Efficient compositional approaches for real-time robust direct visual odometry from RGB-D data

In this paper we give an evaluation of different methods for computing frame-to-frame motion estimates for a moving RGB-D sensor, by means of aligning two images using photometric error minimization. These kind of algorithms have recently shown to be very accurate and robust and therefore provide an attractive solution for robot ego-motion estimation and navigation. We demonstrate three different alignment strategies, namely the Forward-Compositional, the Inverse-Compositional and the Efficient Second-Order Minimization approach, in a general robust estimation framework. We further show how estimating global affine illumination changes, in general improves the performance of the algorithms. We compare our results with recently published work, considered as state-of-the art in this field, and show that our solutions are in general more precise and can perform in real-time on standard hardware.