Efficient Accessibility Analysis in Floor Plans Using Semantic Segmentation

Accessibility features are critical for creating inclusive spaces and supporting diverse mobility needs in architectural design. Extracting these features from raster floor plans supports compliance checking, emergency evacuation planning, and renovation, which makes it a vital process in the architecture, engineering, and construction (AEC) industry. However, existing research on analysis of raster floor plans remains limited because raster plans pose challenges due to their lack of semantics and high variability. Furthermore, existing machine learning-based approaches rely heavily on large-scale datasets, which are scarce in the AEC industry.This paper proposes a novel approach to extract accessibility features from raster floor plans by segmenting and classifying room and door objects using as few as five reference samples. The approach employs similarity maps and a clustering algorithm to generate visual prompts for the Segment Anything Model to segment rooms and doors. The resulting masks are then classified by GPT-4 to facilitate accessiblity featrue extraction. Validated on the CubiCasa5K dataset and demonstrated through a case study, the method effectively addresses raster floor plan analysis challengeswith minimal data requirements.