EchoNeRF: Time-Resolved RF Channel Synthesis with Neural Radiance Fields for Localization
Abstract
The availability of dense spatially consistent wireless channel information is a key premise for future deep-learning-driven wireless systems that exploit environmental knowledge for tasks such as network optimization or wireless localization. Acquiring these data is costly as it requires extensive measurement campaigns or running ray-tracing-based simulations that require highly accurate geometrical models of the environment. Neural radiance fields (NeRFs) are a promising alternative as they can learn implicit representations of 3D-scenes using only a few measurements. However, existing approaches consider only scalar signal quantities, preventing them from capturing the full spatio-temporal structure of multipath propagation. We propose a novel time-resolved wireless EchoNeRF that synthesizes channel measurements in the delay domain rather than single signal magnitudes, effectively capturing fine-grained multipath patterns. To make training efficient, we adapt the Instant-NGP framework to the wireless domain. In a realistic indoor digital twin, our approach synthesizes complex wireless channels with high fidelity (> 28 dB QdB), while remaining robust to sparse training data and noisy directional labels. The resulting synthetic data enhance data-hungry Transformer-based fingerprinting localization.