Swin-Loc: Transformer-Based CSI Fingerprinting Indoor Localization With MIMO ISAC System

With multiple-input multiple-output (MIMO) technologies widely employed in mobile communication systems, wireless signals will have higher resolution in both the time and angle domains. It makes high-precision localization gain increasing attention in MIMO integrated sensing and communication (ISAC) systems. However, the decimeter-level precise indoor localization is full of challenges due to the multi-path fading and additional noise in indoor propagations. Excessive resource overhead and channel state information (CSI) fingerprint distortion in complex channel environments are the main factors that hinder indoor high-precision localization. To solve the CSI fingerprint distortion while reducing resource consumption, we propose a Swin Transformer-based CSI data-driven indoor localization framework called Swin-Loc. In the proposed Swin-Loc, a channel fingerprint extraction scheme is formulated to enhance the CSI features. Moreover, an improved Swin Transformer-based CSI network (SwinCSINet) model is proposed to improve localization precision. Experiments are conducted on the real-world CSI dataset given by the KU Leuven lab and the simulation CSI dataset generated by the DeepMIMO platform. Simulation results demonstrate that the localization precision on all datasets in the metric of root mean squared error (RMSE) is within 0.3 m, which outperforms current deep neural networks (DNNs) based schemeand attention-based scheme. Furthermore, the storage overhead of the improved SwinCSINet model is reduced to about 33<inline-formula><tex-math notation="LaTeX">$\%$</tex-math></inline-formula> of the DNN regression model, and the real-time performance is optimized.