Toward a Real-Time 3D Reconstruction System for Urban Scenes Using Georeferenced and Oriented Images

This paper presents an original 3D reconstruction technique for modeling urban scenes. Such modeling can lead many useful applications: virtual navigation, augmented reality, architectural planification, etc. Our approach uses a hybrid sensor combining a vision-inertial system and a GPS receiver to explore the urban environment and to store interesting data. Each captured image is annotated according GPS position and inertial sensor orientation and used iteratively for 3D reconstruction. This approach allows a metric reconstruction of the scene and thus solves the problem of the scale factor met in the projective reconstruction. Another contribution of our method is about the development of a robust markerless tracking approach. Natural and distinctive image features of the scene are detected and tracked frame-to-frame using the SIFT descriptors. We demonstrate the effectiveness of our approach through experiments on real data in outdoor environment.