基於Transformer多軌跡融合之羽球擊球球種辨識模型

Abstract

擊球球種辨識(Shot Type Recognition)是球類運動的核心任務之一，傳統方法大多仰賴視覺的動作特徵來進行判別，例如透過影像擷取球員的RGB或骨架序列資訊，再輸入至動作辨識模型進行分析。然而對於節奏快速且變化細膩的羽球賽事來說，部份球種的擊球動作極為相似，且常伴隨著假動作，導致難以輕易透過動作去區分。為了突破上述的瓶頸，我們提出一種「基於 Transformer 多軌跡融合之羽球擊球球種辨識模型」，該方法不依賴動作特徵，而是整合多項時序軌跡，將每次擊球過程中所涉及的球飛行軌跡、球員移動軌跡以及球網、球場位置序列資訊，在通道維度拼接為一條時序融合向量，並利用自注意力機制(Self-Attention Mechanism) ，學習不同時間點融合特徵之間的關聯，以進行球種判別，研究結果顯示在AI -CUP 2023公開的羽球資料集上，本模型在九類球種辨識任務下取得96.59%的整體準確率，超越X3D等主流動作辨識模型的81.97%準確率，證實本研究方法不僅能有效提升擊球球種辨識的精度，也具體展現了僅使用純軌跡資訊即可預測球種的可行性，為未來應用於戰術分析奠定技術基礎。

Shot Type Recognition is one of the core tasks in racket sports. Traditional methods mostly rely on visual motion cues for discrimination—e.g., extracting players’ RGB frames or skeleton-sequence data from video and feeding these into an action-recognition model. However, in badminton matches, where the tempo is fast and the variations are subtle, some shot types exhibit nearly identical swinging motions and even include deceptive feints, making it hard to distinguish shots purely by their visual action patterns.To overcome the aforementioned limitations, we propose A Transformer-Based Multi-Trajectory Fusion Model for Badminton Shot Type Recognition. Unlike traditional methods that rely on motion features, our approach integrates multiple temporal trajectories—specifically, the shuttlecock’s flight path, the player's movement trajectory, and positional sequences of the court and net. These data are concatenated along the channel dimension into a unified temporal fusion vector. A self-attention mechanism is then employed to capture the relationships among the fused features across different time steps for stroke classification. Experimental results on the publicly available AI-CUP 2023 badminton dataset demonstrate that our model achieves an overall accuracy of 96.59% across nine stroke categories, significantly outperforming mainstream action recognition models such as X3D, which achieved 81.97%. These results confirm that our method not only enhances the accuracy of stroke recognition but also validates the feasibility of using purely trajectory-based information for stroke prediction, laying a solid technical foundation for future applications in tactical analysis.