NeRA: Neural Reflectance and Attenuation Fields for Radio Map Reconstruction

This paper studies the radio map reconstruction by exploiting the point cloud environmental information. We propose a deep-learning-based model, named Neural Reflectance and Attenuation Fields (NeRA), to represent the attenuation and the reflectance functions to describe the influence of the environment on electromagnetic waves. NeRA learns these two functions directly from the voxelized point cloud of the real environment using three multilayer perceptrons (MLPs) and involves the real reflecting process in the neural network cost function design. NeRA models interactions between the radio frequency (RF) signal and the environment by learning and utilizing the reflectance and attenuation fields, which enables NeRA to reconstruct the radio map precisely even in the scene where both the transmitter and receiver are movable. Experiments show that NeRA outperforms Neural Radio-Frequency Radiance Fields (NeRF<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>) and traditional ray-tracing methods in reconstructing radio maps with less training data, achieving a 0.6 dB (26%) improvement in mean absolute error compared to NeRF<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> when trained with only 20% of the measured data.