Pose-guided spatio-temporal Mamba graph network for vision-based Taekwondo action recognition

Fine-grained Taekwondo action recognition is difficult because decisive cues are brief, visually similar, and concentrated in lower-limb joints. This paper proposes PG-STMGNet, a pose-guided spatio-temporal Mamba graph network that combines a lightweight RGB Vision Transformer, a bounded adaptive skeleton graph encoder, a selective state-space temporal module, and a pose-confidence-conditioned gate. The revision clarifies that the individual backbones are adopted from established architectures, whereas the substantive contribution is their reliability-aware coupling: lower-limb graph residuals are bounded and normalized, and the fusion gate uses pose confidence to suppress unreliable skeleton evidence while preserving RGB appearance cues. On the private eight-class Taekwondo dataset, the existing experiments report 94.3% accuracy, 93.1% macro-F1, 93.7% precision, 92.6% recall, and 97.0% macro-AUC. The revised evaluation protocol additionally specifies subject-independent data partitioning, repeated robustness trials, public-benchmark transfer, loss-weight sensitivity, modern RGB-skeleton baselines, and end-to-end edge-device profiling. The expanded evaluation reports detailed dataset statistics, five-run robustness estimates, public-benchmark validation, loss-weight sensitivity, comparisons with recent RGB-skeleton models, pose-noise analysis, and end-to-end edge-device profiling.