- 學位論文
基於MediaPipe的深蹲訓練輔助系統
Squat Training Assistance System based on MediaPipe
摘要
在忙碌的現代生活中，保持良好的運動習慣對於健康至關重要。居家運動一直是許多人的選擇，不僅適合所有人，包括長輩和行動不便的人，近年來更受COVID-19疫情影響，使更多人選擇在家中進行運動，雖然缺少健身房中專業的器材，但仍然可以透過簡單的動作達到有效的訓練效果。即便是簡單的動作，也會因對動作的認知不足等情況使運動成效不佳，甚至是造成運動傷害，故本研究的目的為設計一基於影像的運動系統輔助一般民眾進行居家運動時，能夠有更好的成效，以及避免運動傷害。 市面上有許多針對居家運動所販售的訓練設備，雖說成效甚佳，但價格不菲，故不利於普及和推廣給大眾。為使居家運動能夠更為完善地普及於大眾之間，本研究將設計一深蹲訓練輔助系統。該系統將藉由OpenCV取得手機或電腦上的視訊鏡頭所拍攝出的訓練即時畫面，或將事先攝得的訓練影片作為輸入，再利用MediaPipe姿態估測模組，獲取全身33個三維骨骼關鍵點的座標，透過計算這些關鍵點之間的向量關係，就可以獲得人體各部位之間的角度變化。兩向量在空間中的角度，會隨著觀察點的不同而不同，於是在設計對於各階段動作的評判標準時，須同時考量所有觀察點上的角度。為了提高準確性，我們將通過對關鍵點座標的平滑化處理、可視度條件設定，以及座標間向量夾角初始設定等方法，這將有助於降低由環境、身體移動、受遮擋等因素而導致模型對運動中所輸出之座標點的誤差，與克服單一鏡頭拍攝造成的拍攝角度問題，最後將於螢幕上顯示當下的運動畫面與部分資訊，並生成相關的運動數據資料。
關鍵字
無並列摘要
It is crucial to maintain good exercise habits for health. Home workouts have always been a choice for many people. Although many home workout training equipment are effective, their high prices hinder their widespread adoption and promotion to the public. To make home workouts more widely adopted among the public, this study will design a squat training assistance system. This system will take the real-time training images captured by the video camera, use the MediaPipe pose estimation module to obtain the coordinates of 33 three-dimensional skeletal landmarks of the whole body, and by calculating the vector relationships between these landmarks, obtain the angle changes between different parts of the human body under different planes. In order to improve accuracy, methods such as smoothing landmark coordinate, visibility condition setting, and initial setting of the angle between coordinate vectors are used in reducing the errors in the coordinate points output by the model during exercise caused by factors such as environment, body movement, and occlusion, and overcoming the shooting angle problem caused by single-camera shooting. Finally, the current exercise image and some information will be displayed on the screen, and relevant exercise data will be provided subsequently.