| 摘要: | 本研究致力於推動電腦視覺在運動分析領域的應用，特別是對網球選手發球技術的提升，過去的發球訓練主要由教練進行的主觀評估，容易受到偏見的影響，因此引入自動化系統是必要的，通過先進的技術，我們提出了多通道發球輔助系統，旨在客觀評估和改進球員的發球技術。這種方法有助於球員優化技能，協助教練進行量身定制的培訓，同時促進全面的分析。研究的方法、實驗評估和結果凸顯了轉變網球發球分析的潛力，為球員和教練提供基於數據的見解。
AceNet 是一個我們提出的多通道深度神經網絡，主要用於分析和改進網球選手的發球技術。通過 MediaPipe從圖片中提取姿勢特徵，AceNet根據開放式動力鏈的關鍵角度計算新的骨架訊息特徵。這些特徵分別用於兩個通道的輸入數據，其中第一通道捕捉基本姿勢動態，處理球員骨架結構的特徵；第二通道專注於角度序列數據，通過分析關鍵點序列理解關節角度的時間變化。提取的特徵經過多通道的 LSTM 和 CNN 進一步處理，透過注意力機制加權先前模型提取的特徵，最後，一個全連接層整合這些特徵，為最終的分類任務做好準備，將發球分為不同的級別水平。AceNet 充分利用發球的時間和空間動態，實現對球員技術的細緻理解，多通道輸入和注意力機制提升分析的精確性和可解釋性，為優化發球表現帶來顯著潛在價值。
在所有評估指標上，多通道發球輔助系統神經網絡（AceNet）展現出卓越的性能，精確度、召回率、F1分數和準確度方面均取得卓越成績，並超越了所有比較方法，達到加權精確率、召回率、F1-score分別為81.98%、81.64％和 80.11%，強烈凸顯 AceNet在發球技術分析上的優越性，為深度學習在網球領域的應用提供有力支持，顯示了對運動員和教練在發球分析和技術提升方面的轉變潛力。透過客觀評估和對姿勢的感知，使運動員和教練能夠更有效地優化發球技術，進而提高整體運動表現水平，本系統的應用為運動領域引入了更深入的洞察和更有效的技術改進手段，有望在訓練和競技中產生積極的影響。
This study aims to enhance computer vision applications in sports analysis, specifically by improving tennis players' serving technique. Traditional subjective coach evaluations are prone to subjective biases, highlighting the need for automated systems. We propose a novel "Multi-Channel Serving Assistance System" called AceNet, which leverages innovative technology to objectively assess and enhance serving techniques. AceNet provides players with tools for skill optimization, assists coaches in tailoring training programs, and enables comprehensive performance analysis.
AceNet is a multi-channel deep neural network that extracts pose features from images using MediaPipe. It calculates new skeletal information features based on key angles of the open kinetic chain and processes these features through two separate input channels. The first channel analyzes the player's skeletal structure to capture basic pose dynamics, while the second channel analyzes keypoints sequences to understand the temporal variation of joint angles. Multi-channel LSTM and CNN further process the extracted features, employing an attention mechanism to prioritize the most relevant features. Finally, a fully connected layer integrates these features, categorizing serving into different quality levels. By maximizing the temporal and spatial dynamics of serving, AceNet provides a detailed understanding of players' technical skills. The multi-channel input and attention mechanisms significantly enhance the precision and interpretability of the analysis, providing valuable insights for optimizing serving performance.
AceNet outperforms all comparative methods across all evaluation metrics, including accuracy, recall, F1 score, and precision. The weighted precision, recall, and F1-score reach 81.98%, 81.64%, and 80.11%, respectively, demonstrating AceNet's superiority in serving technique analysis. This has transformative potential for athletes and coaches, offering deeper insights and more effective methods for technical improvement. AceNet promises positive impacts on training and competition within the sports domain. |
