Model Predictive Control for Vision-Based Quadrotor Guidance

Abstract

A model predictive control algorithm for autonomous approaches of quadrotor helicopters to a window is presented in this paper. The target is selected by an operator in a reference image which is sent to the vehicle. A wide baseline image matching algorithm is used to obtain the target position in the current image. The vehicle guidance for the approach is realized by a nonlinear model predictive control algorithm which calculates the desired thrust and angular velocities while taking the goals of approaching the target orthogonally and keeping the target visible in the camera image into account. Moreover, two novel methods for the estimation of the orthogonal distance and orientation to the façade plane in which the target window is located are presented. The first method relies on laser rangefinder measurements while the second method is based on the decomposition of the estimated homography matrix between the façade plane in the reference and the live image. Monte Carlo simulations of approaches to a house show the accuracy and robustness of the vehicle guidance. The trajectories converge even from difficult start positions, for example approaches with acute angles between the façade and the quadrotor's orientation. As desired, the window is approached perpendicularly towards the end of the approaches.