A Deep Learning Method for Human Sleeping Pose Estimation with Millimeter Wave Radar.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Sleep posture monitoring is vital for individuals with sleep disorders.
• Current methods like contact sensors and cameras have limitations (discomfort, privacy concerns).
• Radar sensors offer a non-invasive, privacy-preserving alternative with deep penetration capabilities.

Purpose of the Study:

• To develop a deep learning-based method for recognizing human sleep postures using single-antenna Frequency-Modulated Continuous Wave (FMCW) radar.
• To introduce and evaluate the ResTCN architecture for capturing essential frequency and sequential features from radar signals.
• To provide a robust and accurate sleep pose recognition system that overcomes the limitations of existing technologies.

Main Methods:

• Utilized a single-antenna Frequency-Modulated Continuous Wave (FMCW) radar device to acquire human sleeping data.
• Developed a novel deep learning architecture, ResTCN, combining Residual blocks and Temporal Convolution Network (TCN).
• Extracted augmented statistical motion features from radar time series data for posture classification.

Main Results:

• Achieved an average classification accuracy of 82.74% in recognizing different sleeping postures.
• The ResTCN model effectively captured both frequency and sequential features from the radar data.
• Demonstrated superior performance compared to existing state-of-the-art methods in sleep pose recognition.

Conclusions:

• The proposed deep learning method using FMCW radar and ResTCN is effective for non-invasive sleep pose recognition.
• Radar-based systems present a viable, privacy-preserving alternative for sleep disorder monitoring.
• Further research can explore advanced signal processing and deep learning techniques for enhanced accuracy and broader applications.