3D Modelling of Large and Narrow Indoor Spaces using Stereo Vision Aided Inertial Navigation and Semi-Global Matching

Abstract

Digital modeling of cultural heritage is usually performed using expensive and heavy equipment (e.g. laser scanners) to achieve 3D models that are accurate and true to scale. In indoor environments the process of measurement can be very time consuming, especially when the object to be examined are large, narrow and labyrinthine, like caves, historical mines, and other underground structures, but also narrow buildings and town centers, ships, deep excavation, etc. In such cases it is almost impossible to generate a 3d model with laser scanners with justifiable effort. This is because of the huge number of scans that is necessary to resolve most of the occlusions and to achieve the measurement overlap that is necessary for the adjustment of the single measurements. An alternative approach is to take many photos and use the so called Structure from Motion (SfM) approach to generate the 3D model directly from the images, using a photogrammetric approach. As a first step towards 3D modelling from photos, the poses where the photos have been taken from have to be determined. Usually this is performed using the image information itself, as shown [Woh13]. But in narrow environments this essential step of the approach becomes critical because the image mosaic becomes unstable due to the few and loose connections between the images. Another problem is that the computational complexity of this step becomes insurmountably high for large objects.