Enhancing Badminton Performance Analytics with CNN-LSTM Shot Recognition

This work proposed a CNN-LSTM model for classifying badminton strokes and recognizing badminton stroke types. The model draws on two signal properties that can be identified by CNN and one that LSTM can learn to pinpoint various shot types from smashes, clears, drops, drives, and net shots. The proposed model achieved an accuracy of 93% with the help of a vast set of high-definition videos of badminton matches. They obtained the precision at the 92%, the recall at the 93%, and F1-score at the 92%. The visualization of the confusion matrix revealed a fair amount of precision in categorizing specific shot types. More specifically, smashes were identified accurately 95% of the time, drops and clears 91% and 92%, respectively. Analyzing the ROC curve for each shot type, it was found that the area under the curve or AUC for each shot type was well above 0. 90, showing that the model can differentiate between the variety of shots. The accuracy of the training and validation curves was constantly increasing; the curve of the validation accuracy coincided with the training accuracy, which showed successful generalization. Reduced training & validation losses also further support the fact that the model is not overfitting too much & has good generalization. These results assert the effectiveness of the suggested hybrid model, further reinforcing that it has an extraordinary lead over standard models and models solely based on CNNs and LSTMs. These satisfactory statistical results assure its potential for more realistic usage in coaching, performance assessment, and other intelligent refereeing services.