Vision-Based Control for Fast 3-D Reconstruction With an Aerial Robot

Abstract

We propose an active perception controller to drive an aerial robot to localize 3-D features in an environment using an onboard monocular camera. The robot estimates feature positions with either an extended Kalman filter (EKF) or an unscented Kalman filter (UKF). For each filter, we derive a controller that seeks the most valuable robot motions for estimating the 3-D positions of features in the local robot body frame. The control algorithm uses an explicit expression for the gradient of the error covariance matrices for both the EKF and UKF at the next time step. This gradient approach is demonstrated in both simulations and real-world experiments on a quadrotor with a downward facing camera where it is shown to outperform a wide variety of the state-of-the-art active sensing strategies. Our proposed control algorithms are computationally efficient, making them well-suited for fast, online reconstructions onboard microaerial vehicles.