Increasing the Convergence Domain of RGB-D Direct Registration Methods for Vision-based Localization in Large Scale Environments

Developing autonomous vehicles capable of dealing with complex and dynamic unstructured environments over large-scale distances, remains a challenging goal. One of the major difficulties in this objective is the precise localization of the vehicle within its environment so that autonomous navigation techniques can be employed. In this context, this paper presents a methodology to map building and to efficient pose computation which is specially adapted for cases of large displacements. Our method uses hybrid robust RGB-D cost functions that have different convergence properties, whilst exploiting the visibility rotation invariance given by panoramic spherical images. The proposed registration model is composed of a RGB and point-to-plane ICP cost in a multi-resolution framework. We close up the paper presenting mapping and localization results in real outdoor scenes.