Effects of Different Marker Set Configurations on Tennis Stroke Recognition Performance: A Three-Dimensional Kinematic Study Based on MiniRocket

Three-dimensional motion capture provides high-fidelity kinematic trajectories, but the extent to which different marker configurations retain discriminative information for tennis stroke recognition remains unclear. This study combined 3D motion capture with MiniRocket to compare seven upper-limb and racket marker configurations using data from 40 tennis-trained participants, 13 stroke types, and 10,133 valid movement samples. The configurations were the full-information group (ALL), full-arm group (AB), forearm group (FA), forearm + racket group (FR), upper-arm group (UA), racket group (RK), and watch group (WT). Model performance was evaluated using leave-one-subject-out cross-validation. All configurations achieved high performance in the three-class task. In the 13-class task, Macro-F1 was highest for ALL (78.76%), followed closely by FR (78.71%) and RK (78.28%). Holm-adjusted pairwise comparisons showed no significant differences among ALL, FR, and RK. FR significantly outperformed AB, FA, UA, and WT, whereas RK significantly outperformed FA, UA, and WT. Serve and overhead strokes were easiest to recognize, while several fine-grained baseline and net-play strokes showed lower F1-scores. These findings indicate that discriminative information is more strongly represented in the forearm-wrist-racket chain rather than simply increasing with marker count.