Convergence-enhanced dense RGB-D odometry with a rotational motion prior from a gyroscope

In this paper, we address a convergence-enhanced dense RGB-D visual odometry with a rotational motion prior derived from a gyroscopic data. We focus on two main problems of dense methods, namely high computational loads and existence of many local minima. To deal with these problems, we use the rotational prior from the gyroscope and the general structure of the photometric error minimization procedure is divided into two-fold steps, 3-DoF partial problem and 6-DoF refinement problem, efficiently utilizing the rotational prior from the gyroscope. We evaluate performances of the proposed algorithm using various datasets and compare with the popular dense visual odometry, DVO. Our algorithm not only operates faster than DVO, but also relieves the high non-linearity of the cost function, so frequent motion estimation jumps are suppressed considerably.