Spatial Content-Based Scene Matching Using a Relaxation Method

Abstract

Scene matching is a fundamental task for a variety of geospatial analysis applications. As we move towards multi-source data analysis, constantly increasing amounts of generated geospatial datasets and the diversification of data sources are the two major forces driving the need for novel and more efficient matching solutions. Despite the great effort within the geospatial and computer science communities, automated scene matching still remains crucial and challenging when vector data are involved such as image-to-map registration for change detection. In this context, features extracted from vector data contain no intensity information which typically is the significant component in current promising approaches for registration. This problem becomes increasingly complicated as the two or more datasets usually present differences in coverage, scale, or orientation in general, and accordingly corresponding objects in the two or more datasets may also differ to a certain extent. This dissertation developed a novel methodology for automatic image-to-vector matching, based on contextual information among salient spatial features (e.g. road networks and buildings) in a scene. In this work, we model the road networks extracted from the two to-be-matched datasets as attributed graphs. The developed attribute metric measures the geometric and topological properties of the road network, which are invariant to the differences of the two datasets in scale, orientation, area of coverage, physical changes and extraction errors. Road networks comprise line segments (or curves), intersections and loops. Such complex structure suggests versatile attributes derivable from the components themselves of the road networks as well as between these components. It is important to develop attributes that need less computational efforts, while having sufficient descriptive power. We extend the entropy concept to statistically measure the descriptive quality of the attributes under consideration. Subsequently, by considering the spatial distribution and structure similarity in a neighborhood, we formulate a global compatibility in a relaxation manner to measure the overall goodness of correspondence. An optimal matching is achieved by finding an optimal morphism that maximizes this compatibility function. In this work, we further extend the invariant metric to incorporate additional scene content (i.e. buildings) in the form of object configurations present within individual road network loops (e.g. as they may become available from other GIS layers). For the local similarity, we developed an assessment framework to quantitatively measure the similarity of spatial configuration, where there is no need for semantic information (e.g. names) for buildings, a prior information necessary for spatial scene similarity in many alternative approaches. By combining diverse but co-located pieces of information (e.g. roads and buildings) in an integrated process, this multilayer scene matching allows us to integrate information that may become available from different sources, better addressing the evolving needs of the geoinformatics community. This novel integration enables achieving matching under perplexing scenario where the structure of each intersection in networks is identical.