Enhanced Indoor Scene Recognition and Localization for Pedestrian Using Minimum Matching Group of Pedestrian Steps

Indoor scene recognition and transition detection are fundamental to pedestrian localization. To address the limited adaptability of traditional landmark-based methods in complex spatial layouts, diverse walking trajectories, and open entrance scenarios, this study proposes an indoor scene recognition and localization framework based on the Minimum Matching Group of Pedestrian Steps (MMG-PS).

Under the paper's experimental settings, static scene recognition accuracy is 88.89%–100%, scene transition recognition precision in office buildings and shopping malls is 87.50%–100.00%, and average localization errors in corridors and halls are 2.17 m and 1.34 m respectively.

Designed a five-step sliding-window MMG-PS scene transition recognition method that aggregates spatiotemporal features of consecutive gaits, reducing the impact of single-step observation errors and short-term signal fluctuations.

Built a random-forest-driven multi-modal scene recognition framework that fuses multi-source sensor features and uses scene-constrained cosine similarity for dynamic scene transition discrimination.

Proposed a scene-aware AP adaptive selection strategy that optimizes AP subsets for corridors and halls using PCA and ED algorithms to improve localization results in complex signal-propagation environments.

Static scene recognition accuracy reaches 88.89%–100%, and scene transition recognition precision reaches 87.50%–100.00% in office buildings and shopping malls.

Average localization errors in corridors and halls are 2.17 m and 1.34 m respectively.