Badminton Coach AI：基於深度學習之羽球賽事資訊分析平台

Badminton Coach AI: A badminton match data analysis platform based on deep learning

緒論：科技結合體育是近年來的重要方向，其中透過大數據的蒐集與分析，可以有效瞭解運動員本身或對手的動作、跑位、戰術、以及比賽狀況等重要資訊，有效協助教練或運動員在訓練實務與競賽表現參考。然而，過去主要以人工標記分析方法進行羽球競賽資訊的蒐集，需要投入大量人力和時間成本，而無法有效且快速分析大量競賽影片，使得競賽資訊取得的立即性和精準度受到限制。因此，本研究目的為建構一套完整的大數據羽球分析平台Badminton Coach AI，進行賽局數據蒐集，進一步分析出客觀且具立即性的賽況資訊，供教練與運動員參考。方法：透過深度學習模組TrackNet、YOLOv3、OpenPose、影像處理技術與自行開發之演算法來偵測辨識球體、球員位置與球員骨架姿態等。結果：雛型系統包含四個模組：資料預處理模組、物件偵測與分割模組、數據分析與統計模組與視覺化模組，開發的技術能有效且精準的偵測到羽球飛行軌跡、擊球點、分割回合、得失分原因、主被動擊球與球種分類。結論：本研究已建置出一套具有效能的羽球競賽資訊分析平台Badminton Coach AI。據此研究結果，提出具體建議與研究方向，供實務訓練、競賽之運用與未來研究之參考。

Introduction: In recent years, information and communication technology has been widely utilized in sports science research and applications and has become a mainstream tool in sports learning. Microscopic data collection incorporated with big data analysis is an efficient and effective approach for assessing the performance of professional athletes in terms of factors such as motion correctness, movement agility, and tactical analysis. Massive data collection is essential for big data analysis. However, manual labeling is exhausting and time consuming, and in the past, this hindered the development of related research and applications. It should be noted, meanwhile, that computer vision has been applied in posture analysis and injury detection. In this work, deep learning and machine learning techniques are introduced to analyze broadcast video of badminton games in order to develop an automatic data collection system. Using the developed methods, it is possible to extract microscopic data from match videos in a timely manner right after the videos becomes available. As such, a data collection and tactical analysis platform for post-game reviews and tactical analyses called Badminton Coach AI was successfully developed using computer vision based data collection. Methods: Deep learning networks, including TrackNet, YOLOv3, and OpenPose, were adopted and applied to detect the shuttlecock, locate players, and predict player skeletons in every frame of the video images. Results: Four modules are proposed in this work, including the data preprocessing module, feature extraction and segmentation module, statistical analysis module, and visualization module. The developed prototype system can effectively and accurately depict shuttlecock trajectories, detect hit points, segment rallies, judge scores, and classify active or passive shots as well as stroke types. Conclusion: In this study, a complete, effective, and accurate badminton match information analysis platform called Badminton Coach AI was built. The results of the study can not only serve as a reference regarding tactics and training but also indicate potential directions for future research.