Automatic Gigapixel Mosaicing in Large Scale Unstructured Underground Environments

We are currently developing a vision-based system aiming at automatically generating in situ ultra-high resolution mosaics in previously unknown, complex and unstructured underground environments. Nowadays, image stitching algorithms present several limitations when dealing with unstructured environments. The most important to our concern is the ability to deal with feature-less areas. In this paper we describe an automatic on line Gigapixel mosaicing system capable to deal with the absence of reliably detectable and trackable features. The input of our algorithm is a poseannotated sequence of high-resolution images acquired from a common optical center by a calibrated pan-tilt motorized digital camera unit. The proposed mosaicing system is powered by a global-to-local pairwise image alignment which recovers the rotations relating the overlapping images in a coarse-to-fine approach. The local motion procedure outputs a list of locally matched anonymous features which are later injected in a bundle adjustment engine for multi-view fine alignment. The proposed algorithm combines the state of the art mosaicing techniques in a complementary and efficient fashion providing an environment-independent solution for the image mosaicing task. The final output of the algorithm is a Gigapixel spherical mosaic rendering. Tests on real data acquired in a prehistorical cave (Tautavel Cave, France) illustrate mosaicing examples obtained from several hundreds of high-resolution images.