OpenPose-Inspired Reduced-Complexity CSI-Based Wi-Fi Indoor Localization

Inspired by the OpenPose model in computer vision, we propose a reduced-complexity deep learning (DL) model for indoor localization based on WiFi Channel State Information (CSI) preprocessed with a 2D IFFT and transformed into 2D heatmap images. We mitigate timing offset due to transmitter-receiver miss-synchronization for both one-way and two-way CSI scenarios. Compared to state-of-the-art, our method improves accuracy by 72% at $90<^>{th}$ percentile while reducing DL model size by more than 90%. To our best knowledge, this is the first DL model for WiFi indoor localization based on learning and generalizing from CSI features instead of fingerprinting.