Combining the Strengths of Sparse Interest Point and Dense Image Registration for RGB-D Odometry

Abstract:
Visual odometry, i.e., estimating ego-motion from camera images, is frequently used as a building block in robot navigation systems. In this paper, we propose an efficient approach that provides robust and accurate visual odometry from RGB-D cameras in a wide range of settings. We seamlessly combine dense RGB-D image registration with the alignment of sparse interest points. While the former approach is robust and accurate when perceiving the depth towards structures well in less textured parts of an environment, the latter often performs better, if well textured but less structured parts are visible. Our formulation also integrates interest points with strongly uncertain or no depth to make best use of the available images. In experiments, we demonstrate advantages of our approach over methods that either are based on dense image or sparse interest point matching.