Room semantics inference using random forest and relational graph convolutional networks: A case study of research building
Xuke Hu, Hongchao Fan, Alexey Noskov, Zhiyong Wang, Alexander Zipf, Fuqiang Gu, Jianga Shang
Abstract
Abstract Semantically rich maps are the foundation of indoor location‐based services. Many map providers such as OpenStreetMap and automatic mapping solutions focus on the representation and detection of geometric information (e.g., shape of room) and a few semantics (e.g., stairs and furniture) but neglect room usage. To mitigate the issue, this work proposes a general room tagging method for public buildings, which can benefit both existing map providers and automatic mapping solutions by inferring the missing room usage based on indoor geometric maps. Two kinds of statistical learning‐based room tagging methods are adopted: traditional machine learning (e.g., random forests) and deep learning, specifically relational graph convolutional networks (R‐GCNs), based on the geometric properties (e.g., area), topological relationships (e.g., adjacency and inclusion), and spatial distribution characteristics of rooms. In the machine learning‐based approach, a bidirectional beam search strategy is proposed to deal with the issue that the tag of a room depends on the tag of its neighbors in an undirected room sequence. In the R‐GCN‐based approach, useful properties of neighboring nodes (rooms) in the graph are automatically gathered to classify the nodes. Research buildings are taken as examples to evaluate the proposed approaches based on 130 floor plans with 3,330 rooms by using fivefold cross‐validation. The experiments conducted show that the random forest‐based approach achieves a higher tagging accuracy (0.85) than R‐GCN (0.79).