An initialization method for monocular visual localization of miniature aerial robots

This paper proposes a method to initiate the pose of an aerial robot from a glimpse of a monocular view during the rapid take-off. Unlike the existing filter-based methods that failed to estimate pose due to insufficient baseline, this paper utilizes a unified projective parametrization and a progressively suppressed refinement scheme for the non-planar homography estimations to tackle the baseline initialization problem, which has been notoriously and persistently encountered in the field of aerial robotics. Without pruning the elliptical uncertainty spheres iteratively in a filter-based framework as in the existing methods, the proposed method associates the unknown scene points and the pose of the agile aerial robot in a unified projective parametrization, and leverages a hypothesis-and-verify scheme to facilitate the decompositions of the non-planar homographies in line with a series of essential matrices. Moreover, a progressively suppressed strategy is introduced to minimize the hypothetic errors in the homographies and to minimize estimation divergence. Therefore, even with only a glance through a single camera onboard an aerial robot, a referable poses of the aerial robot can be initiated for the immediate navigation. In addition, the empirical results demonstrates that this method not only yields a consistent and referable estimation of the 6DoF pose of the robot in parallel with the agile movement of take-off, but also withstands the loop-closure evaluation in the 3D space.