- 學位論文
基於姿勢辨識技術之籃球裁判手勢訓練系統之研究
Research on Basketball Referee Gesture Training System based on Pose Recognition Technique
摘要
對於剛上手沒有實務上經驗的初學者裁判來說，最基本的就是比裁判手勢動作要正確性，做出六十六種出籃球官方裁判手勢ORS ( Basketball official referee signals)的動作。目前籃球裁判手勢的動作，都是參考裁判書提供的圖片去學習，再進行判決實際操作時，可能會遇到比手勢出錯的狀況，或是姿勢不佳的問題。因此本論文提出開發一個訓練籃球裁判手勢的訓練系統，該系統提供給裁判做手勢的訓練。本方法使用影像辨識方法來建立籃球裁判手勢的訓練系統，拍攝了各種手勢的影片，從影片轉換成影像，並從影像截取姿勢的關鍵點，每一種手勢的影像資料都會在模型中不斷訓練，訓練好的模型會另外測試該模型的準確度。該模型會部署在籃球裁判手勢的訓練系統中，用於辨識手勢別，該系統可以提供訓練籃球裁判手勢。本實驗中只做三種手勢，分別是換人78筆、爭球112筆和犯規122筆的資料，總共用了364張影像去做訓練。實驗結果準確度得出，換人手勢高達93.8%，犯規手勢高達76.9%，爭球手勢高達58.3%。爭球準確率最低，因為手部動作很小，加上無法辨識大拇指，整體姿勢的關鍵點較不容易辨識。換人準確率很高，因為手臂交叉很明顯，姿勢範圍容易取得。犯規手勢因為跟很多其他手勢有類似的手臂舉起動作，所以準確率普通。未來需要克服手勢中較細微的手指動作，加強模型影像的訓練，讓模型更好的辨識出更多籃球裁判手勢。
並列摘要
For junior referees who are just getting started and have no practical experience, the most basic thing is to be more correct than the referee's gestures, and make 66 kinds of basketball official referee signals ORS (Basketball official referee signals). At present, the gestures of basketball referees are learned by referring to the pictures provided in the referee book. When making judgments, you may encounter errors in gestures or poor posture. Therefore, this thesis proposes to develop a training system for training basketball referee gestures, which provides training for referees to make gestures. This method uses the image recognition method to establish a training system for basketball referee gestures. It shoots videos of various gestures, converts the videos into images, and intercepts the key points of the gestures from the images. The image data of each gesture will be continuously displayed in the model. Training, the trained model will additionally test the accuracy of the model. The model will be deployed in the basketball referee gesture training system for gesture recognition, and the system can provide training for basketball referee gestures. In this experiment, only three kinds of gestures were made, which were 78 substitutions, 112 scrums and 122 fouls. A total of 364 images were used for training. The accuracy of the experimental results shows that the substitution gesture is as high as 93.8%, the foul gesture is as high as 76.9%, and the scrimmage gesture is as high as 58.3%. Scrimmage has the lowest accuracy rate, because the hand movement is very small, and the key points of the overall posture are not easy to identify due to the inability to recognize the thumb. Substitutions are accurate because the arms are clearly crossed and the range of poses is easily accessible. The foul gesture is mediocre because it involves raising the arm similar to many other gestures. In the future, it is necessary to overcome the subtler finger movements in gestures and strengthen the training of model images so that the model can better recognize more gestures of basketball referees.