A Novel Indoor Positioning System Utilizing Image-Based Relocalization and Pedestrian Dead Reckoning

High-precision positioning information is critical for indoor personal navigation. Current technology faces challenges, including complex infrastructure deployment and limited application scope, resulting in the lack of cost-effective, rapidly deployable, and user-centric indoor positioning systems. This study proposes a novel indoor positioning system that leverages the complementary strengths of inertial, visual, and map information to achieve continuous and robust global positioning. Specifically, the system constructs a scene-matching framework incorporating spatiotemporal dynamic constraints and single-frame probability estimation, ensuring optimal scene recognition and accurate camera pose estimation under visually challenging conditions. Additionally, prior visual information is fully exploited to extract the geometric and topological constraints of the environment, which are then integrated into the state estimation framework to enhance system reliability. Finally, multimodal data fusion is performed using Monte Carlo inference to eliminate accumulated errors, achieving optimal state estimation and ensuring the long-term tracking performance of the localization system. Experiments conducted in real-world indoor environments demonstrate that the proposed system outperforms existing methods in challenging scenarios, achieving a positioning accuracy of 0.441 m. Moreover, our system demonstrates significant application potential, enabling users to maintain continuous localization in challenging indoor environments using only a smartphone.