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

Area of Science:

• Artificial Intelligence
• Machine Learning
• Wearable Sensor Technology

Background:

• Human activity recognition using sensors is advancing Artificial Intelligence.
• Recognizing complex activities like badminton is challenging due to feature extraction difficulties.
• Current Convolutional Neural Network (CNN) approaches struggle with temporal data and global signal understanding.

Purpose of the Study:

• To develop an advanced deep learning framework for sensor-based badminton activity recognition.
• To address limitations of existing CNN-based methods in capturing temporal dynamics and comprehensive signal features.
• To improve the accuracy and efficiency of recognizing sophisticated human activities from sensor data.

Main Methods:

• Proposed a deep learning framework integrating convolutional layers, Long Short-Term Memory (LSTM) structure, and a self-attention mechanism.
• The framework automatically extracts local sensor signal features in the time domain.
• Utilized LSTM for processing badminton activity data and self-attention for focusing on essential information.

Main Results:

• Achieved a 97.83% accuracy in recognizing 37 badminton activities using a single sensor dataset.
• Demonstrated superior performance compared to existing sensor-based badminton activity recognition methods.
• Exhibited advantages in reduced training time and faster convergence.

Conclusions:

• The proposed deep learning framework effectively recognizes complex badminton activities from sensor data.
• The combination of CNN, LSTM, and self-attention overcomes limitations of traditional methods.
• This approach offers a more accurate, efficient, and faster solution for sensor-based activity recognition.