A Topological Semantic Mapping Method Based on Text-Based Unsupervised Image Segmentation for Assistive Indoor Navigation

Recently, much effort has gone into developing travel technology aids in indoor scenes, intending to increase the autonomy and quality of life of partially sighted or visually impaired (PSVI) people. However, existing indoor navigation methods require accurate prior maps. Yet, such maps are not always available and often difficult for people, especially the PSVI, to acquire or construct on site in real-time. To tackle such important issue, this work proposes a topological semantic mapping method for the PSVI such that they can well navigate such indoor environment as hospitals. The distinctive feature of this approach is its utilization of nonstandard floor plans, which can be easily obtained from real-world settings, as the primary data source. The proposed method that generating a sparse topological semantic map (TSM) from a floor plan includes: 1) preprocessing of captured floor plan images to extract a map area and correct a map view; 2) segmentation of an accessible area through a proposed text-based unsupervised image segmentation (TUIS) network; and 3) node calculation and semantic matching. Experimental results demonstrate the validity of the constructed TSM as well as TUIS's superiority to existing methods.