STFLoc: ShuffleSKNet Network Based on Two-Stage Feature Fusion for Indoor Localization

Device-free indoor localization methods based on Channel State Information (CSI) have become an increasingly important technique. However, environmental noise and strong multipath effects significantly degrade the stability of CSI fingerprints and localization accuracy. In complex environments, effectively combining CSI amplitude and phase information to improve fingerprint stability and localization performance remains a critical challenge. To address these issues, we propose the STFLoc system, which integrates stable and multiscale CSI amplitude and phase information using a two-stage feature fusion method for indoor localization. We propose an unsupervised CSI Outlier Correction algorithm based on a Boxplot Adjustment of the Distance vector considering k nearest neighbors (OCBADk) to reduce the interference of environmental noise. To more effectively utilize both amplitude and phase information, we design a two-stage feature fusion approach. In the first stage, to mitigate multipath interference, the Jenks-based Nonuniform Quantization CSI Entropy (NQCE-Jenks) algorithm is proposed to select more stable features, which are fused to construct a CSI fingerprint image database. In the second stage, we design the ShuffleSKNet network based on ShuffleNet V2, which performs localization by fusing multiscale features and capturing the correlation between CSI amplitude and phase. Experimental results show that the proposed algorithm reduces the average distance error by 0.64m compared to existing methods, significantly improving localization accuracy.