基於Transformer物件關聯模型應用於籃球賽事分析

Abstract

在籃球賽事分析中，準確識別持球者和判斷得分時機對於確定得分者是關鍵挑戰。傳統的分析方法，比如物件重疊度和相對距離測量，往往在識別持球和進球時刻面臨較高的誤判風險。為了解決這一問題，我們對本團隊先前提出的Transformer-based Object Relationship Finder（ORF）架構的輸入特徵進行了改進，重點關注了幾個關鍵因素：與球密切相關的球員、球員的姿勢，以及不同的物件類型。這一策略顯著提高了架構在識別複雜動作和搶球情況下的準確度，使得持球者的識別準確率從原來的80.79%提升至86.18%，有效地展示了精準特徵選擇的重要性。此外，我們還利用Transformer-based Object Relationship Finder架構來識別進球時機，並結合最後接觸球的持球者信息，從而有效地判斷得分者，相較於傳統方法我們將得分者準確率從63.89%提高到了87.50%，這一成績突顯了Transformer-based Object Relationship Finder在籃球分析中的強大效能和廣泛應用前景。最後，我們開發了一款整合了這些技術的應用工具。這不僅讓教練和分析師能更全面地理解比賽情況，還為未來的籃球研究和技術開發提供了堅實的基礎。

In basketball game analysis, accurately identifying the ball handler and determining the scoring opportunity is crucial for pinpointing the scorer. Traditional analysis methods, such as object overlap and distance measurement, often face a high risk of misjudgment in identifying ball handling and scoring moments.To address this issue, we improved the input features of the Transformer-based Object Relationship Finder (ORF) architecture previously proposed by our team, with a focus on several key factors: players closely associated with the ball, their postures, and different types of objects. This strategy significantly increased the accuracy of the architecture in identifying complex actions and ball contest situations, raising the accuracy of ball handler identification from 80.79% to 86.18%, effectively demonstrating the importance of precise feature selection. Moreover, we utilized the Transformer-based Object Relationship Finder architecture to identify the timing of scoring moments, combined with the information of the last player to touch the ball, thereby effectively determining the scorer. Compared to traditional methods, we increased the scorer identification accuracy from 63.89% to 87.50%, highlighting the strong performance and wide application prospects of the Transformer-based Object Relationship Finder in basketball analysis.Finally, we developed an application tool that integrates these technologies. This not only enables coaches and analysts to understand the game more comprehensively but also lays a solid foundation for future basketball research and technological development.