Adaptive Selection of Color Images or Depth to Align RGB-D Point Clouds

Alignment of pairwise image point clouds is an important task in building environment maps with partial information. The combination of depth information and images provided by RGB-D cameras are often used to improve such alignment. However, when the environment is structured and its images show little texture, depth information is more reliable; on the other hand, when the images of the environment have enough texture, better results are achieved when texture information is used. In this paper, we propose a new adaptive approach to make the most effective selection of image or depth information in order to find a better alignment of points and thus better define the rigid transformation between two point clouds. Our approach uses an adaptive parameter based on the degree of texture of the scene, selecting not only FPFH and SURF descriptors, but also weighting the iterative ICP process. Datasets containing RGB-D data with textured and non textured images are used to validate our proposal.