Enhanced Badminton Stroke Recognition Using Hybrid RGB–Skeleton Features and Ensemble Learning

Previous studies on automatic recognition of badminton strokes have shown persistent gaps, particularly uneven classification accuracy across stroke categories and suboptimal overall performance. To overcome these limitations, this study proposes a hybrid framework that combines handcrafted spatiotemporal features with a two-stage feature selection and weighted ensemble learning. From representative video frames, RGB-based descriptors (Histogram of Oriented Gradients (HOG), Histogram of Optical Flow (HOF), Motion Boundary Histogram (MBH)) and skeleton-based features (Range of Motion Index (ROMI) for spatial and Dynamic Time Warping (DTW) for temporal) are extracted. Dimensionality reduction is applied through Autoencoder compression followed by SelectKBest to preserve the most informative features. The refined hybrid features are then classified using a weighted soft voting ensemble that integrates Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), and Adaptive Boosting (AdaBoost) to enhance both accuracy and class-wise balance. Experimental evaluation on a badminton stroke dataset demonstrates the effectiveness of this approach, achieving 98.21% accuracy. The results highlight that the best-performing feature combination is ROMI, DTW, and HOF, confirming that hybrid handcrafted features with ensemble Machine Learning (ML) significantly improve robustness and stability, offering strong potential for practical applications in performance analysis, training systems, and sport analytics.