Inertial-aided sequential 3D metric surface reconstruction from monocular image streams

The self-acting digital reconstruction of 3D objects from monocular image streams, generally known as Structure-from-Motion (SfM), has been a subject of computer vision research for several decades. Most classical SfM approaches are off-line methods which implement a huge optimisation problem, based on a complete image sequence (often referred to as bundle adjustment (BA)). Such an iterative non-linear optimisation is very costly, in terms of computation time and cannot be used under real-time conditions. Recently, ideas from vision-based Simultaneous Localisation and Mapping (SLAM) were used to develop sequential-SfM frameworks for real-time applications. SLAM typically consist of two stages: the generation of an initial 3D scene model and then sequential SfM. BA requires an initial estimate, relatively close to the actual solution, to converge in a reasonable amount of time. This paper suggests a novel concept for sequential 3D scene reconstruction based on the integration of inertial measurements, as an aiding modality, in order to provide a reasonable initial guess for bundle adjustment. This new approach is able to outperform other techniques, in terms of accuracy and computational complexity.