Human Activity Recognition Method Based on FMCW Radar Sensor with Multi-Domain Feature Attention Fusion Network

Area of Science:

• Engineering
• Computer Science
• Signal Processing

Background:

• Human Activity Recognition (HAR) is crucial for various applications.
• Existing HAR methods using radar sensors often rely on single features, limiting performance.
• Frequency-Modulated Continuous Wave (FMCW) radar offers potential for non-intrusive HAR.

Purpose of the Study:

• To propose an advanced HAR method for FMCW radar sensors.
• To overcome the limitations of single-feature reliance in current HAR techniques.
• To enhance the accuracy and robustness of human activity classification.

Main Methods:

• Development of a Multi-Domain Feature Attention Fusion Network (MFAFN).
• Fusion of time-Doppler (TD) and time-range (TR) maps for comprehensive activity representation.
• Implementation of a multi-feature attention fusion module (MAFM) with channel attention.
• Application of a multi-classification focus loss (MFL) function for confusable samples.

Main Results:

• Achieved 97.58% recognition accuracy on the University of Glasgow dataset.
• Demonstrated performance improvement of 0.9-5.5% over existing HAR methods.
• Showcased significant improvement, up to 18.33%, in classifying confusable human activities.

Conclusions:

• The proposed MFAFN method provides a superior approach to HAR using FMCW radar.
• Multi-domain feature fusion and attention mechanisms enhance activity recognition capabilities.
• The method effectively addresses challenges in classifying similar human activities.