FMCW Radar Human Action Recognition Based on Asymmetric Convolutional Residual Blocks

Area of Science:

• Radar Signal Processing
• Machine Learning
• Human Action Recognition

Background:

• Optical and infrared video-based human action recognition is limited by environmental factors and complex feature extraction.
• Traditional methods struggle with feature extraction in challenging conditions.

Purpose of the Study:

• To propose an advanced human action recognition method using Frequency Modulated Continuous Wave (FMCW) radar.
• To overcome the limitations of existing methods in complex environments and feature extraction.

Main Methods:

• Utilized FMCW radar to extract micro-Doppler time-domain spectrograms of human actions.
• Developed an asymmetric convolutional residual network (ResNet18 variant) incorporating asymmetric convolution and Mish activation.
• Integrated an Improved Convolutional Block Attention Module (ICBAM) to enhance feature learning and data comprehension.

Main Results:

• Achieved a high action recognition accuracy of 98.28% in complex scenes.
• Demonstrated superior performance compared to classic deep learning approaches.
• Significantly improved recognition for actions with similar micro-Doppler features and showed excellent anti-noise performance.

Conclusions:

• The proposed FMCW radar-based method offers a robust and accurate solution for human action recognition.
• The asymmetric convolutional residual network with ICBAM effectively enhances feature learning and recognition accuracy.
• This approach provides a promising alternative for reliable human action recognition in diverse and challenging environments.