Multi-View Wireless Sensing via Conditional Generative Learning: Framework and Model Design

Zhaohui Yang, Ziqing Xing, Zhaoyang Zhang, Zirui Chen, Hongning Ruan, Zhaohui Yang 0001arXiv 2025 · DOI · arXiv

Introduction

As we transition toward 6G, the concept of Integrated Sensing and Communication (ISAC) aims to leverage existing wireless infrastructure for high-resolution environmental sensing. However, traditional sensing methods often struggle with "multi-view" data fusion and fail in "high-contrast" scenarios where targets differ significantly from their background. This paper introduces Gen-MV, a novel generative learning framework that transforms multi-view Channel State Information (CSI) into high-precision 4D point clouds.

System model for multi-view wireless sensing in a region of interest (RoI).

The Gen-MV Framework

The Gen-MV framework is built on a bipartite architecture designed to handle the complexities of wave propagation and target reconstruction. It consists of a Multi-View Channel Encoder and a Target Generator based on a conditional diffusion model.

The Gen-MV architecture featuring a multi-view channel encoder and a conditional diffusion model for target reconstruction.

The encoder extracts features from CSI using a physics-informed approach. A key innovation is the use of multiplicative positional embeddings, which better reflect how electromagnetic waves couple target features with transceiver positions compared to standard additive embeddings. The extracted features then condition a diffusion model to generate a 4D point cloud. Each point in the cloud contains spatial coordinates ($x, y$) and electromagnetic (EM) properties: relative permittivity ($\epsilon_r$) and conductivity ($\sigma$).

Physics-Informed Design: The IVT Encoder

The "secret sauce" of the encoder is the Interleaved-View Transformer (IVT). Unlike standard transformers that treat data as a flat sequence, the IVT treats multi-view data as a 2D grid (Base Station views vs. User Equipment views). By using alternating attention mechanisms across these dimensions, the IVT captures the intricate correlations inherent in multi-static sensing. This design allows the model to achieve "zero-shot" scalability, meaning it can handle a variable number of transceivers at inference time without retraining.

Results and Robustness

The researchers evaluated Gen-MV using datasets generated via the Method of Moments (MoM), a rigorous numerical technique for solving Maxwell’s equations.

  • Accuracy: Gen-MV significantly outperformed traditional baselines like the Born Iterative Method (BIM). In high-contrast scenarios ($\epsilon_r = 1.7$), where traditional methods typically fail and produce artifacts, Gen-MV maintained clear shape and accurate EM property reconstruction.
  • Encoder Performance: The proposed IVT encoder achieved a mean log-Chamfer Distance (log-CD) of -12.90 dB, outperforming standard MLPs (-9.30 dB) and BiLSTMs (-11.01 dB).
  • Resilience: The system proved robust against environmental clutter and low Signal-to-Noise Ratios (SNR), demonstrating its potential for real-world 6G deployments.

Conclusion

Gen-MV represents a paradigm shift in wireless sensing by combining physics-informed transformer architectures with generative diffusion models. By representing targets as 4D point clouds, the framework focuses on relevant features and overcomes the limitations of traditional weak-scattering approximations. Future research will explore extending this framework to 3D scenarios and dynamic environments involving high-speed mobility.

Summary updated

BibTeX
@misc{yang2025multi,
  author = {Zhaohui Yang and Ziqing Xing and Zhaoyang Zhang and Zirui Chen and Hongning Ruan and Zhaohui Yang},
  title = {{Multi-View Wireless Sensing via Conditional Generative Learning: Framework and Model Design}},
  year = {2025},
  eprint = {2505.12664},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.12664},
}
Figure