Visual Homography-based Pose Estimation of a Quadrotor using Spectral Features

Pose estimation of Unmanned Aerial Vehicles (UAV) using cameras is currently a very active task in computer and robotic vision. This is mainly because of the use of robots in GPS-denied environments. However, the use of visual information for ego-motion estimation presents several difficulties, such as features search, data association, inhomogeneous features distribution in the image. This work addresses these issues by the use of the so-called spectral features, and a down-looking monocular camera rigidly attached to a quadrotor. We propose a visual position and orientation estimation algorithm based on the discrete homography constraint induced by the presence of planar scenes. This homography constraint results more appropriate than the well-known epipolar constraint, which vanishes for a zero translation and loses rank in the case of planar scenes. The pose estimation algorithm is tested in a simulated dataset and compared with the corresponding ground truth.