Using dense point clouds as environment model for visual localization of mobile robot

TL;DR: This paper proposes to tackle wide field of view camera 3D pose estimation using intensities of the whole image and surrounding datasets previously acquired by a 3D laser scanner, and finds the direct use of image intensities withdraws features detection and matching issues and ensures more consistency than using geometric features.

Abstract: Camera 3D pose estimation, to be consistent and precise, can benefit from two things: a 3D model of the environment, as it is well known, and the photometric appearance of the environment. The latter recently received more attention from the research community. However, it is mainly tackled for conventional cameras and using 3D models obtained from their images and for this purpose. In parallel, recent tools like 3D laser scanners have been more and more improved and are now able to rapidly generate an accurate and colored dense point clouds of a scene. We propose in this paper to tackle wide field of view camera 3D pose estimation using intensities of the whole image and surrounding datasets previously acquired by a 3D laser scanner. The direct use of image intensities withdraws features detection and matching issues and ensures more consistency than using geometric features. The performance of the approach is proven in simulation and real experiments in indoor situations.