Velocity estimator via fusing inertial measurements and multiple feature correspondences from a single camera

In this paper, we present a novel real-time velocity estimation algorithm. A sensor assembly consisting of a monocular camera and an inertial measurement unit with three-axis accelerotmeter and gyroscope is considered. To improve the robustness of the velocity estimator with respective to image noise, we apply a coarse-to-fine structure based on multiple feature correspondences over three consecutive frames. The presented algorithm starts with an initial guess by solving a set of linear equations from modified epipolar constraints, which has increased accuracy and computational efficiency in comparison to previous work. Then, a highly accurate velocity estimation is achieved by non-linear minimization of the reprojection errors using the Levenberg-Marquardt algorithm. We implement our approach and present the results both in simulation and on real data.