Fusion of Monocular Visual-Inertial Measurements for Three Dimensional Pose Estimation

This work describes a novel fusion schema to estimate the pose of a UAV using inertial sensors and a monocular camera. The visual motion algorithm is based on the plane induced homography using so called spectral features. The algorithm is able to operate with images presenting small amount of corner-like features, which gives more robustness to the state estimation. The key contribution of the paper is the use of this visual algorithm in a fusion schema with inertial sensors, exploiting the complementary properties of these two sensors. Results are presented in simulation with six degrees of freedom motion that satisfies dynamic constraints of a quadcopter. Virtual views are generated from this simulated motion cropped from a real floor image. Simulation results show that the presented algorithm would have enough precision to be used in an on-board algorithm to control the UAV in hovering operations.