A Framework for Real-Time Scene Modelling based on Visual-Inertial Cues

The self-acting generation of three-dimensional models by analysing monocular image streams from standard cameras is one fundamental problem in the field of computer vision. A prerequisite for the scene modelling is the former computation of camera poses for the different frames of the sequence. Several techniques and methodologies are introduced during recent decades to solve this classical Structure fromMotion (SfM) problem, which incorporates camera egomotion estimation and subsequent recovery of scene structure. Nevertheless the applicability of those systems in real world devices and applications is still limited due to non-satisfactorily properties in terms of computational costs, accuracy and robustness. This paper suggests a novel framework for visual-inertial scene reconstruction (VISrec!) based on ideas from multi-sensor data fusion (MSDF). The integration of additional modalities (here: inertial measurements) is useful to compensate typical problems of systems which rely only on visual information.