Identifying the physical properties of objects is a crucial technique in several industrial, security, and inspection applications. In this study, we propose a novel method for non-invasively estimating multiple properties of a target object, including scale, shape, and density, by leveraging channel state information (CSI) obtained from a commercial-off-the-shelf (COTS) Wi-Fi device. Our proposed method employs a multi-branch convolutional neural network (CNN) architecture that learns the intricate correspondence between ambient signal propagation and the target properties. The novelty of our method lies in its recursive mechanism: each network branch outputs the classification results for a specific property, which then serves as input to other branches. This multi-branch and recursive strategy allows for a comprehensive understanding of the target, as the estimation of each property is refined by mutually referencing the results of the others. We evaluate our proposed method using a tabletop testbed with COTS Wi-Fi devices and 3D printed objects. We compare our method with vision-based and CSI-based baselines, and demonstrate that our method achieves a better classification accuracy of 99% for all object properties under various illumination conditions and the distance between the object and the Wi-Fi devices.
Ikemura et al. (Thu,) studied this question.