Combining transformer with a latent variable model for radio tomography based robust device-free localization

Radio tomographic imaging (RTI) is a promising device-free localization (DFL) method for reconstructing the signal attenuation caused by physical objects in wireless networks. In this paper, we use the received signal strength (RSS) difference between the current and baseline measurements captured by a wireless network to achieve the RTI based DFL in a predefined monitoring area. RTI is formulated as solving a badly conditioned problem under complex noise. And the end-to-end deep learning method based on Transformers and latent variable models (LVMs) is considered to address the RTI problem. The data grouping strategy is designed to divide the RSS data into multiple spatially-correlated groups, and a Transformer-based convolutional neural network (TCNN) model is firstly developed for RTI, in which the Transformer blocks are able to help the model learn the more expressive feature for the environmental image reconstruction task. The RTI system is influenced by both sensor noise and environmental noise simultaneously. In order to improve the performance of the RTI method, a Transformer-based latent variable model (TLVM) is proposed further, where the robustness to interference can be enhanced by controlling the capacity of the latent variables. The comparative numerical experiments are conducted for RTI based DFL, and the efficacy of the proposed TCNN and TLVM based RTI methods is verified by the experimental results.