Quantifying Shuttlecock Skill Progression via Multimodal Action Recognition and Dynamics‐Guided Trajectory Prediction

Abstract

The fundamental techniques of shuttlecock kicking are characterized by high speed, complex structure, and sensitivity to details, making it difficult for beginners to establish stable movement patterns within a short timeframe. This study proposes a multimodal feature fusion‐based action recognition and trajectory prediction framework to quantitatively evaluate the effectiveness of compensatory exercises in accelerating shuttlecock skill acquisition. The system comprehensively employs TrackNet, Faster R‐CNN, and HRNet to extract shuttlecock trajectories, player positions, and human pose information. It constructs two pose representation methods that keypoint coordinate and heatmap encoding, and unifies multi‐modal features through a multi‐core fusion strategy. Subsequently, sequence models including LSTM, GRU, and Transformer are employed to predict future hitting trajectories. Stability and consistency analyses are conducted through velocity decay fitting, landing point deviation distribution, and action confusion matrices. Experimental results show that compensatory training produced stable exponential decay constants ranging from τ = 0.71 to 0.84 s, indicating smoother energy dissipation and improved movement stability, while landing point deviations became more concentrated and action recognition accuracy increased significantly. These findings demonstrate that the proposed framework effectively enhances the regularity, predictability, and quantitative assessment of shuttlecock skill acquisition.