CSI2PC: 3D Point Cloud Reconstruction Using CSI

Wireless sensing research is underway to generate 2D images and 2D videos corresponding to an object or space using the measured amplitude and phase changes during RF signal propagation. The obtained 2D images and 2D videos can be used for object recognition and distance measurement based on image processing techniques. However, 2D images only contain visual information about the sensing target from a specific viewpoint. This paper proposes Channel State Information to Point Cloud (CSI2PC) to enable the observation of a sensing target from multiple viewpoints. CSI2PC generates a 3D point cloud corresponding to the 3D structure of the sensing target from Channel State Information (CSI), which stores the variation of amplitude and phase. CSI2PC generates a 3D point cloud from the measured CSI using a neural network (NN) architecture based on Generative Adversarial Networks (GAN) and Graph Neural Networks (GNN). To ensure that the generated point clouds accurately represent the sensing target, the proposed scheme designs 1) a two-stage learning of the proposed NN architecture and 2) a loss function considering the 3D point cloud reconstruction. Experimental results using consumer Wi-Fi devices show that the proposed CSI2PC can reconstruct a clean point cloud from the measured CSI and accurately classify the object using the point cloud-based classification model.