Multi-Sensor Fusion-Based Indoor Single-Track Semantic Map Construction and Localization

Abstract

The advancement of smartphones has been a prerequisite for multi-sensor fusion-based indoor mapping due to their high penetration and low cost. This article constructs a sparse single-track semantic map by fusing waypoints, semantic landmarks, and Wi-Fi landmarks, which are scene-representable and navigational in unknown indoor environments. The pedestrian dead reckoning (PDR)-aided visual-inertial simultaneous localization and mapping (PDR-aided VI-SLAM) method uses the PDR velocity as an external observation to constrain the pre-integration of inertial measurements. The semantic objects detected by the YOLO V4 are pre-filtered by the proposed semantic object filtering algorithm before semantics matching. The centroid of feature points clustered by the random-sampling spatial clustering algorithm (R-DBSCAN) characterizes the location of a semantic landmark relative to the waypoints estimated by the PDR-aided VI-SLAM. To enhance the stability of Wi-Fi landmarks, Wi-Fi signals are processed by the proposed sliding window-based Wi-Fi fusion algorithm. By using waypoints as an intermediate quantity, the association relationships between waypoints, semantic landmarks, and Wi-Fi landmarks are established. After that, a lightweight single-track semantic map is constructed. A landmark matching-based localization method is proposed to evaluate the similarity between the local map and the constructed venue map to infer the location of a pedestrian in the venue map. Experiments are conducted in the office building and mall scenes under various illumination conditions and semantic density. Results demonstrate the high quality of the single-track semantic map and high-precision localization of a local map in the constructed venue map.