Multi-Sensor fusion and semantic map-based particle filtering for robust indoor localization

Abstract

To enhance positioning accuracy and ensure long-term stability in large-scale indoor environments, this paper presents a robust particle filtering method that integrates IMU, magnetic field, Bluetooth, and semantic map data, implemented through a handheld portable device. In the particle prediction phase, based on peak candidates extracted in the time domain, the unintentional interference movement of the hand is removed through short-time Fourier transform in the frequency domain to reduce the noise of pedestrian dead reckoning. During the observation phase, we introduce a magnetic interference coefficient to model and quantify the degree of electromagnetic interference in different environments, thereby improving the accuracy of indoor magnetic headings and enhancing the alignment of particle swarms with real-world directional constraints. Furthermore, based on passable semantic information, the indoor map is segmented into distinct probability regions, allowing the particle filter to adaptively adjust the weight of the particle swarm when moving between regions, thus improving the robustness of the overall fusion system by preventing particles from entering infeasible areas. Experiments and validations were carried out in a large shopping mall and parking lot with multiple devices. Results demonstrate that the proposed method improves the average positioning accuracy by approximately 25% compared to the Bluetooth-only positioning system, with a 15% increase in cumulative distribution probability within 3 m.