A WiFi indoor positioning method integrating dual-stream heterogeneous features and hierarchical support for local regression

Abstract

In complex indoor environments, WiFi RSSI signals are susceptible to obstruction, multipath effects, and changes in access point (AP) visibility. As a result, traditional fingerprint-based localization methods face challenges such as severe feature confusion, unstable local matching, and limited accuracy in continuous coordinate estimation in multi-building and multi-floor scenarios. To address these issues, this paper proposes a WiFi indoor positioning method that integrates dual-stream heterogeneous feature extraction, CBAM attention enhancement, hierarchical routing constraints, and support-point-based local regression. The method first partitions APs into operationally termed stable and unstable streams according to AP visibility and aggregate RSSI dispersion to extract complementary heterogeneous features. Subsequently, it constructs a three-channel input comprising absolute RSSI, relative RSSI, and visibility masks, and introduces 1D-CBAM to enhance key fingerprint representations after stream fusion; building upon this, it jointly models building, floor, and building-floor routing group information, constrains the candidate building-floor space through hierarchical priors, and achieves continuous coordinate estimation by combining support point metric retrieval, group-conditioned direct regression, local residual compensation, and gated fusion. Experimental results based on the official UJIIndoorLoc dataset show that the proposed method achieves a building identification accuracy of 99.64%, a floor identification accuracy of 90.01%, an average positioning error of 7.21 m, and a median positioning error of 4.34 m. The results demonstrate that this method achieves competitive positioning accuracy under the official UJIIndoorLoc multi-building and multi-floor validation protocol.