Autoencoder Models for Point Cloud Environmental Synthesis from WiFi Channel State Information: A Preliminary Study

TL;DR

The paper addresses the challenge of producing 3D environmental point clouds from WiFi CSI signals. It introduces a two-stage autoencoder framework that first learns a geometry-driven latent space via a PointNet autoencoder and then maps CSI into a matching latent space with a dedicated CSI encoder, enforcing g ≈ z through a latent-space alignment loss. Experimental results show substantial improvements over direct CSI-to-geometry regression in both Chamfer Distance and Earth Mover’s Distance, validating the cross-modal translation approach and its potential for wireless sensing and environmental mapping. The method highlights the value of modular, latent-space-grounded architectures for converting ubiquitous wireless data into structured 3D representations with practical sensing implications.

Abstract

This paper introduces a deep learning framework for generating point clouds from WiFi Channel State Information data. We employ a two-stage autoencoder approach: a PointNet autoencoder with convolutional layers for point cloud generation, and a Convolutional Neural Network autoencoder to map CSI data to a matching latent space. By aligning these latent spaces, our method enables accurate environmental point cloud reconstruction from WiFi data. Experimental results validate the effectiveness of our approach, highlighting its potential for wireless sensing and environmental mapping applications.

Figures (3)

• Figure 1: Architecture of the Proposed Framework for Generating Point Clouds from WiFi CSI Data. The proposed framework consists of two main components: a PointNet autoencoder with convolutional layers and a CSI encoder. The PointNet autoencoder is trained to generate point clouds from a latent space representation, while the CSI encoder is trained to map CSI data to a latent space that matches the PointNet autoencoder latent space. The PointNet autoencoder processes the input point cloud through multiple 1D convolutional layers and max pooling to extract hierarchical features, which are then aggregated into a global feature vector serving as the latent space representation. The decoder uses fully connected layers to reconstruct the point cloud from this latent space. The CSI encoder processes the amplitude and phase components extracted from CSI, through 1D convolutional layers to extract features and map them to a matching latent space. The alignment of these latent spaces enables the generation of accurate point clouds directly from WiFi CSI data.
• Figure 2: The point clouds generated from the images of acquired room. These point clouds have been used for training the CSI Encoder and to subsequently test it.
• Figure 3: The point clouds generated from the CSI data of a) the first, b) the second, and c) the third room, respectively.