CNN-LSTM hybrid model achieves 93% badminton shot categorization accuracy

Badminton performance analytics reached a 93% classification accuracy milestone after researchers deployed a hybrid convolutional neural network and long short-term memory architecture during the 5th IEEE Global Conference for Advancement in Technology. Previously, the process of shooting identification in badminton relied mainly on basic attributes provided by designers and classical artificial neural network approaches. Even though these strategies provide some level of effectiveness, the approaches suffered from imprecision and unsustainability. Due to the complexity of badminton shots and the fact that they can change over time, higher-level data analysis techniques were required to capture spatial and temporal variations.

To address these issues, the study proposed a dual CNN-LSTM approach to improve model performance in recognizing various types of strokes. The convolutional neural network is utilized to extract spatial features and effectively recognize several features in different sections of videos. To isolate various strokes—including smashes, clears, drops, drives, and net shots—the system relies on one signal property that an LSTM can master alongside two distinct signal properties recognizable by a CNN.

Proposed Model Achieves 93% Accuracy in Badminton Shot Categorization

By leveraging a large collection of high-definition badminton matches, the newly developed framework successfully attained a 93% accuracy rate. Researchers obtained precision at 92%, recall at 93%, and an F1-score of 92%. The visualization of the confusion matrix revealed a fair amount of precision in categorizing specific shot types. Smashes were identified accurately 95% of the time, while drops and clears reached 91% and 92%, respectively. Examining the receiver operating characteristic curve for every category revealed that the area under the curve for each shot type remained well above 0.90, confirming the system’s proficiency in distinguishing between the diverse strokes.

Training and validation curves showed constantly increasing accuracy, and the validation accuracy matched the training accuracy, demonstrating successful generalization. Reduced training and validation losses further supported the fact that the model is not overfitting too much and possesses good generalization. These results assert the effectiveness of the suggested hybrid model, reinforcing that it holds an extraordinary lead over standard models and models solely based on CNNs and LSTMs.

Applications in Coaching and Refereeing

Widely played and heavily scrutinized, badminton is an increasingly prominent sport expanding globally through a growing fan base and expanding competitive championships. This increase in interest has created pressure to enhance performance and formulate working strategies in the sport. The ability to name and differentiate a range of badminton shots is crucial for achieving these aims. Identifying shot occurrences helps analyze players and is mandatory when adapting training to the right skills and developing automated refereeing systems.

These satisfactory statistical results assure the model’s potential for more realistic usage in coaching, performance assessment, and other intelligent refereeing services.