Conformer-Based Human Activity Recognition Using Inertial Measurement Units

Area of Science:

• Computer Science
• Biomedical Engineering
• Signal Processing

Background:

• Human Activity Recognition (HAR) is crucial for IoT applications like healthcare and fitness.
• Inertial Measurement Units (IMUs) offer accurate motion and orientation data for HAR.
• Existing models like CNN, LSTM, and GRU have limitations in capturing complex temporal dynamics.

Purpose of the Study:

• To propose a novel Conformer-based model for enhanced HAR using IMUs.
• To leverage attention mechanisms within the Conformer architecture for improved temporal feature extraction.
• To evaluate the proposed model's performance against established HAR methods.

Main Methods:

• Development of a Conformer-based HAR model incorporating convolutional layers and attention mechanisms.
• Utilizing self-attention and residual connections within multiple Conformer blocks.
• Training and validation on publicly available datasets (WISDM and USCHAD).

Main Results:

• The Conformer-based model achieved superior accuracy compared to CNN, LSTM, and GRU.
• High accuracy rates of 98.1% on WISDM and 96% on USCHAD were recorded.
• Attention mechanisms with residual connections facilitated gradient flow and improved model convergence.

Conclusions:

• Conformer-based models represent a promising advancement for IMU-based HAR.
• The proposed architecture effectively captures temporal dynamics, leading to state-of-the-art recognition accuracy.
• This approach holds potential for real-time IoT applications in various domains.