DoRF++: Spherical Representation Learning over Doppler Radiance Fields for Robust Wi-Fi Sensing

Wi-Fi sensing has emerged as a privacy-preserving alternative to camera-based activity recognition. While recent standards like IEEE 802.11bf have standardized WLAN sensing, achieving robust performance across different users and environments remains difficult. Traditional methods often struggle with hardware impairments, phase noise, and the inherent randomness of multipath propagation in indoor settings. This paper addresses these challenges by introducing a geometry-aware framework that treats Wi-Fi channel state information (CSI) as a collection of virtual observations of human motion.

The authors introduce Doppler Radiance Fields (DoRF), a technique inspired by Neural Radiance Fields (NeRF) from computer vision. Just as NeRF reconstructs 3D scenes from 2D images, DoRF reconstructs a latent 3D motion sequence from 1D Doppler velocity projections extracted from Wi-Fi CSI.

To ensure high-quality input, the researchers developed a "common-RX" CSI phase sanitization method. By forming ratios between streams received at the same antenna but transmitted from different antennas, they effectively cancel out receiver-side synchronization errors and RF-chain mismatches that typically degrade phase-based sensing. Once the Doppler projections are extracted, DoRF models them as sparse virtual-camera views. These are then projected onto an equiangular grid on a unit sphere. Finally, the DoRF++ extension applies spherical Transformers to this representation, allowing the model to leverage rotational symmetries and uniform coverage for more accurate activity classification.

By moving beyond simple aggregation of Doppler estimates, DoRF++ provides a more robust and geometrically consistent way to interpret Wi-Fi sensing data. The experimental results on a hand-gesture dataset demonstrate that this approach significantly improves cross-user generalization compared to state-of-the-art methods. This is particularly important for real-world applications where the environment and user behavior vary, as it allows Wi-Fi systems to maintain high accuracy even when using a single multi-antenna receiver access point.