RFCanvas: Modeling RF Channel by Fusing Visual Priors and Few-shot RF Measurements

Abstract

Accurate and responsive simulation of radio frequency (RF) signal propagation is crucial for designing wireless systems operating in dynamic environments. Conventional ray tracing approaches struggle to accurately model the intricate geometries and material properties of objects that impact propagation. Recently proposed neural scene representations can learn such intricacies from RF data, but they treat the entire scene as implicit neural networks, necessitating retraining with a massive amount of RF data upon any environmental changes. In this paper, we propose RFCanvas, which fuses visual priors and RF measurements to achieve high accuracy for realistic scenes and be responsive to environmental changes. To ensure compatibility between visual priors and RF measurements, we introduce RFCanvas scene representations that model shapes and materials of substantial objects with tensorial fields and signed distance fields. We further extract motion information from visual priors to adapt RFCanvas scene representations to scene dynamics. RFCanvas is built upon an end-to-end optimization framework with differentiable RF simulation. Extensive evaluations across real-world wireless communication and sensing environments demonstrate RFCanvas's superiority over both existing methods.