柔化後的結果 -------------------------------- 19
圖十一、	Z-S演算法流程圖 ------------------------------ 20
圖十二、	(a) 直接細化的結果 --------------------------- 21
圖十二、	(b) 保留最大骨架的結果 ----------------------- 21
圖十三、	(a) 電腦上的斜線示意圖 ----------------------- 22
圖十三、	(b) 圖的結構 --------------------------------- 22
圖十四、	「連接點」、「鄰近連接點」圖例 ----------------- 22
圖十五、	(a)「非連接點」圖例 -------------------------- 23
圖十五、	(b)「新連接點」圖例 --------------------------- 23
圖十五、	(c)取代並刪除後的結果 ------------------------ 23
圖十六、	圖的實際轉變 --------------------------------- 23
圖十七、	(a) 中間出現空白 ----------------------------- 24
圖十七、	(b) 細化後產生迴圈 --------------------------- 24
圖十七、	(c) 移除鄰近分支點後的結果 ------------------- 24
圖十八、	(a)移除鄰近點的結果 -------------------------- 25
圖十八、	(b)去除迴圈後的結果 -------------------------- 25
圖十九、	(a)簡單路徑 ---------------------------------- 26
圖十九、	(b)不是簡單路徑 ------------------------------ 26
圖二十、	(a)採用歐式距離 ------------------------------ 28
圖二十、	(b)採用新方法 -------------------------------- 28
圖二十一、	（a) 初步細化結果 ---------------------------- 29
圖二十一、	（b) 不是一次刪除一條分支 -------------------- 29
圖二十一、	（c) 一次只刪除一條分支 ---------------------- 29
圖二十二、	 特徵向量範例 -------------------------------- 30
圖二十三、	系統所使用的基本貝氏網路架構 ----------------- 31
圖二十四、	動作的時間關係範例 --------------------------- 32
圖二十五、	系統所使用的動態階層式貝氏網路 --------------- 34
圖二十六、	測試範例 ------------------------------------- 36
圖二十七、	特徵向量配對的情形 --------------------------- 36
圖二十八、	系統操作介面 --------------------------------- 39
圖二十九、	取出的物件 ----------------------------------- 40
圖三十、	細化得到的骨架 ------------------------------- 40
圖三十一、	去除鄰近連接點後 ----------------------------- 41
圖三十二、	去除迴圈的結果 ------------------------------- 41
圖三十三、	刪除分支 ------------------------------------- 41
圖三十四、	套回原圖的結果 ------------------------------- 41
圖三十五、	訓練的畫面 ----------------------------------- 42
圖三十六、	用此切割法可以增加更多的區域，讓資訊更多 ----- 43
圖三十七、	用新方法分割的結果 （a）膝蓋未彎曲 ------------- 44
圖三十七、	用新方法分割的結果 （b）膝蓋彎曲 --------------- 44



公式目錄
公式一、------------------------------------------------ 15
公式二、------------------------------------------------ 15
公式三、------------------------------------------------ 15
公式四、------------------------------------------------ 16
公式五、------------------------------------------------ 27

[1]	C.-L. Huang, H.-C. Shih, C.-Y. Chao, “Semantic Analysis of Soccer Video Using Dynamic Bayesian Network,” 749-760, Volume 8, Number 1, February 2006, IEEE Transactions on Multimedia (TMM), Vol. 8, 2006.
[2]	Hui-Huang Hsu, Sheng-Wen Hsieh, Wu-Chou Chen, and Chun-Jung Chen, “Motion Analysis for the Standing Long Jump,” in Proc. the 26th IEEE Int’l Conf. on Distributed Computing Systems Workshops (ICDCS Workshops 2006), Lisboa, Portugal, July 4-7, 2006.
[3]	 Sangho Park, J.K. Aggarwal, “Recognition of Two-person Interactions Using a Hierarchical Bayesian Network,” IWV
S’03, November 7, 2003, Berkeley, California, USA.
[4]	Ibrahim Karliga and Jenq-Neng Hwang, “Analyzing Human Body 3-D Motion of Golf Swing From Single-camera Video Sequenc
es,” Volume: 5,  page(s): V-V, ICASSP 2006.
[5]	Ying Luo, Tzong-Der Wu, and Jenq-Neng Hwang, “Object-based Analysis and Interpretation of Human Motion in Sports Video Sequences by Dynamic Bayesian Networks,” Computer Vision and Image Understanding, vol. 92 , Issue 2-3 , Pages: 196–
216, 2003.
[6]	 Changick Kim and Jenq-Neng Hwang,“Fast and Automatic Video Object Segmentation and Tracking for Content-Based Applications,” IEEE Transactions on Circuits and Systems For Video Technology, VOL. 12, NO. 2, FEBRUARY 2002.
[7]	Di Zhong and Shih-Fu Chang,“An Integrated Approach for Content-Based Video Object Segmentation and Retrieval,” IEEE Transactions on Circuits and Systems For Video Tec
hnology, vol. 9, No. 8, December 1999.
[8]	Sunil S Polmottawegedara, Ranjith Munasinghe, Asad Davari, “Tracking Moving Targets,” Proceedings of the 38th Sout
heastern Symposium on System Theory Tennessee Technologi
cal University Cookeville, TN, USA, March 5-7, 2006.
[9]	Joen-Woei Liaw, Tian-Yuan Shih, Kuen-Tzung Chang, “A Comparison of Thinning Algorithms,”http://nhmrc.nctu.ed
u.tw/People/tyshih/Publications/CadaE/Thinning2001.pdf
[10]	Bala´zs Ke´ gl and Adam Krzy _zak, “Piecewise Linear 
Skeletonization Using Principal Curves,” IEEE Trans
actions On Pattern Analysis And Machine Intelligence,
VOL. 24, NO. 1, JANUARY 2002.
[11]	Bal´azs K´egl, Adam Krzy˙zak, Tam´as Linder, Kenneth 
Zeger, “A Polygonal Line Algorithm for Constructing 
Principal Curves”, Proceedings of the 1998 conference 
on Advances in neural information processing systems 
II, Pages: 501 - 507 , 1999.
[12]	Ellis Hordwitz, Sartaj Sahni, Dinesh Mehta, “Fundamen
tals of DATA STRUCTURES IN C++,”Sixth printing, An 
imprint of W.H. Freeman and Company 
[13]	Arron R. Walker, Binh Pham, Miles Moody,“Spatial Bayes
ian Learning Algorithms for Geographic Information Re
trieval,” GIS’05, November 4-5, 2005,Bremen,Germany.
[14]	D.M. Chickering, “Learning Equivalence Classes of 
Bayesian-Network Structures,” Journal of Machine 
Learning Reserch, vol. 2, pp. 445-498, 2002.
[15]	G.F. Cooper and E. Herskovits, “A Bayesian Method for 
the Induction of Probabilistic Networks from Data,” 
Machine Learning, vol. 9, pp. 309-347, 1992.
[16]	Daniel Grossman, Pedro Domingos, “Learning Bayesian
 Network Classifiers by Maximizing Conditional likeliho
od,” Proceedings of the twenty-first international 
conference on Machine Learning ICML’04.
[17]	Zhang, S. and K. S. Fu., 1984, “A Thinning Algorithm 
for Discrete Binary Images”, Proceedings of the Inter
national Conference on Computers and Application. Beij
ing, China. 879-886.