Self-Attention-Based Deep Convolution LSTM Framework for Sensor-Based Badminton Activity Recognition

Key Points

• A new self-attention-based CNN-LSTM hybrid deep learning framework (SADeepConvLSTM) is proposed for badminton activity recognition, which combines the advantages of convolutional, recurrent, and self-attention layers.
• The SADeepConvLSTM achieves a recognition accuracy of 97.83% on a specialized Badminton Single-Sensor (BSS) dataset, outperforming existing popular deep learning models and conventional machine learning algorithms.
• The SADeepConvLSTM has the advantages of lower training time and faster convergence compared to other deep learning models, making it suitable for lightweight deployment.

Detailed Summary

• Introduction: Sensor-based badminton activity recognition aims to classify badminton activities using sensor data. Conventional machine learning methods have limitations in feature extraction, while deep learning approaches have shown promising results. However, existing deep learning models often lack the ability to capture temporal features and global signal comprehension.
• Proposed Framework: To address these limitations, the authors propose the SADeepConvLSTM framework, which consists of convolutional layers to extract local features, recurrent layers (LSTM) to model temporal dependencies, and self-attention layers to capture crucial context information. The framework is designed for lightweight deployment, with a focus on reducing training time and improving recognition accuracy.
• Experiments and Results: The SADeepConvLSTM framework is evaluated on a specialized BSS dataset containing sensor data from various badminton activities. The results demonstrate that the proposed framework significantly outperforms existing deep learning models and conventional machine learning algorithms, achieving a recognition accuracy of 97.83%. Additionally, the SADeepConvLSTM shows advantages in training time and convergence speed.
• Conclusions: The SADeepConvLSTM framework provides a robust and efficient solution for sensor-based badminton activity recognition. Its high accuracy, low training time, and lightweight design make it suitable for practical applications such as badminton AI coaches and rating systems.

Conclusions

• The proposed SADeepConvLSTM framework effectively combines convolutional, recurrent, and self-attention layers to improve badminton activity recognition.
• The SADeepConvLSTM achieves superior recognition accuracy and outperforms existing methods on the specialized BSS dataset.
• The framework's low training time and fast convergence make it suitable for lightweight deployment and real-time applications.