SELF-LOCALIZATION OF A MULTI-FISHEYE CAMERA BASED AUGMENTED REALITY SYSTEM IN TEXTURELESS 3D BUILDING MODELS

Abstract

Georeferenced images help planners to compare and document the progress of underground construction sites. As underground positioning can not rely on GPS/GNSS, we introduce a solely vision based localization method, that makes use of a textureless 3D CAD model of the construction site. In our analysis-by-synthesis approach, depth and normal fisheye images are rendered from presampled positions and gradient orientations are extracted to build a high dimensional synthetic feature space. Acquired camera images are then matched to those features by using a robust distance metric and fast nearest neighbor search. In this manner, initial poses can be obtained on a laptop in real-time using concurrent processing and the graphics processing unit.