基于多阶段特征融合的跨模态室内定位方法OA

Although multi-modal fusion positioning can partially compensate for the shortcoming of single geomag-netic or Wi-Fi positioning,the positioning accuracy is severely limited by the heterogeneity of features between differ-ent modalities.A cross-modal indoor localization model named MamLoc based on multi-stage feature fusion is pro-posed to address the above issues.Firstly,a feature extraction unit called Mamba Block is constructed with selective state space model of Mamba as the core and introduces layer normalization and residual connection design to learn the hidden location correlation information in geomagnetic and Wi-Fi signals.Secondly,a cross-modal feature fusion mechanism based on multi-head attention and a multi-stage feature fusion method are proposed.At different stages of feature extraction,the geomagnetic and Wi-Fi features are fused twice to gradually construct cross-modal primary and advanced fusion representations.At the same time,through multi-head attention,the correlation information between cross-modal fusion features and single-modal features is further learned,enhancing the identification of cross-modal fu-sion features and reducing the impact of modal heterogeneity on localization accuracy.The experimental results show that the fusion positioning accuracy of MamLoc based on geomagnetic and Wi-Fi is relative improved by 13.8%and 58.7%respectively compared to single geomagnetic and Wi-Fi positioning accuracy.Moreover,the positioning accuracy of MamLoc is 13.7%higher than that of the CNN-LSTM positioning method based on transfer learning and 25.2%higher than that of the MLP-based positioning method.