Abstract
We propose a 3D modelling method from multiple pairs of spherical stereo images. A static environment is captured as a vertical stereo pair with a rotating line scan camera at multiple locations and depth fields are extracted for each pair using spherical stereo geometry. We propose a new PDE-based stereo matching method which handles occlusion and over-segmentation problem in highly textured regions. In order to avoid cumbersome camera calibration steps, we extract a 3D rigid transform using feature matching between views and fuse all models into one complete mesh. A reliable surface selection algorithm for overlapped surfaces is proposed for merging multiple meshes in order to keep surface details while removing outliers. The performances of the proposed algorithms are evaluated against ground-truth from LIDAR scans.