Radar-Based Cross-Domain Human Behavior Recognition Using Physics-Informed Data Augmentation and Multi-Source Domain...

Area of Science:

• Radar Signal Processing
• Machine Learning for Human Behavior Analysis
• Data Augmentation Techniques

Background:

• Falling is a critical safety concern, particularly for the aging population.
• Radar-based human behavior recognition, especially for falls, is extensively studied.
• Current methods suffer from poor generalization due to limited labeled data.

Purpose of the Study:

• To propose a novel physics-informed data augmentation (PIDA) strategy.
• To enhance domain generalization (DG) for human behavior recognition using micro-Doppler signatures (m-DS) from FMCW radar.
• To improve the accuracy and robustness of fall detection systems.

Main Methods:

• Developed PIDA, including Distance Data Augmentation (DDA) and Behavior Pattern Data Augmentation (BPDA).
• DDA simulates electromagnetic attenuation across ranges; BPDA diversifies motion styles while preserving kinematics.
• Integrated PIDA with a multi-source domain adversarial neural network (MSDAN) for transfer learning and invariant feature extraction.

Main Results:

• PIDA significantly improved average DG accuracy compared to baseline, traditional optical data augmentation (TODA), and generative data augmentation (GDA).
• On own datasets, PIDA improved accuracy by 1.52%, 2.48%, and 4.88% over baseline, TODA, and GDA, respectively.
• On public datasets, PIDA achieved improvements of 3.67%, 6.29%, and 9.78% over baseline, TODA, and GDA, respectively, with a small DG gap.

Conclusions:

• PIDA effectively addresses the data scarcity issue in radar-based human behavior recognition.
• The proposed method demonstrates superior domain generalization capabilities for fall detection.
• PIDA offers a promising approach for robust and accurate fall-related human behavior recognition systems.