| |
|
|
|
| |
|
|
|
|
|
|
|
|
| |
In our work [Automatically annotating and integrating spatial datasets], we focused on the problem of accurately integrating geospatial vector data with (satellite or aerial) images. One application for such integration could be for the purpose of automatic recognition and annotation of spatial objects in imagery. We utilized a wide variety of geospatial and textual data available on the Internet in order to efficiently and accurately identify objects in the satellite imagery. To demonstrate the utility of our technique, we built an application that utilizes the satellite imagery from the Microsoft TerraService and the Tigerline vector files from US Census Bureau (as well as some online sources) to annotate buildings on the imagery. |
|
| |
|
|
|
| |
The main challenge is that geospatial data (specifically, vector and image data) obtained from various data sources may have different projections, different accuracy levels, and different inconsistencies. The applications that integrate information from various geospatial data sources must be able to overcome these inconsistencies accurately, in real-time and for large regions. Traditionally, this problem has been in the domain of the image processing and GIS systems. However, the conflation approach (Saalfeld 1993) used in various GIS systems to manually or semi-automatically align two geo-spatial data sets does not scale up to large regions. Image processing techniques to identify objects in the image in order to resolve vector-image inconsistencies require significant CPU time and may result in inaccurate results. |
|
| |
|
|
|
| |
Figure 1. Effects and
Process of Conflation |
|
| |
|
|
|
| |
To explain our approach, we first need to explain the conflation process. The conflation process divides into following tasks: (1) find a set of conjugate point pairs, termed "control point pairs", in both vector and image datasets, (2) filter control point pairs, and (3) utilize algorithms, such as triangulation and rubber-sheeting, to align the rest of the points and lines in two datasets using the control point pairs. Traditionally, human input has been essential to find control point pairs and/or filter control points. Instead, we developed completely automatic techniques to find control point pairs in both datasets and designed novel filtering techniques to remove inaccurate control points. |
|
| |
|
|
|
| |
We developed two different techniques to find accurate control point pairs. Our first technique generates control points using localized image processing. The second technique finds control points by querying information from online web sources. We briefly describe the first technique, which relies only on the imagery and vector data for accurate integration. We find feature points, such as the road intersection points, from the vector dataset. For each intersection point, we perform image processing in a small area around the intersection point to find the corresponding point in the satellite image. The running time for this approach is dramatically lower than traditional image processing techniques due to localized image processing. Furthermore, the road directions information makes detecting edges in the image much easier problem, thus reducing the running time even more. |
|
| |
|
|
|
