A semi-automatic tool to georeference historical landscape images

Abstract

Smapshot is a web-based participatory virtual globe where users can georeference historical images of the landscape by clicking a minimum of six well identifiable correspondence points between the image and a 3D virtual globe. The images database is expected to grow exponentially. In a near future, the work of the web users will no longer be enough. To tackle this issue, we developed a semi-automatic process to georeference images. The volunteers will be shown only images having a maximum number of neighbour images in the matching graph. These neighbour images are the ones with which they share some overlay. This overlap is detected using the SIFT algorithm in a pairewise matching process. For an image pair made of a reference image with a known pose and a query image we want to georeference, we extracted the 3D world coordinates of the tie points from a digital elevation model. Then, by running a perspective-n-point algorithm after having geometrically tested the resulting homography between the two images, we compute the 6 degree of freedom pose, i.e. the position (X,Y,Z) and orientation (azimuth, tilt and roll angles) of the query image. The query image then becomes a reference and the georeference computation can be propagated more deeply in the graph structure.