Automatic Semantic and Geometric Enrichment of CityGML Building Models using HOG-based Template Matching

Semantically rich 3D building models give the potential for a wealth of rich geo-spatially-enabled applications such as cultural heritage augmented reality, urban planning, radio network planning and personal navigation. However, the majority of existing building models lack much if any semantic detail. This work demonstrates a novel method for automatically locating subclasses of windows and doors, using computer vision techniques including the histogram of oriented gradient (HOG) template matching, and automatically creating enriched CityGML content for the matched windows and doors. Good results were achieved for class identification with potential for further refinement of subclasses of windows and doors and other architectural features. It is part of a wider project to bring even richer semantic content to 3D geo-spatial building models.