Automatically and accurately conflating road vector data, street maps and orthoimagery

• Authors: Cyrus Shahabi;Craig Knoblock;Ching-Chien Chen
• Venue: Automatically and accurately conflating road vector data, street maps and orthoimagery
• Year: 2005

Recent growth of the geospatial information on the web has made it possible to access various spatial data. By integrating diverse spatial datasets, one can support the queries that could have not been answered given any of these sets in isolation. However, accurately integrating different geospatial data remains a challenging task because diverse geospatial data may have different projections and different accuracy levels. Most of the existing conflation algorithms only handle vector-vector data integration or require human intervention to accomplish vector-raster or raster-raster data integration. In this dissertation, I propose an approach, named AMS-Conflation, that achieves automatic geospatial data integration by exploiting multiple sources of geospatial information. In particular, I focus on vector-imagery and map-imagery conflation. For vector-imagery conflation, I describe techniques to automatically generate control points by exploiting the information from the road vectors to perform localized image processing on the imagery. I also evaluate various filtering algorithms to eliminate inaccurate control point pairs. Based on the experimental results, these techniques automatically align the roads to orthoimagery, such that in one of my experiments, 85% of the conflated roads are within 4.5 m from the real road axes compared to 55% for the original roads for partial areas in St. Louis, MO. For map-imagery conflation, my approach can take a map of unknown coordinates and automatically align it with an image. My approach first aligns road vectors with imagery using vector-imagery conflation techniques to generate control points on the imagery. For the maps, my approach utilizes image processing techniques to detect intersections. Furthermore, I present an algorithm (called GeoPPM) to compute the matched point pattern from the two point sets. The experimental results show that GeoPPM only misidentified one point pattern from the fifty tested maps. The experimental results also show that my approach can align a set of TIGER maps with imagery for an area in St. Louis, MO, such that 85.2% of the conflated map roads are within 10.8 m from the real road axes compared to 51.7% for the original and geo-referenced TIGER map roads.