Model-aided state estimation for quadrotor micro aerial vehicles
- Publication Type:
- Issue Date:
Due to their manoeuvrability, compactness and vertical take-off and landing capability, quadrotor Micro Aerial Vehicles (MAV) are ideally suited to assist or replace humans in a host of tasks in urban and indoor environments that would otherwise be hazardous, tedious or expensive. However, obtaining reliable pose estimates to perform these tasks safely and efficiently is a significant challenge due to the limited accuracy of GPS in such environments. This thesis presents algorithms for pose estimation of quadrotor Micro Aerial Vehicles (MAVs) operating in GPS-denied environments. The main contributions of the thesis stem from the use of the dynamic model describing the motion of a quadrotor as an additional source of information during state estimation. A state estimator design for quadrotor MAVs that only employs consumer grade inertial sensors is first proposed. Two major improvements to the conventional inertial only state estimators for MAVs are demonstrated. First, it is shown that incorporating an appropriate dynamic model improves the accuracy of the MAV attitude estimate. Second, in contrast to the conventional designs, it is shown that the new estimator provides a drift free estimate of the horizontal components of the quadrotor body frame velocity. These velocity estimates can be exploited to substantially improve the stability and controllability of a quadrotor MAV. In addition to inertial sensors, monocular cameras provide an excellent source of information that can be used for the MAV state estimation task. The complementary nature of visual and inertial information means that a fusion of the two information sources can improve the accuracy and robustness of the state estimation algorithms. This thesis demonstrates that further improvements in accuracy and robustness can be obtained by incorporating the quadrotor dynamic model into visual-inertial fusion algorithms. The resulting state estimator design is capable of producing reliable pose estimates even when the quadrotor MAV is travelling at a constant velocity, a case which is known to be difficult to handle with conventional algorithms. A theoretical analysis using Lie derivatives is presented to verify this improvement in observability. Extensive simulations and experiments in a number of practical situations are presented to demonstrate the effectiveness of the proposed methodology and to demonstrate that it outperforms conventional visual-inertial fusion methods. Employing the dynamic model to aid the state estimation can also be extended to deal with wind disturbances that would otherwise hamper the performance of lightweight quadrotor MAVs. This thesis demonstrates that explicit modelling of the effects of wind on the quadrotor dynamics enables the simultaneous estimation of the vehicle pose and two components of wind velocity, using only a monocular camera and an inertial measurement unit. This design is validated through a non-linear observability analysis and extensive simulations that makes use of a realistic wind model. Experimental results in a controlled lab environment are also presented to demonstrate the effectiveness of the proposed state estimator.
Please use this identifier to cite or link to this item: