Disentangled Representation Learning for mmWave Radar Action Recognition via Hilbert-Schmidt Independence Criterion Optimization

Millimeter-wave (mmWave) radar-based human action recognition (HAR) faces inherent limitations: raw radar signals are difficult to annotate due to their non-intuitive patterns, and motion features are inherently entangled with environmental clutter. We propose a disentangled representation learning framework that addresses both challenges through dual orthogonal projections and hybrid constraints. The dual projections decompose signals into action-specific and environment-associated subspaces, leveraging environmental labels as weak supervision for the latter. Feature independence is enforced via HSIC minimization, with adversarial suppression of residual dependencies in the action classifier. Evaluated on SAD_r2p1e5, a mmWave HAR dataset containing five annotated environments, our method achieves 10.74% higher accuracy than baseline methods, demonstrating that explicit disentanglement via HSIC and adversarial invariance effectively mitigates signal ambiguity.