Robust Pose Estimation for Multirotor UAVs Using Off-Board Monocular Vision

This paper deals with the problem of pose estimation (or motion estimation) for multirotor unmanned aerial vehicles (UAVs) by using only an off-board camera. An extended Kalman filter (EKF) is often adopted to solve this problem. However, the accuracy and robustness of an EKF are limited partly by the usage of an existing linear constant-velocity process model applicable to many rigid objects. For such a reason, a nonlinear constant-velocity process model featured with the characteristics of multirotor UAVs is proposed in this paper, the superiority of which is explained from the perspective of observability. With the new process model and a generic camera model, a practical EKF method suitable for conventional cameras and fish-eye cameras is then proposed. By taking EKF implementation into account, a general correspondence method that could handle any number of feature points is further designed. Simulation and real experiments show that the proposed EKF method is more robust against noise and occlusion than currently employed filtering methods.