- 學位論文
視覺式智慧型高爾夫揮桿動作姿勢分析系統
A Vision-Based Intelligent Golf Swing Posture Analysis System
摘要
全球參與高爾夫這項運動的人口數量正在逐步上升，根據世界高爾夫管理機構皇家古老高爾夫俱樂部(The R&A)公布2021年的全世界高爾夫球人數為6,660萬人，超越了2012年的6,160萬人來到歷史高點，可見高爾夫球己經成為全世界普及的運動。近年來運動科技興起，將運動與科技兩者相互結合，利用智慧化訓練能夠有效幫助運動員提升訓練品質並降低運動傷害發生。本研究以高爾夫運動為基礎，為避免高爾夫揮桿姿勢錯誤導致運動傷害，因此開發出一套視覺式智慧型高爾夫揮桿動作姿勢分析系統，讓使用者能夠隨時隨地將自身和教練兩者的高爾夫揮桿姿勢相互比較，可達到自行修正高爾夫揮桿姿勢之目的。 視覺式智慧型高爾夫揮桿動作姿勢分析系統輸入使用者之高爾夫揮桿影片以及教練之高爾夫揮桿影片進行高爾夫揮桿姿勢比對分析。本系統主要分為兩大步驟：高爾夫揮桿分解動作擷取以及三維人體模型姿勢比對分析。在第一步驟中，本研究使用輕量級網路ShuffleNetV2和循環神經網路Bi-GRU進行改良後擷取出使用者以及教練兩者的高爾夫揮桿八個分解動作。在第二步驟中，利用擷取出使用者以及教練兩者的高爾夫揮桿八個分解動作分別建構出可以表現出豐富人體資訊的三維人體模型，接著使用三維人體模型進行使用者以及教練的高爾夫揮桿姿勢比對分析。 本研究將高爾夫揮桿動作拆解成八個分解動作，依序是擊球準備(address)、起桿(toe-up)、上桿(mid-backswing)、上桿頂點(top)、下桿(mid-downswing)、擊球(impact)、送桿(mid-follow-through)以及收桿(finish)。本研究使用GolfDB資料集[Mcn19]所蒐集的高爾夫揮桿影片進行訓練及測試，實驗結果顯示高爾夫揮桿分解動作擷取之準確率為86.15%。另外，本研究採用之三維人體模型是由6,890個節點所組成的人體網格，該模型將人體分解成24個身體部位，實驗時利用該模型之擬真人體特性能夠更精準地判斷使用者及教練之高爾夫揮桿姿勢差異。如上所述，本研究所提出之視覺式智慧型高爾夫揮桿動作姿勢分析系統具有效性。
並列摘要
Global participation in golf is gradually increasing. The world golf management organization, has already announced that the number of golfers in the world is 66.6 million in 2021, which exceeds the number of 61.6 million in 2012 to all-time high. In recent years, with the rise of sports technology, and the use of intelligent techniques can effectively help athletes improve training quality and reduce sports injuries. This study proposes a vision-based intelligent golf swing posture analysis system allowing the user to compare the golf swing of himself and the coach with each other anytime and anywhere. The proposed system can achieve the purpose of correcting the golf swing by the user and avoid sports injury caused by wrong swing posture. The input of the proposed a vision-based intelligent golf swing posture analysis system is one user's golf swing video and one coach's golf swing video. The system is mainly provided with two stages: golf swing decomposition action extraction and golf swing action comparison analysis. In the first stage, this study modified a lightweight neural network (ShuffleNetV2) and a recurrent neural network (bidirectional GRU) to extract the eight decomposition actions of golf swings of both the user and the coach. In the second stage, the system uses the eight decomposition actions of the golf swing of both the user and the coach to construct 3D human models, respectively. The 3D human models are used to compare and analyze the golf swing of the user and the coach. This study decomposes the golf swing into eight decomposed actions which in the order of address, toe-up, mid-backswing, top, mid-downswing, impact, mid-follow-through and finish. This study uses golf swing videos collected by GolfDB dataset for training and testing and the experimental results show that the accuracy of golf swing determination action capture is 86.15%. In addition, the 3D human model used in this study is composed of 6,890 vertices. The above model decomposes the human body into 24 body parts. In the experiment, the characteristics of the 3D model can be used to more accurately judge the difference in golf swing posture of users and coaches. In conclusion, the vision-based intelligent golf swing posture analysis system proposed in this study is effective and robust.
參考文獻
延伸閱讀
- 賀育民、莊翰伯、陳祥和（2005）。高爾夫揮桿動作電腦化學習系統雛形之建構。人因工程學刊，7(1)，25-30。https://www.airitilibrary.com/Article/Detail?DocID=a0000549-200503-7-1-25-30-a
- 薛雅馨、張曉昀、蔡松樺（2016）。智慧型即時影像功能性動作評估系統之建構及其準確性。物理治療，41(4)，298-299。https://doi.org/10.6215/FJPT.2016.72.O16
- 王建智、楊沛峰、杜錦豐（2003）。高爾夫揮桿動作的多面向實驗概念與展望。大專高爾夫學刊，(1)，9-30。https://doi.org/10.29706/GS.200303.0002
- fu, L. J. (2022). A Study of the Methods and Techniques of Golf Swing Practice. 中華創新發展期刊, 6(1), 33-39. https://doi.org/10.30141/JID.202205_6(1).0004
- 郭癸賓、黃世育、林益亘（2013）。The Study of Golf Teaching System with Move Controller。屏東科大體育學刊，(2)，1-14。https://www.airitilibrary.com/Article/Detail?DocID=22276114-201303-201304010008-201304010008-1-14