D2Sense: Decoupling Multi-Person WiFi Sensing via Virtual Path-Enhanced CSI Decomposition

Multi-person WiFi sensing is essential for ubiquitous real-world applications, yet inferring “who is performing what at which location” from highly mixed multipath CSI remains a formidable challenge. Existing solutions face a fundamental dilemma: statistical approaches extract patterns from aggregated mixtures but suffer from entangled feature, failing to yield physically meaningful per-target signals; conversely, geometric approaches pursue fine-grained parameters (AoA/ToF) but are limited by the resolution of commodity WiFi, resulting in unstable associations. To bridge this gap, we present D2Sense, a system that follows a new “decompose-to-sense” principle. D2Sense asks: can we separate superimposed CSI by target to reuse mature single-person sensing algorithms? While raw CSI lacks the theoretical guarantees for direct separation due to bandwidth limits and static multipath overlap, D2Sense overcomes this via virtual path-based target decomposition (VP-Tep). VP-Tep introduces a controllable virtual path to transform the CSI, theoretically enhancing the sensing signal-to-noise ratio (SSNR) and statistical resolution. This transformation ensures that distinct target subspaces are mathematically separable, enabling the recovery of disentangled per-target path clusters via principal component analysis (PCA). Extensive experiments demonstrate that D2Sense achieves 94.6% accuracy for activity recognition and a median localization error of 0.39 m in multi-person settings, effectively unlocking the reuse of existing single-person algorithms. Notably, by enabling valid signal decomposition, D2Sense improves recognition accuracy by 7.2× and reduces localization error by 14.6× compared to the direct application of PCA on CSI power (which suffers from a 5.7 m median error).