Improving Badminton Action Recognition Using Spatio-Temporal Analysis and a Weighted Ensemble Learning Model

Incredible progress has been made in human action recognition (HAR), significantly impacting computer vision applications in sports analytics. However, identifying dynamic and complex movements in sports like badminton remains challenging due to the need for precise recognition accuracy and better management of complex motion patterns. Deep learning techniques like convolutional neural networks (CNNs), long short-term memory (LSTM), and graph convolutional networks (GCNs) improve recognition in large datasets, while the traditional machine learning methods like SVM (support vector machines), RF (random forest), and LR (logistic regression), combined with handcrafted features and ensemble approaches, perform well but struggle with the complexity of fast-paced sports like badminton. We proposed an ensemble learning model combining support vector machines (SVM), logistic regression (LR), random forest (RF), and adaptive boosting (AdaBoost) for badminton action recognition. The data in this study consist of video recordings of badminton stroke techniques, which have been extracted into spatiotemporal data. The three-dimensional distance between each skeleton point and the right hip represents the spatial features. The temporal features are the results of Fast Dynamic Time Warping (FDTW) calculations applied to 15 frames of each video sequence. The weighted ensemble model employs soft voting classifiers from SVM, LR, RF, and AdaBoost to enhance the accuracy of badminton action recognition. The E2 ensemble model, which combines SVM, LR, and AdaBoost, achieves the highest accuracy of 95.38%.