Automatic semantic and geometric enrichment of CityGML 3D building models of varying architectural styles with HOG-based template matching

Slade, Jonathan 2018. Automatic semantic and geometric enrichment of CityGML 3D building models of varying architectural styles with HOG-based template matching. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF - Accepted Post-Print Version Download (11MB) | Preview
	PDF - Supplemental Material Restricted to Repository staff only Download (2MB)

Abstract

While the number of 3D geo-spatial digital models of buildings with cultural heritage interest is burgeoning, most lack semantic annotation that could be used to inform users of mobile and desktop applications about the architectural features and origins of the buildings. Additionally, while automated reconstruction of 3D building models is an active research area, the labelling of architectural features (objects) is comparatively less well researched, while distinguishing between different architectural styles is less well researched still. Meanwhile, the successful automatic identification of architectural objects, typified by a comparatively less symmetrical or less regular distribution of objects on façades, particularly on older buildings, has so far eluded researchers. This research has addressed these issues by automating the semantic and geometric enrichment of existing 3D building models by using Histogram of Oriented Gradients (HOG)-based template matching. The methods are applied to the texture maps of 3D building models of 20th century styles, of Georgian-Regency (1715-1830) style and of the Norman (1066 to late 12th century) style, where the amalgam of styles present on buildings of the latter style necessitates detection of styles of the Gothic tradition (late 12th century to present day). The most successful results were obtained when applying a set of heuristics including the use of real world dimensions, while a Support Vector Machine (SVM)-based machine learning approach was found effective in obviating the need for thresholds on matchscores when making detection decisions.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Computer Science & Informatics
Date of First Compliant Deposit:	10 May 2018
Last Modified:	12 Apr 2021 15:13
URI:	https://orca.cardiff.ac.uk/id/eprint/111334

Actions (repository staff only)

Edit Item