隨著科技的日新月異，手機所拍攝的影片，解析度也越來越高，雖然解析度有了大幅得提升，但所需的容量也越大，經過實測，以使用手機iPhone13為例，光是拍攝一個1小時的影片就約略需要4GB的容量，在資源有限的設備與網路下，會造成蠻大的負擔。
為了解決這個問題，本研究透過Google所開發的MediaPipe Holistic偵測球員的人體骨架，將原始影片的球員繪製成人體骨架並結合我們所繪製的虛擬場地、球網、球拍，以及利用TrackNetV2模型所繪製的球軌跡。綜合以上，得以保留原始影片的主角動作要素，將原始視訊影像轉化成以人體骨架動畫所展示的效果。
另外，我們也利用MediaPipe Holistic擷取人體33個特徵點數值，並結合LSTM模型進行訓練與預測，目的是為了辨識羽球的揮拍姿勢，其辨識的姿勢為以下五種：正手高遠球、反手高遠球、正手發球、正手切球和正手殺球。本研究的實驗結果顯示，模型準確率可達95.3%。
本研究致力於協助各羽球校隊，在沒有過多資金的情況下，可以利用手機拍攝球員訓練時對打的練習賽，來記錄練習賽狀況，並將影片進行壓縮，以減輕影片容量過大的負擔。之後再透過訓練深度學習模型的方式來協助初學者球員自行練習羽球基本揮拍姿勢。

With the rapid advancement of technology, the resolution of videos captured by smartphones is also increasing. Although the resolution has significantly improved, the required storage capacity has also increased. Based on our testing, using the iPhone 13 as an example, recording a one-hour video alone requires approximately 4GB of storage. This places a significant burden on devices and networks with limited resources.
To address this issue, in this study, we utilize MediaPipe Holistic, developed by Google, to detect the skeletal structure of players. We then render the players in the original video as skeletal animations and combine them with virtual courts, nets, and rackets that we create. Additionally, we use the TrackNetV2 model to generate the trajectory of the ball. By integrating these techniques, we preserve the essential player movements from the original video and transform the visual footage into animations showcasing the skeletal structure.
Furthermore, we employ MediaPipe Holistic to extract 33 feature point values of the human body. These values are then used in conjunction with an LSTM model for training and prediction to recognize badminton swing poses. The recognized poses consist of five types: forehand clear, backhand clear, forehand serve, forehand drop shot, and forehand smash. The experimental results of this study demonstrate an accuracy rate of 95.3% for the model.
This research aims to assist various badminton teams by enabling them to record practice matches of players training against each other using their smartphones, even with limited financial resources. The recorded videos can be compressed to alleviate the burden of large file sizes. Afterwards, we will employ deep learning models to assist beginner players in practicing basic badminton swing postures on their own.

摘要 i
ABSTRACT ii
誌謝 iii
目錄 iv
表目錄 vii
圖目錄 viii
一、 緒論  1
1.1 研究背景 1
1.2 研究動機  1
1.3 研究目的 2
1.4 論文架構 2
二、 文獻探討  3
2.1 遞歸神經網路(Recurrent Neural Networks, RNN) 3
2.1.1 遞歸神經網路(RNN)模型架構 3
2.2 長短期記憶網路(Long Short-Term Memory, LSTM) 4
2.2.1 LSTM架構 5
2.2.2 LSTM類型 6
2.3 MediaPipe 7
2.4 Mediapipe Holistic 7
2.5 BlazePose：即時身體姿勢追蹤 8
2.5.1 姿勢追蹤的ML管道 8
2.5.2 透過延伸BlazePose進行檢測 9
2.5.3 BlazePose實際應用 10
2.7 TrackNet 11
2.8 TrackNetV2 12
三、 研究方法 14
3.1 系統架構 14
3.2 影像數據集介紹 16
3.2.1 實驗一：人體骨架動畫-自製數據集 16
3.2.2 實驗二：羽球揮拍姿勢-自製數據集 16
3.3 實驗一：資料前處理-人體骨架繪製 19
3.3.1 利用Mediapipe Holistic繪製33個landmarks 20
3.3.2 將左右側影像還原為原始影片比例 23
3.3.3 將左右側影像合併 24
3.4 實驗一：資料前處理-場地線、球網 25
3.5 實驗一：資料前處理-利用TrackNetV2追蹤球軌跡 26
3.6 實驗二：資料前處理-取得人體骨架33個特徵點數值 28
3.7 實驗二：資料前處理-存取收集到的特徵點數值 29
3.8 透視變換(Perspective Transformation) 31
3.9 實驗二：資料擴增 31
3.10 實驗二：LSTM模型訓練 33
3.11 實驗二：模型驗證評估 34
四、 實驗結果 36
4.1 實驗研究環境 36
4.2 實驗一：實驗結果 39
4.2.1 檢視羽球追蹤結果 39
4.2.2 原始影像轉換為人體骨架影像之結果 42
4.2.3 影片大小比較 46
4.3 實驗二：實驗結果 48
4.3.1 LSTM模型訓練結果之分析 48
4.3.2 實際應用展示 51
五、 結論與未來工作 55
5.1 結論 55
5.2 未來工作 55
六、 參考文獻 56

[1]全台瘋小戴、麟洋羽球為什麼能成為台灣「新國球」？https://www.cw.com.tw/article/5117624
[2]Hopfield network
http://www.scholarpedia.org/article/Hopfield_network
[3]Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1985). Learning internal representations by error propagation. California Univ San Diego La Jolla Inst for Cognitive Science.
[4]遞迴神經網路(recurrent neural networks，RNN)
https://www.lalilaza.org/larry891001/2022/05/28/ML/RNN/
[5]Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural computation, 9(8), 1735-1780.
[6]以神經網絡進行時間序列預測—LSTM
https://ithelp.ithome.com.tw/articles/10273512?sc=rss.iron
[7]Understanding LSTM Networks
https://colah.github.io/posts/2015-08-Understanding-LSTMs/
[8]The Unreasonable Effectiveness of Recurrent Neural Networks
http://karpathy.github.io/2015/05/21/rnn-effectiveness/
[9]MediaPipe
https://developers.google.com/mediapipe
[10]Google AI
https://ai.google/
[11]MediaPipe Holistic — Simultaneous Face, Hand and Pose Prediction, on Device
https://ai.googleblog.com/2020/12/mediapipe-holistic-simultaneous-face.html
[12]Bazarevsky, V., Grishchenko, I., Raveendran, K., Zhu, T., Zhang, F., & Grundmann, M. (2020). Blazepose: On-device real-time body pose tracking. arXiv preprint arXiv:2006.10204.
[13]On-device, Real-time Body Pose Tracking with MediaPipe BlazePose
https://ai.googleblog.com/2020/08/on-device-real-time-body-pose-tracking.html
[14]Chen, H. T., Tsai, W. J., Lee, S. Y., & Yu, J. Y. (2012). Ball tracking and 3D trajectory approximation with applications to tactics analysis from single-camera volleyball sequences. Multimedia Tools and Applications, 60, 641-667.
[15]Wang, X., Ablavsky, V., Shitrit, H. B., & Fua, P. (2014). Take your eyes off the ball: Improving ball-tracking by focusing on team play. Computer Vision and Image Understanding, 119, 102-115.
[16]Fu, T. S., Chen, H. T., Chou, C. L., Tsai, W. J., & Lee, S. Y. (2011, November). Screen-strategy analysis in broadcast basketball video using player tracking. In 2011 Visual Communications and Image Processing (VCIP) (pp. 1-4). IEEE.
[17]Myint, H., Wong, P., Dooley, L., & Hopgood, A. (2015, May). Tracking a table tennis ball for umpiring purposes. In 2015 14th IAPR International Conference on Machine Vision Applications (MVA) (pp. 170-173). IEEE.
[18]Huang, Y. C., Liao, I. N., Chen, C. H., İk, T. U., & Peng, W. C. (2019, September). TrackNet: A deep learning network for tracking high-speed and tiny objects in sports applications. In 2019 16th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS) (pp. 1-8). IEEE.
[19]Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556.
[20]Noh, H., Hong, S., & Han, B. (2015). Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE international conference on computer vision (pp. 1520-1528).
[21]Sun, N. E., Lin, Y. C., Chuang, S. P., Hsu, T. H., Yu, D. R., Chung, H. Y., & İk, T. U. (2020, December). TrackNetV2: Efficient shuttlecock tracking network. In 2020 International Conference on Pervasive Artificial Intelligence (ICPAI) (pp. 86-91). IEEE.
[22]Pose landmark detection guide
https://developers.google.com/mediapipe/solutions/vision/pose_landmarker/
[23][Day 24] 激勵函數(Activation Function)
https://ithelp.ithome.com.tw/articles/10304438
[24]Action Detection for SignLanguage
https://github.com/nicknochnack/ActionDetectionforSignLanguage/blob/main/Action%20Detection%20Refined.ipynb
[25]Cricket Shot Identification
https://github.com/gowthambalachandhiran/CricketShotIdentification/blob/main/Cricket%20Shot%20Detection.ipynb
[26]Lin, Y., Jiao, X., & Zhao, L. (2023). Detection of 3D Human Posture Based on Improved Mediapipe. Journal of Computer and Communications, 11(2), 102-121.
[27]Kwon, Y., & Kim, D. (2022). Real-Time Workout Posture Correction using OpenCV and MediaPipe. The Journal of Korean Institute of Information Technology, 20(1), 199-208.
[28]Garg, S., Saxena, A., & Gupta, R. (2022). Yoga pose classification: a CNN and MediaPipe inspired deep learning approach for real-world application. Journal of Ambient Intelligence and Humanized Computing, 1-12.
[29]Yoshikawa, Y., Shishido, H., Suita, M., Kameda, Y., & Kitahara, I. (2021, March). Shot detection using skeleton position in badminton videos. In International Workshop on Advanced Imaging Technology (IWAIT) 2021 (Vol. 11766, pp. 312-317). SPIE.
[30]Promrit, N., & Waijanya, S. (2019, December). Model for practice badminton basic skills by using motion posture detection from video posture embedding and one-shot learning technique. In Proceedings of the 2019 2nd artificial intelligence and cloud computing conference (pp. 117-124).
[31]Hsu, T. H., Chen, C. H., Jut, N. P., Ik, T. U., Peng, W. C., Wang, Y. S., ... & Lin, Y. H. (2019, September). CoachAI: A project for microscopic badminton match data collection and tactical analysis. In 2019 20th Asia-Pacific Network Operations and Management Symposium (APNOMS) (pp. 1-4). IEEE.
[32]Chu, Z., & Li, M. (2021). Image recognition of badminton swing motion based on single inertial sensor. Journal of Sensors, 2021, 1-12.
[33]Shi, B., & Lv, B. Accuracy of badminton swing action recognition based on fractional time network. Journal of Intelligent & Fuzzy Systems, (Preprint), 1-10.