U-Net Based Indoor Radio Map Prediction under Sparse Sampling

Tianxiang Xing, Leyi Zou, Tejas Bharadwaj, Rushabha Balaji, Danijela Cabric

Keywords: Radip map prediction, U-Net, sparse sampling, path loss

Abstract: In this paper, we present a runtime-efficient method for 2D path loss (PL) map prediction in complex indoor environments, based on the U-Net convolutional neural network. The proposed approach reconstructs full PL maps from sparse measurements in highly-cluttered environments. Using the radial symmetry of the PL, we construct several environment-aware geometrical features for the network to process. We empirically show that such features help the network not only generalize to unseen points in the same environment but to different environments as well. Some of these features include, the obstruction count map, accumulated transmittance maps, free-space path loss and log-scaled distance map, which are collectively used as input features to the network. Our method is evaluated in the context of MLSP 2025 The Sampling-Assisted Pathloss Radio Map Prediction Data Competition. The evaluation results demonstrate that the proposed method achieves a weighted final root mean square error of 4.80 dB with an average total runtime of 14 milliseconds.

Submission Number: 6