Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

One Walk is All You Need: Data-Efficient 3D RF Scene Reconstruction with Human Movements

Yiheng Bian, Zechen Li, Lanqing Yang, Hao Pan, Yezhou Wang, Longyuan Ge, Jeffery Wu, Ruiheng Liu, Yongjian Fu, Yichao chen, Guangtao xue

TL;DR

This work tackles the data-inefficiency of static RF scene reconstruction by showing that a single 60-second walk can reveal occluded geometry when human motion is treated as informative. It introduces a two-stage framework based on composite 3D Gaussian Splatting that linearly combines a frozen static background with a sparse, learned dynamic perturbation, enabling high-fidelity PAS reconstruction with minimal data. Key contributions include the theoretical justification for motion-enabled observability, a practical two-stage training strategy, and extensive empirical validation showing substantial SSIM gains over heavily-sampled baselines and strong generalization to moving TX/RX tasks. The approach promises on-the-fly 3D RF mapping in unseen environments, reducing data collection bottlenecks and expanding the practicality of RF sensing for robotics and AR applications.

Abstract

Reconstructing 3D Radiance Field (RF) scenes through opaque obstacles is a long-standing goal, yet it is fundamentally constrained by a laborious data acquisition process requiring thousands of static measurements, which treats human motion as noise to be filtered. This work introduces a new paradigm with a core objective: to perform fast, data-efficient, and high-fidelity RF reconstruction of occluded 3D static scenes, using only a single, brief human walk. We argue that this unstructured motion is not noise, but is in fact an information-rich signal available for reconstruction. To achieve this, we design a factorization framework based on composite 3D Gaussian Splatting (3DGS) that learns to model the dynamic effects of human motion from the persistent static scene geometry within a raw RF stream. Trained on just a single 60-second casual walk, our model reconstructs the full static scene with a Structural Similarity Index (SSIM) of 0.96, remarkably outperforming heavily-sampled state-of-the-art (SOTA) by 12%. By transforming the human movements into its valuable signals, our method eliminates the data acquisition bottleneck and paves the way for on-the-fly 3D RF mapping of unseen environments.

One Walk is All You Need: Data-Efficient 3D RF Scene Reconstruction with Human Movements

TL;DR

This work tackles the data-inefficiency of static RF scene reconstruction by showing that a single 60-second walk can reveal occluded geometry when human motion is treated as informative. It introduces a two-stage framework based on composite 3D Gaussian Splatting that linearly combines a frozen static background with a sparse, learned dynamic perturbation, enabling high-fidelity PAS reconstruction with minimal data. Key contributions include the theoretical justification for motion-enabled observability, a practical two-stage training strategy, and extensive empirical validation showing substantial SSIM gains over heavily-sampled baselines and strong generalization to moving TX/RX tasks. The approach promises on-the-fly 3D RF mapping in unseen environments, reducing data collection bottlenecks and expanding the practicality of RF sensing for robotics and AR applications.

Abstract

Reconstructing 3D Radiance Field (RF) scenes through opaque obstacles is a long-standing goal, yet it is fundamentally constrained by a laborious data acquisition process requiring thousands of static measurements, which treats human motion as noise to be filtered. This work introduces a new paradigm with a core objective: to perform fast, data-efficient, and high-fidelity RF reconstruction of occluded 3D static scenes, using only a single, brief human walk. We argue that this unstructured motion is not noise, but is in fact an information-rich signal available for reconstruction. To achieve this, we design a factorization framework based on composite 3D Gaussian Splatting (3DGS) that learns to model the dynamic effects of human motion from the persistent static scene geometry within a raw RF stream. Trained on just a single 60-second casual walk, our model reconstructs the full static scene with a Structural Similarity Index (SSIM) of 0.96, remarkably outperforming heavily-sampled state-of-the-art (SOTA) by 12%. By transforming the human movements into its valuable signals, our method eliminates the data acquisition bottleneck and paves the way for on-the-fly 3D RF mapping of unseen environments.

Paper Structure

This paper contains 82 sections, 12 theorems, 87 equations, 6 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Background
  4. Primer on Electromagnetics
  5. Propagation Formulations
  6. Power Angular Spectrum (PAS)
  7. Preliminary:Why Human Motion Helps
  8. Human Body as Electromagnetic Relay
  9. Why Weak Scattering Suffices?
  10. Spatial Diversity from Motion
  11. Conclusion
  12. Method Design
  13. Overview
  14. Theoretical Foundation: Two-Stage Training Rationale
  15. Why Two-Stage Training?
  16. ...and 67 more sections

Key Result

Lemma 1

If $\mathbf{\Phi}_{\text{bg}}$ is estimated first using human-free data: this gradient contains no $\Delta\mathbf{\Phi}$ contamination, leading to more stable convergence. Subsequently, freezing $\mathbf{\Phi}_{\text{bg}}$ in Stage 2: focuses gradient flow exclusively on the residual perturbation without corrupting the learned background.

Figures (6)

  • Figure 1: An example of the power angular spectrum.
  • Figure 2: Human motion creates new RF propagation paths. The upper: no humans; the lower: with human movements. Although the body absorbs the signal, the scattered portion illuminates these otherwise invisible regions, and movement across multiple positions creates vast spatial diversity equivalent to thousands of virtual measurements.
  • Figure 3: Architecture of the system.
  • Figure 4: Overall PAS visualization at different positions.
  • Figure 5: $3$ scenes used in the experiments. The black square indicates the position of TX.
  • ...and 1 more figures

Theorems & Definitions (25)

  • Lemma 1: Gradient Cleanness via Separation
  • Lemma 2: Low-Rank Scattering Model
  • Definition 3: Three-Level SNR Hierarchy
  • Definition 4: Composite Optimization Problem
  • Proposition 5: Gradient Structure
  • proof
  • Theorem 6: Hessian Block Structure
  • proof
  • Definition 7: Two-Stage Strategy
  • Lemma 8: Stage 2 Gradient
  • ...and 15 more