Automated Geo-referencing of Images Dr. Ronald Briggs Yan Li GeoSpatial Information Sciences The University of Texas at Dallas
Spatial datasets from different sources need to be accurately aligned geographically in order to be viewed or analyzed together.
Image Geo-referencing –To align a raw image with a real world map coordinate system
Drawings (e.g. As-Builts) Satellite Imagery (e.g., Google map screenshot) Maps Historical photo Arial photo City of Dallas Street Centerline Reference Map Typical images & Reference Vector Map
Geo-Referencing
Manual Geo-Referencing Process Key step: Manually identify a set of corresponding linked pairs – i.e, Control Points Pairs (CPPs) Image Reference map
Issues Manually finding CPPs –Time Consuming – Tedious – Sometimes impossible –Must know a priori the approximate location
City of Dallas StreetCenterline.shp 68,000 street segments An unknown aerial photo The impossible: Finding the location ?
Scaled Skewed Image might be arbitrarily rotated And, of course, translated away from its real world origin More Problems – Distorted Images Which makes it even harder to identify the location
Objectives Automated geo-referencing solutions are in demand –Automatically find an unknown image’s location anywhere in a city, county, state …. –Identify the right transformation to correct an image’s deformation –A practical solution Error tolerant Accurate result Fast processing speed
Rationale Release users from tedious manual work Increase productivity significantly Handle cases that are impossible to be geo- referenced by hand Support high accuracy, consistency, and stability Batch processing
Transformations Handled Raster images can be under any combinations of –Similarity transformations Rotation Translation Uniform Scaling –Affine transformations Scaling Rotation Translation Skew Uniform scalingTranslationRotation Differential scalingSkew Similarity transformations +
Our Solution - TPPM Street intersections are plentiful and easy to identify General idea: Automatically search for corresponding CPPs in image and vector map –Topological Point Pattern Matching (TPPM)
Theory behind TPPM TPPM under similarity transformations Shape preserving property: relative distance and angles are invariant TPPM under affine transformations Area preserving property: Shape is not preserved, but the ratio of areas is constant θ= θ' OB/OA = OB' /OA' Area(OAB)/Area(OA'B' ) = c
Automated Scheme Extract intersection points from a vector map and calculate its TPP (a one time task) Extract intersection points from a raster image and calculate its TPP (topological point pattern) Compare image’s TPP with vector’s TPP to find candidate sets of CPPs Matching verification and optimal result generation Transform and resample Image Batch Processing TPP = Topological Point PatternCPP = Control Point Pairs
Application Implementation
Automated Geo-referencing Result Total RMS error: 6.23 for 14 CPPs Image points include false, missing, inaccurate points
Features Our algorithm and application can handle: –Very large vector map –Affine distorted images –Unknown location –Errors such as missing, spurious, inaccurate image points, or mismatched points Fast - Processing time is down to seconds for large geo-referencing area Achieves small total RMS ( “quality of the match”) Highly scalable. Requires only a small subset of image points for matching
Future Work Address CPP identification under higher order transformations (distortions). Incorporate automated point selection from the image. Adaptive pattern matching –Urban area vs rural area
Questions? Thank you! Ron Briggs: Yan Li: