A Radar System-Agnostic (RSA) Learning Architecture for Human Activity Recognition

In recent years, radar-based Human Activity Recognition (HAR) has been widely applied across various fields. However, differences in radar setup, such as frequency band and waveform of transmitted signals, across various radar devices may lead to data incompatibility, limiting the collaborative capabilities of multi-radar detection systems. To address this issue, this paper proposes a radar system-agnostic learning architecture for HAR. The framework enhances HAR performance by employing adversarial training of the Gradient Reversal Layer (GRL) and the Auxiliary Classifier Generative Adversarial Network (ACGAN), with the constraints of each module mutually reinforcing effectiveness. The proposed RSA architecture is evaluated through extensive experiments using radar datasets from three devices across different frequency bands, covering 11 types of human activities. The experimental results demonstrate that the algorithm performs well in single-domain and multi-domain scenarios. In single-domain training, HAR accuracy improves by at least 1.5% over the baseline. Multi-domain training significantly surpasses other methods, achieving approximately 97% accuracy with three domains. An ablation study further validates the contributions of the GRL and ACGAN components, confirming that their integration is essential for achieving optimal performance. These findings highlight the practicality and advantages of RSA for robust cross-radar frequency HAR.