Informasi Lainnya
Abstraksi
<p>This study proposes a badminton motion classification system that leverages two-dimensional body landmark coordinates extracted using MediaPipe-based pose estimation. A CNN–LSTM architecture is employed to simultaneously model spatial posture information and temporal motion dynamics, allowing the system to accurately learn fine-grained movement patterns across time. The preprocessing pipeline consists of landmark coordinate normalization, sequence standardization to 40 frames, and velocity feature integration to capture inter-frame joint movement. The system classifies six badminton actions, including goodsmash, goodlob, goodserve, badsmash, badlob, and badserve. Experimental evaluation demonstrates a validation accuracy of 0.76 and a macro F1-score of 0.77, indicating balanced performance across motion categories and stable model convergence. Evaluations on real-game footage further confirm that the classifier reliably identifies correct and well-controlled movements, while variations in joint coordination lead to higher fluctuations in prediction confidence. These findings highlight the sensitivity of landmark-based analysis toward biomechanical consistency. By utilizing 2D landmarks rather than raw pixel data, computational requirements are significantly reduced, enabling lightweight and efficient deployment suitable for real-time applications on resource-limited devices. Overall, the proposed method provides an effective and practical approach for intelligent badminton motion analysis and holds promising potential for use in athlete training feedback systems, performance monitoring platforms, and sports technology solutions focused on enhancing skill development and biomechanical understanding.</p>
Koleksi & Sirkulasi
Tersedia 1 dari total 1 Koleksi
