Outdoor Environment Reconstruction with Deep Learning on Radio Propagation Paths

TL;DR

This work tackles outdoor environment reconstruction using ambient RF signals to enable lightweight mapping suitable for wearable devices. It develops two DL pipelines—a convolutional U-Net and a CLIP+-based Vision Transformer—fed by two RF input representations that encode AoA/AoD rays and per-path descriptors. On the WAIR-D synthetic dataset, the transformer-based approach with CLIP pretraining and full-layer integration delivers the strongest performance (IoU around $0.42$) while maintaining reasonable geometric fidelity, illustrating the feasibility of RF-driven 3D-like maps. The study highlights both the potential and current limits of RF-only reconstruction, emphasizing data bottlenecks and the value of scale-aware, pretrained transformer architectures for wireless-based spatial mapping with practical implications for low-power AR navigation.

Abstract

Conventional methods for outdoor environment reconstruction rely predominantly on vision-based techniques like photogrammetry and LiDAR, facing limitations such as constrained coverage, susceptibility to environmental conditions, and high computational and energy demands. These challenges are particularly pronounced in applications like augmented reality navigation, especially when integrated with wearable devices featuring constrained computational resources and energy budgets. In response, this paper proposes a novel approach harnessing ambient wireless signals for outdoor environment reconstruction. By analyzing radio frequency (RF) data, the paper aims to deduce the environmental characteristics and digitally reconstruct the outdoor surroundings. Investigating the efficacy of selected deep learning (DL) techniques on the synthetic RF dataset WAIR-D, the study endeavors to address the research gap in this domain. Two DL-driven approaches are evaluated (convolutional U-Net and CLIP+ based on vision transformers), with performance assessed using metrics like intersection-over-union (IoU), Hausdorff distance, and Chamfer distance. The results demonstrate promising performance of the RF-based reconstruction method, paving the way towards lightweight and scalable reconstruction solutions.

Figures (1)

• Figure 1: Outputs of the best model (CLIP-initialized ViT-L/14 with UPerNet decoder). White and grey pixels indicate buildings. Grey pixels indicate parts of the buildings not detected by our model (false negatives). Dark red pixels indicate false positives. Orange and blue crosses indicate locations of base stations and user equipments, respectively.